WO2021139327A1 - 一种音频信号处理方法、模型训练方法以及相关装置 - Google Patents

一种音频信号处理方法、模型训练方法以及相关装置 Download PDF

Info

Publication number
WO2021139327A1
WO2021139327A1 PCT/CN2020/124244 CN2020124244W WO2021139327A1 WO 2021139327 A1 WO2021139327 A1 WO 2021139327A1 CN 2020124244 W CN2020124244 W CN 2020124244W WO 2021139327 A1 WO2021139327 A1 WO 2021139327A1
Authority
WO
WIPO (PCT)
Prior art keywords
audio
signal
howling
input signal
audio input
Prior art date
Application number
PCT/CN2020/124244
Other languages
English (en)
French (fr)
Inventor
张金亮
余涛
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Priority to EP20911413.1A priority Critical patent/EP3998557B1/en
Publication of WO2021139327A1 publication Critical patent/WO2021139327A1/zh
Priority to US17/700,862 priority patent/US20220215853A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/003—Changing voice quality, e.g. pitch or formants
    • G10L21/007—Changing voice quality, e.g. pitch or formants characterised by the process used
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04R—LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; DEAF-AID SETS; PUBLIC ADDRESS SYSTEMS
    • H04R3/00—Circuits for transducers, loudspeakers or microphones
    • H04R3/02—Circuits for transducers, loudspeakers or microphones for preventing acoustic reaction, i.e. acoustic oscillatory feedback
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0208—Noise filtering
    • G10L21/0216—Noise filtering characterised by the method used for estimating noise
    • G10L21/0232—Processing in the frequency domain
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0208—Noise filtering
    • G10L21/0216—Noise filtering characterised by the method used for estimating noise
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0208—Noise filtering
    • G10L21/0264—Noise filtering characterised by the type of parameter measurement, e.g. correlation techniques, zero crossing techniques or predictive techniques
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0316—Speech enhancement, e.g. noise reduction or echo cancellation by changing the amplitude
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0316—Speech enhancement, e.g. noise reduction or echo cancellation by changing the amplitude
    • G10L21/0324—Details of processing therefor
    • G10L21/0332—Details of processing therefor involving modification of waveforms
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
    • G10L25/18—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being spectral information of each sub-band
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
    • G10L25/21—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters the extracted parameters being power information
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04M—TELEPHONIC COMMUNICATION
    • H04M9/00—Arrangements for interconnection not involving centralised switching
    • H04M9/08—Two-way loud-speaking telephone systems with means for conditioning the signal, e.g. for suppressing echoes for one or both directions of traffic
    • H04M9/082—Two-way loud-speaking telephone systems with means for conditioning the signal, e.g. for suppressing echoes for one or both directions of traffic using echo cancellers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Definitions

  • This application relates to the field of computer technology, in particular to audio signal processing.
  • a frequency shifter or a phase shifter can be used to process the input audio at the local end, that is, to destroy the audio information that is consistent with the phase of howling, so as to achieve howling suppression.
  • the present application provides an audio signal processing method, which can effectively locate and suppress howling points, and improve the accuracy of the audio signal processing process.
  • an embodiment of the present application provides an audio signal processing method, which can be applied to a system or program containing an audio signal processing function in a terminal device, which specifically includes:
  • the first audio input signal is input to a machine learning model to obtain a first howling point, and a first gain value is obtained according to the first howling point, wherein the first howling point is used to indicate the first howling point
  • the howling point of the frequency band corresponding to the effective audio signal in an audio input signal is used to indicate the suppression parameter of the first howling point
  • the first audio input signal is processed according to the first gain value , To obtain the second audio input signal;
  • the second audio input signal is detected to obtain a second howling point, and a second gain value is obtained according to the second howling point, and the second howling point is used to indicate that the second audio input signal is The howling point of the effective audio signal corresponding to the frequency band;
  • the second audio input signal is processed according to the second gain value to obtain an audio output signal.
  • an audio signal processing device including:
  • the input unit is configured to input the first audio input signal into a machine learning model to obtain a first howling point, and obtain a first gain value according to the first howling point, wherein the first howling point is Is used to indicate the howling point of the frequency band corresponding to the effective audio signal in the first audio input signal; the first gain value is used to indicate the suppression parameter of the first howling point, and the first gain value is used to process the The first audio input signal to obtain the second audio input signal;
  • the detecting unit is configured to detect the second audio input signal to obtain a second howling point, and obtain a second gain value according to the second howling point, and the second howling point is used to indicate the second howling point.
  • the processing unit is configured to process the second audio input signal according to the second gain value to obtain an audio output signal.
  • an embodiment of the present application provides a method for training a machine learning model, including: collecting a reference signal and a voice sample signal, where the reference signal is a howling signal determined based on at least two variable elements, and the variable elements include Program category, program running time period or program running position, the collected signal is used to indicate the effective voice during the call;
  • the feature training set is input to a machine learning model for at least one cycle of training to obtain a trained machine learning model, and the trained machine learning model is used to determine the corresponding howling point and gain value according to the audio input signal.
  • a fourth aspect of the present application provides a machine learning model training device, including: a collection unit, configured to collect a reference signal and a voice sample signal, the reference signal is a howling signal determined based on at least two variable elements, and the variable Elements include program category, program running time period or program running position, and the collected signal is used to indicate the effective voice during the call;
  • a generating unit configured to generate a feature training set according to the reference signal and the collected signal
  • the training unit is configured to input the feature training set into a machine learning model for at least one cycle of training to obtain a trained machine learning model, and the trained machine learning model is used to determine the corresponding howling according to the audio input signal Point and gain value.
  • a fifth aspect of the present application provides a computer device, including: a memory, a processor, and a bus system; the memory is used to store program code; the processor is used to execute the audio according to the instructions in the program code The signal processing method, or the machine model training method described in the above aspect.
  • a sixth aspect of the present application provides a computer-readable storage medium, the storage medium is used to store a computer program, the computer program is used to execute the audio signal processing method described in the above aspect, or the machine model described in the above aspect Training method.
  • embodiments of the present application provide a computer program product including instructions, which when run on a computer, cause the computer to execute the audio signal processing method described in the above aspect, or the machine described in the above aspect The training method of the model.
  • the first audio input signal Acquire the first audio input signal; then input the first audio input signal into the machine learning model to obtain the first gain value for processing the effective audio signal frequency band; and process the first audio input signal according to the first gain value to Obtain a second audio input signal; then detect the second audio input signal to obtain a second howling point, which is used to indicate the howling point of the frequency band corresponding to the ineffective audio signal in the second audio input signal ; Further processing the second audio input signal according to the second gain value to obtain an audio output signal, the second gain value is used to indicate the suppression parameter of the second howling point.
  • Figure 1 is a network architecture diagram of the operation of the audio signal processing system
  • FIG. 2 is a flow chart of audio signal processing provided by an embodiment of the application
  • FIG. 3 is a flowchart of an audio signal processing method provided by an embodiment of the application.
  • FIG. 5 is a schematic diagram of a scene of audio signal processing provided by an embodiment of this application.
  • FIG. 6 is a schematic diagram of another audio signal processing scenario provided by an embodiment of the application.
  • FIG. 7 is a flowchart of another audio signal processing method provided by an embodiment of the application.
  • FIG. 8 is a comparison diagram of audio signal processing provided by an embodiment of this application.
  • FIG. 9 is a flowchart of another audio signal processing method provided by an embodiment of the application.
  • FIG. 10 is a schematic diagram of an interface of an audio signal processing method provided by an embodiment of the application.
  • FIG. 11 is a schematic diagram of an interface of another audio signal processing method provided by an embodiment of the application.
  • FIG. 12 is a flowchart of a method for training a machine learning model provided by an embodiment of this application.
  • FIG. 13 is a schematic diagram of a process of training a machine learning model provided by an embodiment of this application.
  • FIG. 14 is a schematic structural diagram of an audio signal processing device provided by an embodiment of the application.
  • 15 is a schematic structural diagram of a machine learning model training device provided by an embodiment of the application.
  • FIG. 16 is a schematic structural diagram of a terminal device provided by an embodiment of this application.
  • FIG. 17 is a schematic structural diagram of a server provided by an embodiment of the application.
  • the embodiments of the present application provide an audio signal processing method and related device, which can be applied to a system or program containing an audio signal processing function in a terminal device by acquiring a first audio input signal; and then the first audio input signal Input the machine learning model to obtain the first gain value for processing the effective audio signal frequency band; and process the first audio input signal according to the first gain value to obtain the second audio input signal; then detect the second audio input signal , In order to obtain the second howling point, the second howling point is used to indicate the howling point of the frequency band corresponding to the non-effective audio signal; and then the second audio input signal is processed according to the second gain value to obtain the audio output Signal, the second gain value is used to indicate the suppression parameter of the second howling point.
  • the sound signal collected by the microphone is amplified by the speaker, and then picked up by the microphone.
  • the signal is continuously superimposed and amplified in the feedback loop, and the positive feedback generates an oscillation cycle, and then a phenomenon.
  • Howling point the frequency point where the loop gain is greater than or equal to 1 in the audio signal.
  • Effective audio signal indicates the target audio in the audio signal, such as the voice signal during a voice call.
  • Invalid audio signal Indicates the interference audio in the audio signal, such as environmental noise, echo, etc.
  • Gain value the degree of change in the processing of the audio signal of the specified frequency band, used in the howling suppression scene to indicate the reduction multiple of the corresponding audio signal for the howling point.
  • Pitch period the period of time each time the vocal cords are opened and closed in the process of human vocalization, which can be used to indicate the parameters of the effective audio signal.
  • Machine learning model parameter adjustment is performed by a given sample, so that the output model has similar characteristics of the given sample.
  • Power spectrum The change of signal power with frequency, that is, the distribution of signal power in the frequency domain.
  • ADC Analog-to-Digital Converter
  • Recurrent Neural Network A type of recurrent neural network that takes sequence data as input, recursively in the evolution direction of the sequence, and all nodes (recurrent units) are connected in a chain.
  • Convolutional Neural Networks have characterization learning capabilities and can classify input information according to its hierarchical structure.
  • the audio signal processing method provided in this application can be applied to a system or program containing audio signal processing functions in a terminal device, for example, as a voice plug-in for a game.
  • the audio signal processing system can run as shown in FIG. 1
  • Figure 1 shows the network architecture of the audio signal processing system.
  • the audio signal processing system can provide audio signal processing with multiple information sources, and the terminal establishes a connection with the server through the network. , And then receive audio signals sent by other terminals, and suppress howling by the audio signal processing method provided in this application on the received signals to obtain audio output, thereby realizing the audio interaction process between multiple terminals; understandable
  • Figure 1 shows a variety of terminal devices.
  • the audio signal processing method provided in this embodiment can also be performed offline, that is, without the participation of the server, at this time the terminal interacts with other terminals in the audio signal locally, and then performs the audio signal processing process between the terminals. .
  • the above-mentioned audio signal processing system can be run on personal mobile terminals, for example: as an application such as a game voice plug-in, can also run on a server, and can also be run on a third-party device to provide audio signal processing to obtain information
  • the audio signal processing result of the source can be run in the above-mentioned device in the form of a program, can also be run as a system component in the above-mentioned device, and can also be used as a kind of cloud service program,
  • the specific operation mode depends on the actual scenario and is not limited here.
  • a frequency shifter or a phase shifter can be used to process the input audio at the local end, that is, to destroy the audio information that is consistent with the phase of howling, so as to achieve howling suppression.
  • this application proposes an audio signal processing method, which is applied to the audio signal processing process framework shown in FIG. 2.
  • an audio signal processing method provided by an embodiment of this application is shown in FIG. The signal processing process architecture diagram.
  • the terminal device collects the user's voice and converts it into an audio signal, and then inputs the trained machine learning model to filter and suppress the howling points, and further gain the unprocessed howling points Control, so as to obtain the audio signal after howling suppression as output.
  • the method provided in this application can be a kind of program writing, as a kind of processing logic in a hardware system, or as an audio signal processing device, which can be integrated or externally implemented to realize the aforementioned processing. logic.
  • the audio signal processing device obtains a first audio input signal; then inputs the first audio input signal into a machine learning model to obtain a first gain value for processing an effective audio signal frequency band; and according to the first audio input signal;
  • the gain value processes the first audio input signal to obtain a second audio input signal; then the second audio input signal is detected to obtain a second howling point, the second howling point is used to indicate that the audio signal is not valid Corresponding to the howling point of the frequency band; further processing the second audio input signal according to a second gain value to obtain an audio output signal, the second gain value being used to indicate the suppression parameter of the second howling point.
  • FIG. 3 is a flowchart of an audio signal processing method provided by an embodiment of this application.
  • the embodiments of this application include at least the following step:
  • the first audio input signal can be the initial audio signal when the voice call is started; it can also be the audio signal after a certain period of time.
  • the howling is caused by the continuous gain of the audio signal in the feedback loop.
  • the process that is, the process of feedback gain accumulation, where the feedback loop is the loop composed of the local microphone and the opposite speaker; therefore, the accumulated feedback gains of audio signals in different time periods may be different, which can immediately evoke the audio processing method provided by this application , You can also wait for the feedback gain to be greater than or equal to 1 before awakening the audio processing method provided in this application. This is because the feedback gain in the loop of the audio signal needs to be greater than or equal to 1 for howling.
  • the acquired first audio input signal may be a signal that has been preliminarily amplified.
  • the acquisition signal is acquired first, and the acquisition signal may be acquired by a microphone or other acquisition equipment; then the acquisition signal is converted into a digital signal For example, the conversion is performed by an ADC; the digital signal is further input to an amplifier to obtain the first audio input signal. Since the first audio input signal is amplified, it is convenient for users to listen to it on the one hand, and on the other hand, it is convenient for the subsequent selection process of howling points in this application.
  • the collected signal collected by the collecting device may contain obvious noise, for example: a signal with a frequency far greater than the voice range; at this time, preliminary noise filtering can be performed.
  • the digital signal is input to an amplifier to obtain an amplified signal; then the amplified signal is processed according to the filter parameters to obtain a filtered amplified signal; the filtered amplified signal is further Fourier transformed to the frequency domain to obtain The first audio input signal.
  • the filter parameter may be a fixed value, or may be a targeted setting based on the frequency band corresponding to the common noise in the historical record.
  • the first howling point is used to indicate the howling point of the frequency band corresponding to the effective audio signal in the first audio input signal;
  • the first gain value is used to indicate the suppression parameter of the first howling point, according to the first gain value Process the first audio input signal to obtain a second audio input signal;
  • the machine learning model is trained based on multiple training signals, the training signal contains multiple howling point samples, and the first howling point is used to indicate The effective audio signal corresponds to the howling point of the frequency band.
  • the howling point Since there are some differences in the frequency band distribution or energy characteristics of the howling point and the effective speech, it can be extracted by determining multiple characteristics in the audio input signal. These characteristics can be selected based on the characteristics of the effective audio signal, such as: frequency band distribution, pitch The location of the cycle, the frequency of signal fluctuations, etc.; and then input the machine learning model to determine the corresponding first howling point; and then determine the corresponding first gain value according to the first howling point.
  • the feature extraction can be based on the effective audio signal.
  • the parameter feature information of the effective audio signal such as the Mel frequency cepstrum coefficient of the effective audio signal, or the mathematical deformation of the coefficient
  • It can be based on the biological characteristic information of the effective audio signal, such as the pitch period. This is because the audio signal within 500 Hz of the human voice has a pitch period, but the howling signal does not exist; the three aspects can also be based on the waveform characteristic information of the effective audio signal, for example The judgment is made according to the fluctuation of the effective audio signal in a specific frequency band, which is due to the short-term stability of the effective audio signal.
  • the examples of the above-mentioned features are for illustration only.
  • the specific features can be features that indicate valid audio signals, or features that indicate howling signals, and can also be features that indicate the distinguishing features between effective audio signals and howling signals, which will not be described here. limited.
  • the first audio input signal can be adjusted to the target frequency to convert to the frequency domain; for example, the general mobile phone voice call uses the sampling rate of 16KHz, so the The target frequency is adjusted to 16KHz; then, multiple sampling points in the first audio input signal converted to the frequency domain are determined; and multiple audio features are extracted based on the sampling points.
  • the input signal is processed in multiple threads, and the efficiency of audio processing is improved.
  • the input signal in order to make the time domain signal better meet the periodicity requirements in the Fourier transform process and reduce signal omissions, can be divided Performing based on a window function, that is, dividing the first audio input signal converted to the frequency domain based on the window function to obtain a plurality of subbands; and then determining a plurality of the sampling points in the subband.
  • the window function can be a rectangular window, a Gaussian window or a Kaiser window, etc. The specific function form depends on the actual scene.
  • the first gain value corresponds to multiple howling points in the first audio input signal, and each howling point corresponds to multiple frequency bands, and the collection of these frequency bands is called a subband; therefore, the first gain value It can include multiple howling suppression gain values, and each howling suppression gain is a floating-point number from 0 to 1; the first gain value is input into the first audio input signal, and each subband is multiplied by the howling of the corresponding subband. Called the suppression attenuation gain, you can get the result of the machine learning howling suppression processing, that is, the second audio input signal.
  • the second howling point is used to indicate the howling point of the frequency band corresponding to the ineffective audio signal in the second audio input signal, that is, the howling point that does not belong to the frequency band corresponding to the effective audio signal. Since there may be howling points corresponding to the frequency band of the unprocessed ineffective audio signal in the machine model training, a secondary gain process is performed, that is, the second howling point can be detected.
  • detecting the second howling point may be by obtaining the power spectrum corresponding to the second audio input signal; then detecting the extreme value in the power spectrum, for example: the maximum power in the power spectrum, or setting based on the maximum power Determine the value range; then determine the corresponding candidate frequency points according to the extreme value, that is, these frequency points may be howling points; and then determine the second howling point according to the candidate frequency points. That is, the phase and feedback gain information of the candidate frequency point are detected, and if the phase is the same and the feedback gain is greater than or equal to 1, it is determined as the second howling point.
  • the gain change of the frequency point can be judged intuitively. This is because the power corresponding to the howling point is often greater than the power of the general frequency point, thereby improving the accuracy of howling point identification.
  • the second howling point can be judged according to the peak-to-average ratio, that is, multiple frequency points adjacent to the candidate frequency point are obtained to determine the candidate range; and then the average frequency average value of the frequency points in the candidate range is determined .
  • the peak-to-average ratio when the peak-to-average ratio is greater than the howling threshold, the candidate frequency point is determined to be the second howling point.
  • the howling points can be judged through the peak-to-average comparison, thereby expanding the range of data reference and further improving the accuracy of howling point recognition.
  • the judgment of the howling points can also be obtained by statistical analysis based on historical records.
  • the howling points are easily concentrated above 2KHz
  • the energy of the voice signal is mainly concentrated in the frequency band below 2KHz.
  • the second detection of the range is performed in the identification of howling points.
  • the specific detection method can refer to the identification method of the power spectrum extreme value or the peak-to-average ratio.
  • the frequency band where the howling points are concentrated depends on specific scenarios, that is, in different scenarios, the frequency band where the howling points are concentrated can be higher or lower.
  • the frequency band where the howling points are concentrated can be higher or lower.
  • only the historical records are analyzed to obtain the howling. The method of ordering is explained without limitation.
  • the second gain value is used to indicate the suppression parameter of the second howling point, that is, the reduction multiple of the frequency band corresponding to the second howling point. After the second selection of howling points, the accuracy of howling suppression and the significance of howling suppression effect are guaranteed.
  • the second gain value can be empirically set a floating point value in the range of 0 to 1, or can be calculated based on the energy of the upper and lower adjacent subbands.
  • the processed signal after processing the second audio input signal according to the second gain value, the processed signal can also be converted to the time domain and notched, which is a kind of filter, to further eliminate howling points. .
  • the first audio input signal is obtained; then the first audio input signal is input into the machine learning model to obtain the first gain value for processing the effective audio signal frequency band; and the first gain value is processed according to the first gain value.
  • the howling point of the audio signal corresponding to the frequency band and further processing the second audio input signal according to a second gain value to obtain an audio output signal, and the second gain value is used to indicate the suppression parameter of the second howling point.
  • FIG. 4 is an embodiment of the application.
  • a flowchart of another audio signal processing method is provided. The embodiment of the present application includes at least the following steps:
  • steps 401 to 404 are similar to steps 301 to 304 of the embodiment indicated in FIG. 3, and related feature descriptions can be referred to, which will not be repeated here.
  • the voice protection is to ensure the integrity of the effective audio signal. Specifically, first obtain the feature information in the effective audio signal, which is determined based on the waveform feature indicated by the effective audio signal. For example, the voiced sound has a formant in the waveform feature indicated by the effective audio signal, and the high frequency energy of the unvoiced sound is large and pressurized.
  • the energy slope of the frequency axis is stable; then the corresponding valid audio signal in the second audio input signal is detected according to the characteristic information; further a locking operation is performed on the valid audio signal, and the locking operation is used to indicate the non-effect of the second gain value
  • the object, that is, the processing frequency band corresponding to the second gain value may contain the effective audio signal locked in this step, but no gain processing is performed on the signals in these frequency bands.
  • the voice protection for the effective audio signal can also be performed based on the historical record used to indicate the voice frequency band, that is, the frequency band distribution of the effective audio signal is counted, and the frequency bands with a large distribution weight are detected and screened one by one.
  • ⁇ is a smoothing factor from 0 to 1; Is the suppression gain of the previous frame; m is the frame index; k is the frequency index.
  • the above formula adjusts the gain difference between adjacent frames to make the gain between adjacent frames closer to the linear distribution, reducing sudden audio changes, making the audio output signal smoother in the auditory sense, and improving user experience.
  • the gain parameter in step 407 is used Multiply by the value of the corresponding frequency point to get the audio output signal.
  • FIG. 5 it is a schematic diagram of an audio signal processing scene provided by an embodiment of the present application.
  • the picture shows the scene where the microphone collects voice signals and amplifies them for playback; because the sound source (microphone) is too close to the amplification equipment (speaker), the sound signal collected by the microphone is amplified by the speaker and then picked up by the microphone, and the signal is in the feedback loop
  • the positive feedback produces an oscillating cycle, and then produces howling.
  • the function of positive feedback generating oscillation can be:
  • condition of howling requires that the phase of the input signal collected by the microphone in the feedback loop is the same as the phase of the sound wave signal fed back to the speaker, namely:
  • G(s) is the input signal collected by the microphone; F(s) is the sound wave signal fed back to the speaker; G(w0) is the phase of the input signal collected by the microphone; F(w0) is the feedback to the speaker The phase of the acoustic signal in; n is an integer parameter.
  • the audio signal processing method provided by the present application can be executed in the amplifier, that is, after the audio signal collected by the microphone is transmitted to the amplifier, the audio signal processing process of the embodiment shown in FIG. 3 or FIG. 4 is immediately performed. Then the output signal is transmitted to the loudspeaker for playback, and so on, the effect of howling suppression can be achieved.
  • FIG. 6 it is a schematic diagram of another audio signal processing scenario provided by an embodiment of the present application, and the figure shows a loop in the scenario where the terminal is put outside.
  • the sound from the right terminal speaker comes out and is picked up by the left terminal microphone.
  • the network After pre-processing and signal conversion, it is sent to the right terminal through the network. Played out through the speaker, and then picked up by the left terminal microphone.
  • the loop gain is greater than or equal to 1 at a certain frequency point, and the phase is positive, then this point will form a howling point.
  • FIG. 7 is a flowchart of another audio signal processing method provided in an embodiment of the application.
  • the embodiment of the application includes at least the following steps:
  • the target frequency can be set to 16KHz.
  • the audio signal is transformed into the frequency domain, and is windowed for Fourier transformation to the frequency domain.
  • the window function may be a rectangular window, a Gaussian window, or a Kaiser window, etc.
  • the specific function form depends on the actual scene.
  • Extract 42 feature values.
  • the characteristic value may include 22 Mel-scale Frequency Cepstral Coefficients (MFCC), and the coefficient may refer to the parameter in the speech recognition process, that is, the effective audio signal; the characteristic value may also include the previous The first or second derivative of the 6 coefficients is used to indicate the characteristics of the voice; the characteristic value can also include the gene period, which is because the voiced sound of the voice signal has a gene period within 500 Hz, but the howling signal does not; the characteristic value can also include The detection of non-stationary eigenvalues is because the speech is short-term stationary.
  • MFCC Mel-scale Frequency Cepstral Coefficients
  • the machine learning model adopts a recurrent neural network model, which is for modeling time series, rather than just considering input and output frames.
  • the specific process of obtaining the first gain value is similar to step 302 in the embodiment described in FIG. 3, and details are not described here.
  • steps 705-708 are similar to steps 303-305 of the embodiment shown in FIG. 3, and related feature descriptions can be referred to, and details are not repeated here.
  • FIG. 8 is a comparison diagram of audio signal processing provided by an embodiment of this application; the upper picture is the spectrogram of the input signal before howling suppression, and the lower The spectrogram of the signal after howling suppression processing.
  • the clutter around the wave crest of the sample is significantly reduced, that is, before the howling, the howling has been suppressed by the audio processing method provided in this application.
  • FIG. 9 is a flowchart of another audio signal processing method provided by an embodiment of the application. The embodiment includes at least the following steps:
  • the start instruction of the game may be the start of the game, or the triggering of a certain scene thread in the game, for example: entering a battle scene.
  • the characteristic element is a physical or virtual button that activates the voice call function.
  • Figure 10 it is a schematic interface diagram of an audio signal processing method provided by an embodiment of the present application; the figure shows the interface in the game interface.
  • the characteristic element A1 when any of the buttons is triggered, evokes the audio processing method described in the embodiment of FIG. 3 or FIG. 4.
  • the audio processing method in this application is not only used in the voice call process of two users, but also can be applied in the voice call process of multiple users; as shown in FIG. 11, it is another example provided by the embodiment of the present application.
  • An interface schematic diagram of an audio signal processing method In the figure, the user is in a public voice scene B2. At this time, if the feature element B1 is triggered, the audio processing method described in the embodiment of FIG. 3 or FIG. 4 is invoked.
  • the above-mentioned howling suppressed audio signal is used for input to realize a clear voice call process between two or more users.
  • FIG. 12 is a flowchart of a method for training a machine learning model provided by an embodiment of the application.
  • the embodiment of the application includes at least the following steps:
  • the reference signal is a howling signal determined based on at least two variable elements, and the variable elements include program category, program running period, or program running position, and the voice sample signal is used to indicate that the voice call is in progress Effective voice.
  • the program categories in the variable elements can be different games, such as training samples in different game scenarios such as the glory of the king and the elite of peace.
  • the program running time period indicates the time period when training samples are collected.
  • the game usually has a voice call function from 8 pm to 9 pm, and the call voice is more intense, so additional annotations can be made and training samples can be generated.
  • the location of the program is the geographic information collected by voice, for example, the training samples are collected in different geographic locations such as the market, the teacher, or the bedroom.
  • the machine learning model is The howling point of the frequency band has good recognition ability.
  • corresponding labels are set and classified based on the signals collected under the above-mentioned different factors; and the corresponding howling points are marked to generate a feature training set.
  • the trained machine learning model is used to determine the corresponding howling point and gain value according to the audio input signal.
  • FIG. 13 it is a schematic diagram of a machine learning model training process provided by an embodiment of the present application.
  • the figure shows an RNN model, which includes a 3-level gate recurrence unit (gated recurrence unit, GRU).
  • GRU gate recurrence unit
  • the reset gate determines whether to memorize the current state for calculating the new state
  • the update gate determines how much the current state will change according to the new input.
  • the update door is closed, the GRU can remember the training information for a long time.
  • the first layer GRU inputs 42 dimensions, outputs 24 dimensions and a voice activity detection (VAD) flag.
  • the GRU of the second layer inputs the initial 42-dimensional features and the 24-dimensional features output by the first layer to output 48 dimensions, which are used to estimate the howling signal.
  • the third layer inputs the initial 42-dimensional features and the second-layer output 42-dimensional features to obtain the output; and adjusts the output according to the gain value in the training sample to update the model parameters, thereby realizing the training of the RNN model.
  • training process in this application can also be applied to a deep neural network model or a convolutional neural network model, which will not be repeated here.
  • the howling point distribution and the corresponding first gain value can be obtained after the audio signal is input to the machine learning model, thereby ensuring the accuracy of howling suppression in the voice frequency band.
  • FIG. 14 is a schematic structural diagram of an audio signal processing device according to an embodiment of the application.
  • the audio signal processing device 1400 includes:
  • the obtaining unit 1401 is configured to obtain the first audio input signal
  • the input unit 1402 is configured to input the first audio input signal into a machine learning model to obtain a first gain value, where the first gain value is used to indicate the first howling point in the first audio signal Suppression parameter, the first howling point is used to indicate the howling point of the frequency band corresponding to the effective audio signal;
  • the detecting unit 1403 is configured to process the first audio input signal according to the first gain value to obtain a second audio input signal
  • the processing unit 1404 is configured to process the second audio input signal according to a second gain value to obtain an audio output signal, where the second gain value is used to indicate the suppression parameter of the second howling point, and the second howling The point is used to indicate the howling point of the frequency band that is not the effective audio signal.
  • the input unit 1402 is specifically configured to convert the audio input signal to the frequency domain to extract multiple audio features, and the audio features are based on the effective audio Determining the characteristics of the signal or the howling sample;
  • the input unit 1402 is specifically configured to input the audio feature into the machine learning model to determine the first howling point;
  • the input unit 1402 is specifically configured to determine the corresponding first gain value according to the first howling point.
  • the input unit 1402 is specifically configured to process the first audio input signal according to the first gain value to obtain a second audio input signal
  • the detection unit 1403 is configured to detect the second audio input signal to obtain a second howling point, and obtain a second gain value according to the second howling point, and the second howling point is used to indicate the The howling point of the frequency band corresponding to the ineffective audio signal in the second audio input signal;
  • the processing unit 1404 is configured to process the second audio input signal according to the second gain value to obtain an audio output signal.
  • the input unit 1402 is specifically configured to convert the audio input signal to the frequency domain to extract a plurality of audio features, and the audio feature is determined based on the effective audio signal or the feature of the howling sample ;
  • the input unit 1402 is specifically configured to input the audio feature into the machine learning model to determine the first howling point;
  • the input unit 1402 is specifically configured to obtain the corresponding first gain value according to the first howling point.
  • the input unit 1402 is specifically configured to adjust the first audio input signal to a target frequency to convert to the frequency domain;
  • the input unit 1402 is specifically configured to determine multiple sampling points in the first audio input signal converted to the frequency domain;
  • the input unit 1402 is specifically configured to extract multiple audio features based on the sampling points.
  • the input unit 1402 is specifically configured to divide the first audio input signal converted to the frequency domain based on a window function to obtain multiple subbands;
  • the input unit 1402 is specifically configured to determine the multiple sampling points in the subband.
  • the detection unit 1403 is specifically configured to obtain a power spectrum corresponding to the second audio input signal
  • the detection unit 1403 is specifically configured to detect extreme values in the power spectrum and determine the corresponding candidate frequency points
  • the detection unit 1403 is specifically configured to determine the second howling point according to the candidate frequency point
  • the detection unit 1403 is specifically configured to process the second howling point according to the second gain value to obtain the audio output signal.
  • the detection unit 1403 is specifically configured to obtain multiple adjacent frequency points of the candidate frequency point to determine the candidate range
  • the detection unit 1403 is specifically configured to determine the average frequency average of the mid-frequency points in the candidate range to obtain the peak-to-average ratio;
  • the detection unit 1403 is specifically configured to determine that the candidate frequency point is the second howling point if the peak-to-average ratio is greater than the howling threshold.
  • the detection unit 1403 is further configured to obtain feature information in a valid audio signal, and the feature information is determined based on the waveform feature indicated by the valid audio signal.
  • the effective audio signal is used to indicate voice samples;
  • the detection unit 1403 is specifically configured to detect a corresponding valid audio signal in the second audio input signal according to the characteristic information
  • the detection unit 1403 is specifically configured to perform a locking operation on the effective audio signal, and the locking operation is used to indicate an inactive object of the second gain value.
  • the processing unit 1404 is specifically configured to determine multiple gain frames corresponding to the second howling point
  • the processing unit 1404 is specifically configured to process the gain frame according to a smoothing formula to update the audio output signal.
  • the obtaining unit 1401 is specifically configured to obtain a collection signal
  • the acquiring unit 1401 is specifically configured to convert the collected signal into a digital signal
  • the obtaining unit 1401 is specifically configured to input the digital signal into an amplifier to obtain the first audio input signal.
  • the acquiring unit 1401 is specifically configured to input the digital signal into an amplifier to obtain an amplified signal
  • the acquiring unit 1401 is specifically configured to process the amplified signal according to filter parameters to obtain a filtered amplified signal
  • the acquiring unit 1401 is specifically configured to perform Fourier transform on the filtered amplified signal to the frequency domain to obtain the first audio input signal.
  • the audio signal processing method is applied in a game voice call process, and the acquiring unit 1401 is specifically configured to detect the triggering of a characteristic element, and the characteristic element is Elements in the game interface;
  • the acquiring unit 1401 is specifically configured to acquire the first audio input if the characteristic element is triggered.
  • the first audio input signal Acquire the first audio input signal; then input the first audio input signal into the machine learning model to obtain the first gain value for processing the effective audio signal frequency band; and process the first audio input signal according to the first gain value to Obtain a second audio input signal; then detect the second audio input signal to obtain a second howling point, which is used to indicate the howling point of the frequency band corresponding to the ineffective audio signal in the second audio income signal ; Further processing the second audio input signal according to the second gain value to obtain an audio output signal, the second gain value is used to indicate the suppression parameter of the second howling point.
  • the present application also provides a machine learning model training device 1500.
  • FIG. 15 it is a schematic structural diagram of a machine learning model training device provided in an embodiment of the present application.
  • the device includes a collection unit 1501 for collecting reference signals and A voice sample signal, the reference signal is a howling signal determined based on at least two variable elements, the variable elements including program category, program running period or program running position, and the collected signal is used to indicate the effective voice during the call ;
  • a generating unit 1502 configured to generate a feature training set according to the reference signal and the collected signal
  • the training unit 1503 is configured to input the feature training set into a machine learning model for at least one cycle of training to obtain a trained machine learning model, and the trained machine learning model is used to determine the corresponding howling according to the audio input signal. Call points and gain values.
  • the embodiment of the present application also provides a terminal device.
  • FIG. 16 it is a schematic structural diagram of another terminal device provided by the embodiment of the present application.
  • the terminal can be any terminal device including a mobile phone, a tablet computer, a personal digital assistant (PDA), a point of sales (POS), a car computer, etc. Take the terminal as a mobile phone as an example:
  • FIG. 16 shows a block diagram of a part of the structure of a mobile phone related to a terminal provided in an embodiment of the present application.
  • the mobile phone includes: a radio frequency (RF) circuit 1610, a memory 1620, an input unit 1630, a display unit 1640, a sensor 1650, an audio circuit 1660, a wireless fidelity (WiFi) module 1670, and a processor 1680 , And power supply 1690 and other components.
  • RF radio frequency
  • the RF circuit 1610 can be used for receiving and sending signals during the process of sending and receiving information or talking. In particular, after receiving the downlink information of the base station, it is processed by the processor 1680; in addition, the designed uplink data is sent to the base station.
  • the RF circuit 1610 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low noise amplifier (LNA), a duplexer, and the like.
  • the RF circuit 1610 can also communicate with the network and other devices through wireless communication.
  • the memory 1620 may be used to store software programs and modules.
  • the processor 1680 executes various functional applications and data processing of the mobile phone by running the software programs and modules stored in the memory 1620.
  • the input unit 1630 can be used to receive inputted digital or character information, and generate key signal input related to user settings and function control of the mobile phone.
  • the input unit 1630 may include a touch panel 1631 and other input devices 1632.
  • the display unit 1640 can be used to display information input by the user or information provided to the user and various menus of the mobile phone.
  • the mobile phone may also include at least one sensor 1650, such as a light sensor, a motion sensor, and other sensors. .
  • the audio circuit 1660, the speaker 1661, and the microphone 1662 can provide an audio interface between the user and the mobile phone.
  • the audio circuit 1660 can transmit the electrical signal converted from the received audio data to the speaker 1661, which is converted into a sound signal for output by the speaker 1661; After being received, it is converted into audio data, and then processed by the audio data output processor 1680, and then sent to another mobile phone via the RF circuit 1610, or the audio data is output to the memory 1620 for further processing.
  • WiFi is a short-range wireless transmission technology.
  • FIG. 16 shows the WiFi module 1670, it is understandable that it is not a necessary component of a mobile phone and can be omitted as needed without changing the essence of the invention.
  • the processor 1680 is the control center of the mobile phone. It uses various interfaces and lines to connect various parts of the entire mobile phone. It executes by running or executing software programs and/or modules stored in the memory 1620, and calling data stored in the memory 1620. Various functions and processing data of the mobile phone can be used to monitor the mobile phone as a whole.
  • the processor 1680 included in the terminal also has the function of executing each step of the above-mentioned audio signal processing method or model training method.
  • FIG. 17 is a schematic diagram of a structure of a server provided by an embodiment of the present application.
  • the server 1700 may have relatively large differences due to different configurations or performances, and may include a Or one or more central processing units (CPU) 1722 and memory 1732, and one or more storage media 1730 for storing application programs 1742 or data 1744.
  • CPU central processing units
  • the server 1700 may also include one or more power supplies 1726, one or more wired or wireless network interfaces 1750, one or more input and output interfaces 1758, and/or one or more operating systems 1741.
  • the steps performed by the model training device in the foregoing embodiment may be based on the server structure shown in FIG. 17.
  • An embodiment of the present application also provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and the computer program is used to execute the audio frequency in the method described in the embodiments shown in FIG. 2 to FIG. 13. The steps performed by the signal processing device.
  • the embodiment of the present application also provides a computer program product including audio signal processing instructions.
  • the computer can execute the method described in the method described in the above-mentioned embodiments shown in FIGS. 2 to 13. Steps performed.
  • the embodiment of the present application also provides an audio signal processing system.
  • the audio signal processing system may include the audio signal processing apparatus in the embodiment described in FIG. 14 or the terminal device described in FIG. 16.
  • the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the technical solution of the present application essentially or the part that contributes to the existing technology or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , Including several instructions to make a computer device (which can be a personal computer, an audio signal processing device, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Acoustics & Sound (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Quality & Reliability (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Otolaryngology (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Circuit For Audible Band Transducer (AREA)
  • Telephone Function (AREA)

Abstract

本申请公开了一种音频信号处理方法、模型训练方法以及相关装置，通过对第一音频输入信号输入机器学习模型，以得到第一啸叫点以及对应的第一增益值；并根据该第一增益值处理该第一音频输入信号，以得到第二音频输入信号；接下来检测该第二音频输入信号，以得到第二啸叫点，然后根据第二增益值对第二音频输入信号中的第二啸叫点进行处理，以得到音频输出信号。从而实现了对于音频输入信号中啸叫的抑制，使得啸叫在初始阶段就被抑制，无法再进行增益循环；由于机器学习模型中指示啸叫点的对应性以及模型计算的便捷性，使得上述处理过程迅速，且不会对有效音频信号产生影响，提高了音频处理的准确性及效率。

Description

一种音频信号处理方法、模型训练方法以及相关装置
本申请要求于2020年01月09日提交中国专利局、申请号为202010023045.6、申请名称为“一种音频信号处理方法、模型训练方法以及相关装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，尤其涉及音频信号处理。
背景技术
随着移动终端相关技术的发展，越来越多的智能设备出现在人们的生活中，其中，通过智能设备进行语音通话尤为突出，然而由于通话过程中本端麦克风也会采集到对端扬声器的音频信号，这些音频信号可能在本端与对端的语音过程中循环，尤其在近距离的语音过程中，音频信号会不断的循环增益，从而产生啸叫。
一般，可以采用移频器或移相器对本端的输入音频进行处理，即破坏与啸叫产生的相位一致的音频信息，从而实现啸叫抑制。
发明内容
有鉴于此，本申请提供一种音频信号处理的方法，可以有效定位啸叫点并进行啸叫抑制，提高音频信号处理过程的准确性。
一方面，本申请实施例提供一种音频信号处理的方法，可以应用于终端设备中包含音频信号处理功能的***或程序中，具体包括：
获取第一音频输入信号；
将所述第一音频输入信号输入机器学习模型，以得到第一啸叫点，根据所述第一啸叫点获得第一增益值，其中，所述第一啸叫点用于指示所述第一音频输入信号中有效音频信号对应频段的啸叫点；所述第一增益值用于指示所述第一啸叫点的抑制参数，根据所述第一增益值处理所述第一音频输入信号，以得到第二音频输入信号；
检测所述第二音频输入信号，以得到第二啸叫点，根据所述第二啸叫点获得第二增益值，所述第二啸叫点用于指示所述第二音频输入信号中非有效音频信号对应频段的啸叫点；
根据第二增益值对第二音频输入信号进行处理，以得到音频输出信号。
另一方面，本申请实施例提供一种音频信号处理的装置，包括：
获取单元，用于获取第一音频输入信号；
输入单元，用于将所述第一音频输入信号输入机器学习模型，以得到第一啸叫点，根据所述第一啸叫点获得第一增益值，其中，所述第一啸叫点用于指示所述第一音频输入信号中有效音频信号对应频段的啸叫点；所述第一增益值用于指示所述第一啸叫点的抑制参数，根据所述第一增益值处理所述第一音频输入信号，以得到第二音频输入信号；
检测单元，用于检测所述第二音频输入信号，以得到第二啸叫点，根据所述第二啸叫点获得第二增益值，所述第二啸叫点用于指示所述第二音频输入信号中非有效音频信号对应频段的啸叫点；
处理单元，用于根据第二增益值对第二音频输入信号进行处理，以得到音频输出信号。
另一方面，本申请实施例提供一种机器学习模型训练的方法，包括：采集参考信号和 语音样本信号，所述参考信号为基于至少两种变量元素确定的啸叫信号，所述变量元素包括程序类别、程序运行时段或程序运行位置，所述采集信号用于指示通话过程中的有效语音；
根据所述参考信号和采集信号生成特征训练集；
将所述特征训练集输入机器学习模型进行至少一个循环的训练，以得到训练后的机器学习模型，所述训练后的机器学习模型用于根据音频输入信号确定对应的啸叫点以及增益值。
本申请第四方面提供一种机器学习模型训练的装置，包括：采集单元，用于采集参考信号和语音样本信号，所述参考信号为基于至少两种变量元素确定的啸叫信号，所述变量元素包括程序类别、程序运行时段或程序运行位置，所述采集信号用于指示通话过程中的有效语音；
生成单元，用于根据所述参考信号和采集信号生成特征训练集；
训练单元，用于将所述特征训练集输入机器学习模型进行至少一个循环的训练，以得到训练后的机器学习模型，所述训练后的机器学习模型用于根据音频输入信号确定对应的啸叫点以及增益值。
本申请第五方面提供一种计算机设备，包括：存储器、处理器以及总线***；所述存储器用于存储程序代码；所述处理器用于根据所述程序代码中的指令执行上述方面所述的音频信号的处理方法，或上述方面所述的机器模型的训练方法。
本申请第六方面提供一种计算机可读存储介质，所述存储介质用于存储计算机程序，所述计算机程序用于执行上述方面所述的音频信号的处理方法，或上述方面所述的机器模型的训练方法。
又一方面，本申请实施例提供了一种包括指令的计算机程序产品，当其在计算机上运行时，使得所述计算机执行上述方面所述的音频信号的处理方法，或上述方面所述的机器模型的训练方法。
从以上技术方案可以看出，本申请实施例具有以下优点：
通过获取第一音频输入信号；然后将该第一音频输入信号输入机器学习模型，以得到处理有效音频信号频段的第一增益值；并根据该第一增益值处理该第一音频输入信号，以得到第二音频输入信号；接下来检测该第二音频输入信号，以得到第二啸叫点，该第二啸叫点用于指示第二音频输入信号中非有效音频信号对应频段的啸叫点；进而根据第二增益值对该第二音频输入信号进行处理，以得到音频输出信号，该第二增益值用于指示该第二啸叫点的抑制参数。从而实现了对于音频输入信号中啸叫的抑制，使得啸叫在初始阶段就被抑制，无法再进行增益循环；由于机器学习模型中指示啸叫点的对应性以及模型计算的便捷性，以及进一步的对未处理的啸叫点进行第二增益值的处理，使得上述啸叫点抑制过程迅速且全面，且不会对有效音频信号产生影响，提高了音频处理的准确性及效率。
附图说明
图1为音频信号处理***运行的网络架构图；
图2为本申请实施例提供的一种音频信号处理的流程架构图；
图3为本申请实施例提供的一种音频信号处理的方法的流程图；
图4为本申请实施例提供的另一种音频信号处理的方法的流程图；
图5为本申请实施例提供的一种音频信号处理的场景示意图；
图6为本申请实施例提供的另一种音频信号处理的场景示意图；
图7为本申请实施例提供的另一种音频信号处理的方法的流程图；
图8为本申请实施例提供的一种音频信号处理对比图；
图9为本申请实施例提供的另一种音频信号处理的方法的流程图；
图10为本申请实施例提供的一种音频信号处理方法的界面示意图；
图11为本申请实施例提供的另一种音频信号处理方法的界面示意图；
图12为本申请实施例提供的一种机器学习模型训练的方法的流程图；
图13为本申请实施例提供的一种机器学习模型训练的流程示意图；
图14为本申请实施例提供的一种音频信号处理装置的结构示意图；
图15为本申请实施例提供的一种机器学习模型训练装置的结构示意图；
图16为本申请实施例提供的一种终端设备的结构示意图；
图17为本申请实施例提供的服务器一种结构示意图。
具体实施方式
本申请实施例提供了一种音频信号处理的方法以及相关装置，可以应用于终端设备中包含音频信号处理功能的***或程序中，通过获取第一音频输入信号；然后将该第一音频输入信号输入机器学习模型，以得到处理有效音频信号频段的第一增益值；并根据该第一增益值处理该第一音频输入信号，以得到第二音频输入信号；接下来检测该第二音频输入信号，以得到第二啸叫点，该第二啸叫点用于指示非该有效音频信号对应频段的啸叫点；进而根据第二增益值对该第二音频输入信号进行处理，以得到音频输出信号，该第二增益值用于指示该第二啸叫点的抑制参数。从而实现了对于音频输入信号中啸叫的抑制，使得啸叫在初始阶段就被抑制，无法在进行增益循环；由于机器学习模型中指示啸叫点的对应性以及模型计算的便捷性，以及进一步的对未处理的啸叫点进行第二增益值的处理，使得上述啸叫点抑制过程迅速且全面，且不会对有效音频信号产生影响，提高了音频处理的准确性及效率。
首先，对本申请实施例中可能出现的一些名词进行解释。
啸叫：麦克风采集的声音信号经过扬声器放大，再被麦克风拾取，信号在反馈回路中不断的叠加放大，正反馈产生震荡循环，进而产生的现象。
啸叫点：在音频信号中循环增益大于等于1的频点。
有效音频信号：指示音频信号中的目标音频，例如语音通话过程中的语音信号。
非有效音频信号：指示音频信号中的干扰音频，例如环境噪声、回声等。
增益值：对于指定频段的音频信号的处理变化程度，在啸叫抑制场景中用于指示对于啸叫点对应音频信号的缩小倍数。
基音周期：人发声过程中声带每开启和闭合一次的周期时间，即可以用于指示有效音频信号的参数。
机器学习模型：通过给定样本进行参数调节，以使得输出具有给定样本相似特征的模型。
功率谱：信号功率随着频率的变化情况，即信号功率在频域的分布状况。
模数转换器(Analog-to-Digital Converter，ADC)：一种将模拟信号转变为数字信号的电子元件。
循环神经网络模型(Recurrent Neural Network，RNN)：一类以序列数据为输入，在序列的演进方向进行递归且所有节点(循环单元)按链式连接的递归神经网络。
卷积神经网络模型(Convolutional Neural Networks，CNN)：卷积神经网络具有表征学***移不变分类。
应理解，本申请提供的音频信号处理方法可以应用于终端设备中包含音频信号处理功能的***或程序中，例如作为游戏的语音插件，具体的，音频信号处理***可以运行于如图1所示的网络架构中，如图1所示，是音频信号处理***运行的网络架构图，如图可知，音频信号处理***可以提供与多个信息源的音频信号处理，终端通过网络建立与服务器的连接，进而接收其他终端发送的音频信号，通过对接收到的信号进行本申请提供的音频信号处理方法进行啸叫抑制，以得到音频输出，从而实现了多个终端之间的音频互动过程；可以理解的是，图1中示出了多种终端设备，在实际场景中可以有更多或更少种类的终端设备参与到音频信号处理的过程中，具体数量和种类因实际场景而定，此处不做限定，另外，图1中示出了一个服务器，但在实际场景中，也可以有多个服务器的参与，特别是在多内容应用交互的场景中，具体服务器数量因实际场景而定。
应当注意的是，本实施例提供的音频信号处理方法也可以离线进行，即不需要服务器的参与，此时终端在本地与其他终端进行音频信号互动，进而进行终端之间的音频信号处理的过程。
可以理解的是，上述音频信号处理***可以运行于个人移动终端，例如：作为游戏语音插件这样的应用，也可以运行于服务器，还可以作为运行于第三方设备以提供音频信号处理，以得到信息源的音频信号处理处理结果；具体的音频信号处理***可以是以一种程序的形式在上述设备中运行，也可以作为上述设备中的***部件进行运行，还可以作为云端服务程序的一种，具体运作模式因实际场景而定，此处不做限定。
随着移动终端相关技术的发展，越来越多的智能设备出现在人们的生活中，其中，通过智能设备进行语音通话尤为突出，然而由于通话过程中本端麦克风也会采集到对端扬声器的音频信号，这些音频信号可能在本端与对端的语音过程中循环，尤其在近距离的语音过程中，音频信号会不断的循环增益，从而产生啸叫。
一般，可以采用移频器或移相器对本端的输入音频进行处理，即破坏与啸叫产生的相位一致的音频信息，从而实现啸叫抑制。
但是，使用移频器或移相器的方法处理时间较长，不适用于语音通话的实时处理的场景，且由于对于啸叫点固定的相位移除，也会对有效音频的音质产生损伤，影响音频处理的准确性。
为了解决上述问题，本申请提出了一种音频信号处理的方法，该方法应用于图2所示的 音频信号处理的流程框架中，如图2所示，为本申请实施例提供的一种音频信号处理的流程架构图，首先终端设备收集用户的语音，并转换为音频信号，然后输入训练好的机器学习模型进行啸叫点的筛选并进行抑制，进一步的对于未处理的啸叫点进行增益控制，从而得到啸叫抑制后的音频信号以作为输出。
可以理解的是，本申请所提供的方法可以为一种程序的写入，以作为硬件***中的一种处理逻辑，也可以作为一种音频信号处理装置，采用集成或外接的方式实现上述处理逻辑。作为一种实现方式，该音频信号处理装置通过获取第一音频输入信号；然后将该第一音频输入信号输入机器学习模型，以得到处理有效音频信号频段的第一增益值；并根据该第一增益值处理该第一音频输入信号，以得到第二音频输入信号；接下来检测该第二音频输入信号，以得到第二啸叫点，该第二啸叫点用于指示非该有效音频信号对应频段的啸叫点；进而根据第二增益值对该第二音频输入信号进行处理，以得到音频输出信号，该第二增益值用于指示该第二啸叫点的抑制参数。从而实现了对于音频输入信号中啸叫的抑制，使得啸叫在初始阶段就被抑制，无法在进行增益循环；由于机器学习模型中指示啸叫点的对应性以及模型计算的便捷性，以及进一步的对未处理的啸叫点进行第二增益值的处理，使得上述啸叫点抑制过程迅速且全面，且不会对有效音频信号产生影响，提高了音频处理的准确性及效率。
结合上述流程架构，下面将对本申请中音频信号处理的方法进行介绍，请参阅图3，图3为本申请实施例提供的一种音频信号处理的方法的流程图，本申请实施例至少包括以下步骤：
301、获取第一音频输入信号。
本实施例中，第一音频输入信号可以是开始语音通话时的初始音频信号；也可以是通话一段时间后的音频信号，具体的，由于啸叫的产生为音频信号在反馈回路中不断增益的过程，即反馈增益累计的过程，其中，反馈回路即为本端麦克风与对端扬声器组成的回路；故不同时间段的音频信号可能累计的反馈增益不同，可以立即唤起本申请提供的音频处理方法，也可以等待反馈增益大于或等于1后再唤起本申请提供的音频处理方法。这是由于啸叫的产生需要在音频信号的回路中的反馈增益大于或等于1。
可选的，获取的第一音频输入信号可以是经过初步放大的信号，具体的，首先获取采集信号，该采集信号可以是由麦克风或其他采集设备采集的；然后将该采集信号转换为数字信号，例如通过ADC进行转换；进一步的将该数字信号输入放大器，从而得到该第一音频输入信号。由于对第一音频输入信号进行了放大，一方面便于用户收听，另一方面便于本申请后续的啸叫点的筛选过程。
另外，考虑到采集设备收集到的采集信号可能含有明显的杂音，例如：频率远远大于语音范围的信号；此时可以进行初步的噪声筛除。具体的，将该数字信号输入放大器，以得到放大信号；然后根据滤波参数处理该放大信号，以得到滤波后的放大信号；进一步的将该滤波后的放大信号进行傅立叶变换到频域，以得到该第一音频输入信号。其中，滤波参数可以是固定值，也可以是根据历史记录中常见噪声对应的频带进行的针对性设定。
302、将第一音频输入信号输入机器学习模型，以得到第一啸叫点，根据第一啸叫点获 得第一增益值。
本实施例中，第一啸叫点用于指示第一音频输入信号中有效音频信号对应频段的啸叫点；第一增益值用于指示第一啸叫点的抑制参数，根据第一增益值处理第一音频输入信号，以得到第二音频输入信号；另外，该机器学习模型基于多个训练信号训练所得，该训练信号中包含多个啸叫点样本，该第一啸叫点用于指示有效音频信号对应频段的啸叫点。
由于啸叫点与有效语音在频带分布或能量特征在存在一些差异，可以通过确定音频输入信号中的多个特征进行提取，这些特征可以基于有效音频信号的特征进行选择，例如：频段分布、基音周期的位置、信号波动频率等；进而输入该机器学习模型，以确定对应的第一啸叫点；然后根据该第一啸叫点确定对应的第一增益值。
具体的，特征的提取可以是基于有效音频信号进行的，一方面可以基于有效音频信号的参数特征信息，例如有效音频信号的梅尔频率倒谱系数，或基于该系数的数学变形；二方面还可以基于有效音频信号的生物特征信息，例如基音周期，这是由于人声在500Hz内的音频信号存在基音周期，而啸叫信号不存在；三方面还可以基于有效音频信号的波形特征信息，例如根据有效音频信号在特定频段内的波动情况进行判断，这是由于有效音频信号存在短时平稳的特征。通过上述特征的提取，可以很好的区别出有效音频信号和啸叫点对应的信号，使得有效音频信号的特征可以被机器学习模型进行学习，从而提高机器学习模型对于有效音频信号提取的准确度。
对于上述特征的举例仅为示意，具体的特征可以是指示有效音频信号的特征，也可以是指示啸叫信号的特征，还可以是指示有效音频信号和啸叫信号的区别特征，此处不做限定。
可选的，考虑到不同场景下输入信号的采集频率不同，可以将该第一音频输入信号调整至目标频率，以转换至频域；例如：一般手机语音通话都采用16KHz的采样率，故将目标频率调整为16KHz；然后确定转换至频域后的第一音频输入信号中的多个采样点；并基于该采样点提取多个该音频特征。从而对输入信号进行多线程的处理，提高音频处理的效率。
另外，在将第一音频输入信号由时域转换至频域的过程中，为使得时域信号更好地满足傅里叶变换过程中的周期性要求，减少信号遗漏，对于输入信号的划分可以基于窗函数进行，即基于窗函数对转换至频域后的第一音频输入信号进行划分，以得到多个子带；然后确定该子带中的多个该采样点。其中，窗函数可以是矩形窗、高斯窗或Kaiser窗等，具体的函数形式因实际场景而定。
303、检测第二音频输入信号，以得到第二啸叫点，根据第二啸叫点获得第二增益值。
本实施例中，第一增益值对应于第一音频输入信号中的多个啸叫点，而每个啸叫点对应于多个频带，这些频带的集合称为子带；故第一增益值可以包括多个啸叫抑制增益值，且每个啸叫抑制增益都是0～1的浮点数；将第一增益值输入第一音频输入信号中，其每个子带乘以对应子带的啸叫抑制衰减增益，即可得到机器学习啸叫抑制处理后的结果，即第二音频输入信号。
可以理解的是，该第二啸叫点用于指示所述第二音频输入信号中非有效音频信号对应 频段的啸叫点，即不属于该有效音频信号对应频段的啸叫点。由于机器模型训练中可能存在遗留未处理的非有效音频信号对应频段的啸叫点，故进行二次增益过程，即可以对第二啸叫点进行检测。
具体的，检测第二啸叫点可以是通过获取该第二音频输入信号对应的功率谱；然后检测该功率谱中的极值，例如：功率谱中的功率最大值，或基于功率最大值设定的取值范围；然后根据极值确定对应的候选频点，即这些频点可能是啸叫点；进而根据该候选频点确定该第二啸叫点。即检测该候选频点的相位和反馈增益信息，若相位一致且反馈增益大于等于1，则确定为第二啸叫点。通过功率谱中极值的判断，可以直观的判断出频点的增益变化情况，这是由于啸叫点对应的功率往往大于一般频点的功率，从而提高了啸叫点识别的准确性。
可选的，还可以根据峰值均值比进行第二啸叫点的判断，即获取该候选频点相邻的多个频点，以确定候选范围；然后确定该候选范围中频点的平均频率平均值，以获取峰值均值比；当该峰值均值比大于啸叫阈值时，则确定该候选频点为该第二啸叫点。为避免偶发情况造成的极值对于识别过程的影响，可以通过峰值均值比对啸叫点进行判断，从而扩展了数据参考的范围，进一步提高了啸叫点识别的准确性。
可选的，由于啸叫点存在周期性出现的行为特征，对于啸叫点的判断还可以基于历史记录进行统计从而分析得到，例如在一种可能的场景中，啸叫点容易集中在2KHz以上的频带，而语音信号的能量主要集中在2KHz以下的频带。再根据峰值均值比，判断是否是啸叫点；还可以根据历史记录中啸叫点出现的位置进行进一步的检查，例如：历史记录中记录了啸叫点集中在2KHz-3KHz，则在接下来的啸叫点识别中对该范围进行二次检测，具体的检查方式可以参考上述功率谱极值或峰值均值比的识别方式。
可以理解的是，上述啸叫点集中的频带因具体场景而定，即为不同的场景中，啸叫点集中的频带可以更高也可以更低，此处仅对历史记录进行分析以得到啸叫点的方法进行说明，并不进行限定。
304、根据第二增益值对第二音频输入信号进行处理，以得到音频输出信号。
本实施例中，该第二增益值用于指示该第二啸叫点的抑制参数，即对于第二啸叫点对应频带的缩小倍数。由于经过了第二次的啸叫点的筛选，保证了啸叫抑制的准确性，以及啸叫抑制效果的显著性。其中第二增益值可以按经验设置0～1范围内的浮点数值，也可以根据上下相邻子带的能量计算。
另外，根据第二增益值对该第二音频输入信号进行处理之后，还可以将处理后的信号转换至时域，并进行陷波处理，即滤波器的一种，以进一步的消除啸叫点。
结合上述实施例可知，通过获取第一音频输入信号；然后将该第一音频输入信号输入机器学习模型，以得到处理有效音频信号频段的第一增益值；并根据该第一增益值处理该第一音频输入信号，以得到第二音频输入信号；接下来检测该第二音频输入信号，以得到第二啸叫点，该第二啸叫点用于指示所述第二音频输入信号中非有效音频信号对应频段的啸叫点；进而根据第二增益值对该第二音频输入信号进行处理，以得到音频输出信号，该第二增益值用于指示该第二啸叫点的抑制参数。从而实现了对于音频输入信号中啸叫的抑 制，使得啸叫在初始阶段就被抑制，无法再进行增益循环；由于机器学习模型中指示啸叫点的对应性以及模型计算的便捷性，以及进一步的对未处理的啸叫点进行第二增益值的处理，使得上述啸叫点抑制过程迅速且全面，且不会对有效音频信号产生影响，提高了音频处理的准确性及效率。
上述实施例介绍了一种音频信号处理的过程，但是，在第二次啸叫抑制中可能对有效音频信号产生影响，为避免该情况的发生，请参阅图4，图4为本申请实施例提供的另一种音频信号处理的方法的流程图，本申请实施例至少包括以下步骤：
401、获取第一音频输入信号。
402、将所述第一音频输入信号输入机器学习模型，以得到第一增益值。
403、根据所述第一增益值处理所述第一音频输入信号，以得到第二音频输入信号。
404、检测所述第二音频输入信号，以得到第二啸叫点。
本实施例中，步骤401-404与图3指示的实施例步骤301-304相似，相关特征描述可以进行参考，此处不做赘述。
405、检测所述第二音频输入信号，以进行语音保护。
本实施例中，语音保护即保证有效音频信号的完整性。具体的，首先获取有效音频信号中的特征信息，该特征信息基于该有效音频信号指示的波形特征确定，例如：有效音频信号指示的波形特征中浊音有共振峰，另外清音高频能量大且按频率轴能量斜率稳定；然后根据该特征信息检测该第二音频输入信号中对应的有效音频信号；进一步的对该有效音频信号进行锁定操作，该锁定操作用于指示该第二增益值的非作用对象，即第二增益值对应的处理频带中可能包含了本步骤加锁的有效音频信号，但对这些频带的信号不进行增益处理。
可选的，对于有效音频信号的语音保护还可以基于用于指示语音频带的历史记录进行，即统计有效音频信号的频带分布，对于分布权重大的频段进行逐一检测筛选。
406、根据第二增益值对所述第二音频输入信号进行处理。
407、对根据第二增益值处理后的帧进行平滑处理。
本实施例中，为防止帧间抑制增益差值过大导致音频输出信号听起来刺耳，即音频变化突兀，可以对第二增益值对应啸叫点对应的多个增益帧进行平滑处理。具体的，可以采用如下公式对增益帧以及相邻的帧进行处理：
Figure PCTCN2020124244-appb-000001
其中α为0～1的平滑因子；
Figure PCTCN2020124244-appb-000002
为上一帧抑制增益；m为帧索引；k为频点索引。上述公式通过调整相邻帧间的增益差值，使得相邻帧间的增益更加接近线性分布，减少了音频变化突兀的情况，使得音频输出信号在听觉感官上更加平滑，提高用户体验。
408、获取音频输出信号。
本实施例中，通过上述步骤407中的增益参数
Figure PCTCN2020124244-appb-000003
乘以对应频点的值，即得 到音频输出信号。
结合上述实施例可见，通过对于有效音频信号的检测并加锁保护，提高了音频输出信号的准确性以及清晰度；另外，通过平滑相邻增益帧的增益参数，使得相邻帧间的增益更加接近线性分布，减少了音频变化突兀的情况，使得音频输出信号在听觉感官上更加平滑，提高了用户体验。
上述实施例介绍了啸叫抑制的音频处理方法，下面结合具体的场景对于啸叫抑制的音频处理方法进行说明，如图5所示，是本申请实施例提供的一种音频信号处理的场景示意图。图中示出了麦克风收集语音信号并放大播放的场景；由于声源(麦克风)与扩音设备(扬声器)距离太近，麦克风采集的声音信号经过扬声器放大，再被麦克风拾取，信号在反馈回路中不断的叠加放大，正反馈产生震荡循环，进而产生啸叫。其中，正反馈产生震荡的函数可以是：
Figure PCTCN2020124244-appb-000004
对应的，啸叫产生的条件需要反馈回路中麦克风采集的输入信号的相位与反馈到扬声器中的声波信号的相位相同，即：
∠G(ω 0)F(ω 0)＝n*2π
且反馈回路增益大于等于1，即
|G(ω 0)F(ω 0)|≥1
在上述公式中，G(s)为麦克风采集的输入信号；F(s)为反馈到扬声器中的声波信号；G(w0)为麦克风采集的输入信号的相位；F(w0)为反馈到扬声器中的声波信号的相位；n为整数参数。
在该场景中，可以在放大器中执行本申请提供的音频信号处理的方法，即通过麦克风采集的音频信号传输到放大器后，立即进行上述图3或图4所示实施例的音频信号处理过程，然后输出信号再传输至扬声器播放，如此循环，即可达到啸叫抑制的效果。
在另一种可能的场景中，如图6所示，是本申请实施例提供的另一种音频信号处理的场景示意图，图中示出了终端外放场景下的一条回路。当两部终端距离比较近的时候，右边终端扬声器声音出来，被左边终端麦克风拾取。经过前处理和信号转换，通过网络发到右边终端。经过扬声器播放出来，再被左边终端麦克风拾取。如此不断循环，如果环路在某个频点增益大于等于1，且相位是正向的，那么这一点就会形成啸叫点。
下面结合一种具体的示例对啸叫抑制进行说明，请参阅图7，图7为本申请实施例提供的另一种音频信号处理的方法的流程图，本申请实施例至少包括以下步骤：
701、输入目标频率的音频信号，并分为20毫秒每帧。
本实施例中，考虑到手机语音通话一般是16KHz采样率处理，可以设置目标频率为16KHz。
702、转换到频域。
本实施例中，将音频信号变换到频域，并加窗做傅里叶变换到频域，窗函数可以是矩形窗、高斯窗或Kaiser窗等，具体的函数形式因实际场景而定。
703、提取42个特征值。
本实施例中，特征值可以包括22个梅尔频率倒谱系数(Mel-scale Frequency Cepstral Coefficients，MFCC)，该系数可以参考语音识别过程中的参数，即有效音频信号；特征值还可以包括前6个系数的一阶或二阶导数，用于指示语音特征；特征值还可以包括基因周期，这是由于语音信号的浊音在500Hz以内有基因周期，而啸叫信号没有；特征值还可以包括非平稳特征值的检测，这是由于语音是短时平稳的。
704、通过循环神经网络模型计算第一增益值。
本实施例中，机器学习模型采用循环神经网络模型，这是为了对时间序列建模，而不是仅仅考虑输入和输出帧。具体的第一增益值的获取过程与图3所述实施例的步骤302类似，此处不做赘述。
705、根据第一增益值对第一啸叫点进行啸叫抑制。
706、检测第二啸叫点，并获取第二增益值。
707、根据第二增益值进行啸叫抑制。
708、转换入时域并输出音频信号。
本实施例中，步骤705-708与图3所示实施例的步骤303-305相似，相关特征描述可以进行参考，此处不做赘述。
通过上述实施例，可以得到如图8所示的啸叫抑制结果，图8为本申请实施例提供的一种音频信号处理对比图；上图为啸叫抑制前输入信号的语谱图，下为啸叫抑制处理后信号的语谱图。对比可以看到样本的波峰周围的杂峰明显的减弱了，即在啸叫起来之前，经过本申请提供的音频处理方法已经将啸叫进行了抑制。
上述实施例介绍了音频信号处理的过程，下面，结合游戏应用作为具体场景进行介绍，请参阅图9，图9为本申请实施例提供的另一种音频信号处理的方法的流程图，本申请实施例至少包括以下步骤：
901、获取游戏启动指令。
本实施例中，游戏的启动指令可以是游戏开始运行，或者游戏中某一特定场景线程的触发，例如：进入战斗场景。
902、若特征元素被触发，则进行啸叫抑制。
本实施例中，特征元素为启动语音通话功能的实体或虚拟按钮，如图10所示，是本申请实施例提供的一种音频信号处理方法的界面示意图；图中示出了游戏界面中的特征元素A1，当其中任意按钮被触发时，即唤起上述图3或图4实施例所述的音频处理的方法。
另外，本申请中的音频处理的方法不仅仅用于两个用户的语音通话过程中，还可以应用于多个用户的语音通话过程中；如图11所示，是本申请实施例提供的另一种音频信号处理方法的界面示意图，图中用户处于公共语音场景中B2，此时，若特征元素B1被触发，即唤起上述图3或图4实施例所述的音频处理的方法。
903、输出处理后的音频信号。
本实施例中，通过上述啸叫抑制后的音频信号进行输入，以实现两个或多个用户之间的清晰的语音通话过程。
通过对于游戏过程中用户之间的音频信号进行啸叫抑制，使得用户可以更加清晰的进行语音通话，不会产生因啸叫影响沟通的情况，保证了在游戏这种需要高效高质量的语音场景中的用户体验以及语音通话的准确性。
上述实施例中还涉及了机器学习模型的应用，该机器学习模型是经过预先训练后的模型；该方法可以通过音频处理设备执行，该音频处理设备可以是终端设备，也可以是服务器。训练好的机器学习模型可以应用于前述的音频信号处理方案中。下面，对场景进行介绍，请参阅图12，图12为本申请实施例提供的一种机器学习模型训练的方法的流程图，本申请实施例至少包括以下步骤：
1201、采集参考信号和语音样本信号。
本实施例中，所述参考信号为基于至少两种变量元素确定的啸叫信号，所述变量元素包括程序类别、程序运行时段或程序运行位置，所述语音样本信号用于指示语音通话过程中的有效语音。
可以理解的是，变量元素中的程序类别可以是不同的游戏，例如：王者荣耀、和平精英等不同游戏场景下的训练样本。而程序运行时段则指示的是采集训练样本时的时段，例如游戏一般在晚上8点至9点这一时间段进行语音通话功能，且通话语音较为激烈，可以进行额外的标注并生成训练样本。另外，程序运行位置即语音采集的地理信息，例如：训练样本采集于市场、教师或卧室等不同的地理位置。
通过对于上述多种不同条件下的训练样本的获取，并标记训练样本里的啸叫点，从而保证了训练样本的泛化能力；由于采集信号作为语音样本的参与，使得该机器学习模型对于语音频段的啸叫点具有良好的识别能力。
1202、根据所述参考信号和采集信号生成特征训练集。
本实施例中，基于上述不同因素下采集的信号设定对应的标签，并分类；且标注对应的啸叫点以生成特征训练集。
1203、将所述特征训练集输入机器学习模型进行至少一个循环的训练，以得到训练后的机器学习模型。
本实施例中，所述训练后的机器学习模型用于根据音频输入信号确定对应的啸叫点以及增益值。具体的，如图13所示，是本申请实施例提供的一种机器学习模型训练的流程示 意图，图中示出了一种RNN模型，其包括有3层门循环控制单元(gated recurrence unit，GRU)。与简单的循环单元相比，GRU有两个额外的门；其中，复位门决定是否将当前状态记忆，以用于计算新状态；而更新门决定当前状态将根据新输入改变多少。当更新门关闭时，可以使得GRU长时间地记住训练信息。首先第一层GRU输入42维，输出24维和一个语音活动检测(voice activity detection，VAD)标志。第二层GRU输入初始的42维特征和第一层输出的24维特征，以输出48维，用来估计啸叫信号。第三层输入初始的42维特征和第二层输出的42维特征，以得到输出；并根据训练样本中的增益值对输出进行调整以更新模型参数，从而实现RNN模型的训练。
应当注意的是，本申请中的训练过程也可以应用于深度神经网络模型或卷积神经网络模型中，此处不做赘述。
通过上述机器学习模型的训练过程，使得音频信号在输入机器学习模型后可以得到啸叫点分布以及对应的第一增益值，从而保证了语音频段的啸叫抑制的准确性。
为了更好的实施本申请实施例的上述方案，下面还提供用于实施上述方案的相关装置。请参阅图14，图14为本申请实施例提供的一种音频信号处理装置的结构示意图，音频信号处理装置1400包括：
获取单元1401，用于获取第一音频输入信号；
输入单元1402，用于将所述第一音频输入信号输入机器学习模型，以得到第一增益值，其中，所述第一增益值用于指示所述第一音频信号中第一啸叫点的抑制参数，所述第一啸叫点用于指示有效音频信号对应频段的啸叫点；
检测单元1403，用于根据所述第一增益值处理所述第一音频输入信号，以得到第二音频输入信号；
处理单元1404，用于根据第二增益值对第二音频输入信号进行处理，以得到音频输出信号，所述第二增益值用于指示第二啸叫点的抑制参数，所述第二啸叫点用于指示非所述有效音频信号对应频段的啸叫点。
可选的，在本申请一些可能的实现方式中，所述输入单元1402，具体用于将所述音频输入信号转换至频域，以提取多个音频特征，所述音频特征基于所述有效音频信号或所述啸叫样本的特征确定；
所述输入单元1402，具体用于将所述音频特征输入所述机器学习模型，以确定所述第一啸叫点；
所述输入单元1402，具体用于根据所述第一啸叫点确定对应的第一增益值。
所述输入单元1402，具体用于根据所述第一增益值处理所述第一音频输入信号，以得到第二音频输入信号；
所述检测单元1403，用于检测所述第二音频输入信号，以得到第二啸叫点，根据所述第二啸叫点获得第二增益值，所述第二啸叫点用于指示所述第二音频输入信号中非有效音 频信号对应频段的啸叫点；
所述处理单元1404，用于根据第二增益值对第二音频输入信号进行处理，以得到音频输出信号。
可选的，所述输入单元1402，具体用于将所述音频输入信号转换至频域，以提取多个音频特征，所述音频特征基于所述有效音频信号或所述啸叫样本的特征确定；
所述输入单元1402，具体用于将所述音频特征输入所述机器学习模型，以确定所述第一啸叫点；
所述输入单元1402，具体用于根据所述第一啸叫点获得对应的第一增益值。
可选的，在本申请一些可能的实现方式中，所述输入单元1402，具体用于将所述第一音频输入信号调整至目标频率，以转换至频域；
所述输入单元1402，具体用于确定转换至频域后的第一音频输入信号中的多个采样点；
所述输入单元1402，具体用于基于所述采样点提取多个所述音频特征。
可选的，在本申请一些可能的实现方式中，所述输入单元1402，具体用于基于窗函数对转换至频域后的第一音频输入信号进行划分，以得到多个子带；
所述输入单元1402，具体用于确定所述子带中的多个所述采样点。
可选的，在本申请一些可能的实现方式中，所述检测单元1403，具体用于获取所述第二音频输入信号对应的功率谱；
所述检测单元1403，具体用于检测所述功率谱中的极值，并确定对应的候选频点；
所述检测单元1403，具体用于根据所述候选频点确定所述第二啸叫点；
所述检测单元1403，具体用于根据所述第二增益值对所述第二啸叫点进行处理，以得到所述音频输出信号。
可选的，在本申请一些可能的实现方式中，所述检测单元1403，具体用于获取所述候选频点相邻的多个频点，以确定候选范围；
所述检测单元1403，具体用于确定所述候选范围中频点的平均频率平均值，以获取峰值均值比；
所述检测单元1403，具体用于若所述峰值均值比大于啸叫阈值，则确定所述候选频点为所述第二啸叫点。
可选的，在本申请一些可能的实现方式中，所述检测单元1403，还用于获取有效音频信号中的特征信息，所述特征信息基于所述有效音频信号指示的波形特征确定，所述有效音频信号用于指示语音样本；
所述检测单元1403，具体用于根据所述特征信息检测所述第二音频输入信号中对应的有效音频信号；
所述检测单元1403，具体用于对所述有效音频信号进行锁定操作，所述锁定操作用于指示所述第二增益值的非作用对象。
可选的，在本申请一些可能的实现方式中，所述处理单元1404，具体用于确定所述第二啸叫点对应的多个增益帧；
所述处理单元1404，具体用于根据平滑公式对所述增益帧进行处理，以对所述音频输出信号进行更新。
可选的，在本申请一些可能的实现方式中，所述获取单元1401，具体用于获取采集信号；
所述获取单元1401，具体用于将所述采集信号转换为数字信号；
所述获取单元1401，具体用于将所述数字信号输入放大器，以得到所述第一音频输入信号。
可选的，在本申请一些可能的实现方式中，所述获取单元1401，具体用于将所述数字信号输入放大器，以得到放大信号；
所述获取单元1401，具体用于根据滤波参数处理所述放大信号，以得到滤波后的放大信号；
所述获取单元1401，具体用于将所述滤波后的放大信号进行傅立叶变换到频域，以得到所述第一音频输入信号。
可选的，在本申请一些可能的实现方式中，所述音频信号的处理方法应用于游戏语音通话过程中，所述获取单元1401，具体用于检测特征元素的触发情况，所述特征元素为游戏界面中的元素；
所述获取单元1401，具体用于若所述特征元素被触发，则获取所述第一音频输入。
通过获取第一音频输入信号；然后将该第一音频输入信号输入机器学习模型，以得到处理有效音频信号频段的第一增益值；并根据该第一增益值处理该第一音频输入信号，以得到第二音频输入信号；接下来检测该第二音频输入信号，以得到第二啸叫点，该第二啸叫点用于指示第二音频收入信号中非有效音频信号对应频段的啸叫点；进而根据第二增益值对该第二音频输入信号进行处理，以得到音频输出信号，该第二增益值用于指示该第二啸叫点的抑制参数。从而实现了对于音频输入信号中啸叫的抑制，使得啸叫在初始阶段就被抑制，无法再进行增益循环；由于机器学习模型中指示啸叫点的对应性以及模型计算的便捷性，以及进一步的对未处理的啸叫点进行第二增益值的处理，使得上述啸叫点抑制过程迅速且全面，且不会对有效音频信号产生影响，提高了音频处理的准确性及效率。
本申请还提供一种机器学习模型训练的装置1500，如图15所示，是本申请实施例提供的一种机器学习模型训练装置的结构示意图，包括：采集单元1501，用于采集参考信号和语音样本信号，所述参考信号为基于至少两种变量元素确定的啸叫信号，所述变量元素包括程序类别、程序运行时段或程序运行位置，所述采集信号用于指示通话过程中的有效语音；
生成单元1502，用于根据所述参考信号和采集信号生成特征训练集；
训练单元1503，用于将所述特征训练集输入机器学习模型进行至少一个循环的训练，以得到训练后的机器学习模型，所述训练后的机器学习模型用于根据音频输入信号确定对应的啸叫点以及增益值。
本申请实施例还提供了一种终端设备，如图16所示，是本申请实施例提供的另一种终端设备的结构示意图，为了便于说明，仅示出了与本申请实施例相关的部分，具体技术细节未揭示的，请参照本申请实施例方法部分。该终端可以为包括手机、平板电脑、个人数字助理(personal digital assistant，PDA)、销售终端(point of sales，POS)、车载电脑等任意终端设备，以终端为手机为例：
图16示出的是与本申请实施例提供的终端相关的手机的部分结构的框图。参考图16，手机包括：射频(radio frequency，RF)电路1610、存储器1620、输入单元1630、显示单元1640、传感器1650、音频电路1660、无线保真(wireless fidelity，WiFi)模块1670、处理器1680、以及电源1690等部件。本领域技术人员可以理解，图16中示出的手机结构并不构成对手机的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件布置。
下面结合图16对手机的各个构成部件进行具体的介绍：
RF电路1610可用于收发信息或通话过程中，信号的接收和发送，特别地，将基站的下行信息接收后，给处理器1680处理；另外，将设计上行的数据发送给基站。通常，RF电路1610包括但不限于天线、至少一个放大器、收发信机、耦合器、低噪声放大器(low noise amplifier，LNA)、双工器等。此外，RF电路1610还可以通过无线通信与网络和其他设备通信。
存储器1620可用于存储软件程序以及模块，处理器1680通过运行存储在存储器1620的软件程序以及模块，从而执行手机的各种功能应用以及数据处理。
输入单元1630可用于接收输入的数字或字符信息，以及产生与手机的用户设置以及功能控制有关的键信号输入。具体地，输入单元1630可包括触控面板1631以及其他输入设备1632。
显示单元1640可用于显示由用户输入的信息或提供给用户的信息以及手机的各种菜单。
手机还可包括至少一种传感器1650，比如光传感器、运动传感器以及其他传感器。。
音频电路1660、扬声器1661，传声器1662可提供用户与手机之间的音频接口。音频电路1660可将接收到的音频数据转换后的电信号，传输到扬声器1661，由扬声器1661转换为声音信号输出；另一方面，传声器1662将收集的声音信号转换为电信号，由音频电路1660接收后转换为音频数据，再将音频数据输出处理器1680处理后，经RF电路1610以发送给比如另一手机，或者将音频数据输出至存储器1620以便进一步处理。
WiFi属于短距离无线传输技术，虽然图16示出了WiFi模块1670，但是可以理解的是，其并不属于手机的必须构成，完全可以根据需要在不改变发明的本质的范围内而省略。
处理器1680是手机的控制中心，利用各种接口和线路连接整个手机的各个部分，通过 运行或执行存储在存储器1620内的软件程序和/或模块，以及调用存储在存储器1620内的数据，执行手机的各种功能和处理数据，从而对手机进行整体监控。
在本申请实施例中，该终端所包括的处理器1680还具有执行如上述音频信号处理方法或模型训练方法的各个步骤的功能。
本申请实施例还提供了一种服务器，请参阅图17，图17是本申请实施例提供的服务器一种结构示意图，该服务器1700可因配置或性能不同而产生比较大的差异，可以包括一个或一个以***处理器(central processing units，CPU)1722和存储器1732，一个或一个以上存储应用程序1742或数据1744的存储介质1730。
服务器1700还可以包括一个或一个以上电源1726，一个或一个以上有线或无线网络接口1750，一个或一个以上输入输出接口1758，和/或，一个或一个以上操作***1741。
上述实施例中由模型训练装置所执行的步骤可以基于该图17所示的服务器结构。
本申请实施例中还提供一种计算机可读存储介质，该计算机可读存储介质中存储有计算机程序，所述计算机程序用于执行如前述图2至图13所示实施例描述的方法中音频信号处理装置所执行的步骤。
本申请实施例中还提供一种包括音频信号处理指令的计算机程序产品，当其在计算机上运行时，使得计算机执行如前述图2至图13所示实施例描述的方法中音频信号处理装置所执行的步骤。
本申请实施例还提供了一种音频信号处理***，所述音频信号处理***可以包含图14所描述实施例中的音频信号处理装置，或者图16所描述的终端设备。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的***，装置和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的***，装置和方法，可以通过其它的方式实现。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，音频信号处理装置，或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。
以上所述，以上实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围。

Claims (16)

  1. 一种音频信号的处理方法，所述方法由终端设备执行，所述方法包括：
    获取第一音频输入信号；
    将所述第一音频输入信号输入机器学习模型，以得到第一啸叫点，根据所述第一啸叫点获得第一增益值，其中，所述第一啸叫点用于指示所述第一音频输入信号中有效音频信号对应频段的啸叫点；所述第一增益值用于指示所述第一啸叫点的抑制参数；
    根据所述第一增益值处理所述第一音频输入信号，以得到第二音频输入信号；
    检测所述第二音频输入信号，以得到第二啸叫点，根据所述第二啸叫点获得第二增益值，所述第二啸叫点用于指示所述第二音频输入信号中非有效音频信号对应频段的啸叫点；
    根据所述第二增益值对所述第二音频输入信号进行处理，以得到音频输出信号。
  2. 根据权利要求1所述的方法，所述将所述第一音频输入信号输入机器学习模型，以得到第一啸叫点，根据所述第一啸叫点获得第一增益值，包括：
    将所述第一音频输入信号转换至频域，以提取多个音频特征，所述音频特征基于所述有效音频信号的特征确定；
    将所述第一音频特征输入所述机器学习模型，以确定所述第一啸叫点；
    根据所述第一啸叫点获得对应的第一增益值。
  3. 根据权利要求2所述的方法，所述将所述第一音频输入信号转换至频域，以提取多个音频特征，包括：
    将所述第一音频输入信号调整至目标频率，以转换至频域；
    确定转换至频域后的第一音频输入信号中的多个采样点；
    基于所述采样点提取多个所述音频特征。
  4. 根据权利要求3所述的方法，所述确定转换至频域后的第一音频输入信号中的多个采样点，包括：
    基于窗函数对转换至频域后的第一音频输入信号进行划分，以得到多个子带；
    确定所述子带中的多个所述采样点。
  5. 根据权利要求1所述的方法，所述根据所述第二增益值对所述第二音频输入信号进行处理，以得到音频输出信号，包括：
    获取所述第二音频输入信号对应的功率谱；
    检测所述功率谱中的极值，并确定对应的候选频点；
    根据所述候选频点确定所述第二啸叫点；
    根据所述第二增益值对所述第二啸叫点进行处理，以得到所述音频输出信号。
  6. 根据权利要求5所述的方法，所述根据所述候选频点确定所述第二啸叫点，包括：
    获取所述候选频点相邻的多个频点，以确定候选范围；
    确定所述候选范围中频点的平均频率平均值，以获取峰值均值比；
    若所述峰值均值比大于啸叫阈值，则确定所述候选频点为所述第二啸叫点。
  7. 根据权利要求5所述的方法，在所述根据所述候选频点确定所述第二啸叫点之后，所述方法还包括：
    获取所述有效音频信号中的特征信息，所述特征信息基于所述有效音频信号指示的波形特征确定；
    根据所述特征信息检测所述第二音频输入信号中对应的有效音频信号；
    对所述有效音频信号进行锁定操作，所述锁定操作用于指示所述第二增益值的非作用对象。
  8. 根据权利要求5所述的方法，所述方法还包括：
    确定所述第二啸叫点对应的多个增益帧；
    根据平滑公式对所述增益帧进行处理，以对所述音频输出信号进行更新。
  9. 根据权利要求1-8任一项所述的方法，所述音频信号的处理方法应用于游戏语音通话过程中，所述获取第一音频输入信号，包括：
    检测特征元素的触发情况，所述特征元素为游戏界面中的元素；
    若所述特征元素被触发，则获取所述第一音频输入信号。
  10. 根据权利要求1-8任一项所述的方法，所述机器学习模型为循环神经网络模型，所述第一音频输入信号和所述音频输出信号应用于所述终端设备的语音通话过程中。
  11. 一种机器学习模型的训练方法，所述方法由音频处理设备执行，所述方法包括：
    采集参考信号和语音样本信号，所述参考信号为基于至少两种变量元素确定的啸叫信号，所述变量元素包括程序类别、程序运行时段或程序运行位置，所述采集信号用于指示通话过程中的有效语音；
    根据所述参考信号和采集信号生成特征训练集；
    将所述特征训练集输入机器学习模型进行至少一个循环的训练，以得到训练后的机器学习模型，所述训练后的机器学习模型用于根据音频输入信号确定对应的啸叫点以及增益值。
  12. 一种音频信号的处理装置，包括：
    获取单元，用于获取第一音频输入信号；
    输入单元，用于将所述第一音频输入信号输入机器学习模型，以得到第一啸叫点，根据所述第一啸叫点获得第一增益值，其中，所述第一啸叫点用于指示所述第一音频输入信号中有效音频信号对应频段的啸叫点；所述第一增益值用于指示所述第一啸叫点的抑制参数，根据所述第一增益值处理所述第一音频输入信号，以得到第二音频输入信号；
    检测单元，用于检测所述第二音频输入信号，以得到第二啸叫点，根据所述第二啸叫点获得第二增益值，所述第二啸叫点用于指示所述第二音频输入信号中非有效音频信号对 应频段的啸叫点；
    处理单元，用于根据第二增益值对第二音频输入信号进行处理，以得到音频输出信号。
  13. 一种机器学习模型的训练装置，包括：
    采集单元，用于采集参考信号和语音样本信号，所述参考信号为基于至少两种变量元素确定的啸叫信号，所述变量元素包括程序类别、程序运行时段或程序运行位置，所述采集信号用于指示通话过程中的有效语音；
    生成单元，用于根据所述参考信号和采集信号生成特征训练集；
    训练单元，用于将所述特征训练集输入机器学习模型进行至少一个循环的训练，以得到训练后的机器学习模型，所述训练后的机器学习模型用于根据音频输入信号确定对应的啸叫点以及增益值。
  14. 一种计算机设备，所述计算机设备包括处理器以及存储器：
    所述存储器用于存储程序代码；所述处理器用于根据所述程序代码中的指令执行权利要求1至10任一项所述的音频信号处理的方法，或权利要求11所述的机器学习模型的训练方法。
  15. 一种计算机可读存储介质，所述计算机可读存储介质中存储有计算机程序，所述计算机程序用于执行权利要求1至10任一项所述的音频信号的处理方法，或权利要求11所述的机器学习模型的训练方法。
  16. 一种包括指令的计算机程序产品，当其在计算机上运行时，使得所述计算机执行权利要求1至10任一项所述的音频信号的处理方法，或权利要求11所述的机器学习模型的训练方法。
PCT/CN2020/124244 2020-01-09 2020-10-28 一种音频信号处理方法、模型训练方法以及相关装置 WO2021139327A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP20911413.1A EP3998557B1 (en) 2020-01-09 2020-10-28 Audio signal processing method and related apparatus
US17/700,862 US20220215853A1 (en) 2020-01-09 2022-03-22 Audio signal processing method, model training method, and related apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010023045.6A CN111210021B (zh) 2020-01-09 2020-01-09 一种音频信号处理方法、模型训练方法以及相关装置
CN202010023045.6 2020-01-09

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/700,862 Continuation US20220215853A1 (en) 2020-01-09 2022-03-22 Audio signal processing method, model training method, and related apparatus

Publications (1)

Publication Number Publication Date
WO2021139327A1 true WO2021139327A1 (zh) 2021-07-15

Family

ID=70789591

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/124244 WO2021139327A1 (zh) 2020-01-09 2020-10-28 一种音频信号处理方法、模型训练方法以及相关装置

Country Status (4)

Country Link
US (1) US20220215853A1 (zh)
EP (1) EP3998557B1 (zh)
CN (1) CN111210021B (zh)
WO (1) WO2021139327A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116233697A (zh) * 2022-04-19 2023-06-06 中国科学院声学研究所 一种基于深度学习的声反馈的抑制方法及***

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111210021B (zh) * 2020-01-09 2023-04-14 腾讯科技（深圳）有限公司 一种音频信号处理方法、模型训练方法以及相关装置
CN111768795A (zh) * 2020-07-09 2020-10-13 腾讯科技（深圳）有限公司 语音信号的噪声抑制方法、装置、设备及存储介质
CN112309417B (zh) * 2020-10-22 2023-07-07 瓴盛科技有限公司 风噪抑制的音频信号处理方法、装置、***和可读介质
CN112669868B (zh) * 2020-12-23 2021-11-26 广州智讯通信***有限公司 基于指挥调度***的语音啸叫抑制方法、设备、存储介质
CN113473316B (zh) * 2021-06-30 2023-01-31 苏州科达科技股份有限公司 音频信号处理方法、装置及存储介质
CN113891217B (zh) * 2021-11-08 2024-05-31 易兆微电子(杭州)股份有限公司 啸叫抑制方法、装置、电子设备及存储介质
CN113921023B (zh) * 2021-12-14 2022-04-08 北京百瑞互联技术有限公司 一种蓝牙音频啸叫抑制方法、装置、介质及蓝牙设备
CN115223584B (zh) * 2022-09-19 2022-12-09 腾讯科技（深圳）有限公司 音频数据处理方法、装置、设备及存储介质
CN115567864B (zh) * 2022-12-02 2024-03-01 浙江华创视讯科技有限公司 麦克风增益的调整方法和装置、存储介质及电子设备
CN117316175B (zh) * 2023-11-28 2024-01-30 山东放牛班动漫有限公司 一种动漫数据智能编码存储方法及***
CN117351925B (zh) * 2023-12-06 2024-04-12 科大讯飞(苏州)科技有限公司 啸叫抑制方法、装置、电子设备和存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108449493A (zh) * 2018-03-12 2018-08-24 广东欧珀移动通信有限公司 语音通话数据处理方法、装置、存储介质及移动终端
CN109637552A (zh) * 2018-11-29 2019-04-16 河北远东通信***工程有限公司 一种抑制音频设备啸叫的语音处理方法
CN110536215A (zh) * 2019-09-09 2019-12-03 普联技术有限公司 音频信号处理的方法、装置、计算设置及存储介质
CN111210021A (zh) * 2020-01-09 2020-05-29 腾讯科技（深圳）有限公司 一种音频信号处理方法、模型训练方法以及相关装置

Family Cites Families (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2728120B2 (ja) * 1995-01-24 1998-03-18 日本電気株式会社 エコーキャンセラ装置
SG98435A1 (en) * 2000-12-08 2003-09-19 Nanyang Polytechnic A method for detecting and removing howling
JP4681163B2 (ja) * 2001-07-16 2011-05-11 パナソニック株式会社 ハウリング検出抑圧装置、これを備えた音響装置、及び、ハウリング検出抑圧方法
US8144895B2 (en) * 2005-05-18 2012-03-27 Panasonic Corporation Howling control apparatus and acoustic apparatus
KR20150043858A (ko) * 2013-10-15 2015-04-23 한국전자통신연구원 하울링 제거 장치 및 방법
CN105872910A (zh) * 2016-03-23 2016-08-17 成都普创通信技术股份有限公司 一种音频信号啸叫检测方法
EP3301675B1 (en) * 2016-09-28 2019-08-21 Panasonic Intellectual Property Corporation of America Parameter prediction device and parameter prediction method for acoustic signal processing
EP3474280B1 (en) * 2017-10-19 2021-07-07 Goodix Technology (HK) Company Limited Signal processor for speech signal enhancement
CN109121057B (zh) * 2018-08-30 2020-11-06 北京聆通科技有限公司 一种智能助听的方法及其***
CN109461455B (zh) * 2018-11-30 2022-08-30 维沃移动通信（深圳）有限公司 一种消除啸叫的***及方法
CN109788400B (zh) * 2019-03-06 2020-12-18 哈尔滨工业大学（深圳） 一种用于数字助听器的神经网络啸叫抑制方法、***及存储介质
CN110234051B (zh) * 2019-06-27 2021-08-27 中科上声（苏州）电子有限公司 一种基于深度学习的防啸叫扩声方法及***

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108449493A (zh) * 2018-03-12 2018-08-24 广东欧珀移动通信有限公司 语音通话数据处理方法、装置、存储介质及移动终端
CN109637552A (zh) * 2018-11-29 2019-04-16 河北远东通信***工程有限公司 一种抑制音频设备啸叫的语音处理方法
CN110536215A (zh) * 2019-09-09 2019-12-03 普联技术有限公司 音频信号处理的方法、装置、计算设置及存储介质
CN111210021A (zh) * 2020-01-09 2020-05-29 腾讯科技（深圳）有限公司 一种音频信号处理方法、模型训练方法以及相关装置

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116233697A (zh) * 2022-04-19 2023-06-06 中国科学院声学研究所 一种基于深度学习的声反馈的抑制方法及***
CN116233697B (zh) * 2022-04-19 2023-09-05 中国科学院声学研究所 一种基于深度学习的声反馈的抑制方法及***

Also Published As

Publication number Publication date
EP3998557A4 (en) 2022-10-26
CN111210021B (zh) 2023-04-14
EP3998557A1 (en) 2022-05-18
US20220215853A1 (en) 2022-07-07
EP3998557B1 (en) 2023-08-16
CN111210021A (zh) 2020-05-29

Similar Documents

Publication Publication Date Title
WO2021139327A1 (zh) 一种音频信号处理方法、模型训练方法以及相关装置
US10504539B2 (en) Voice activity detection systems and methods
CN107799126B (zh) 基于有监督机器学习的语音端点检测方法及装置
CN109087669B (zh) 音频相似度检测方法、装置、存储介质及计算机设备
CN110310623B (zh) 样本生成方法、模型训练方法、装置、介质及电子设备
US20200227071A1 (en) Analysing speech signals
CN104168353B (zh) 蓝牙耳机及其语音交互控制方法
WO2020181824A1 (zh) 声纹识别方法、装置、设备以及计算机可读存储介质
CN106782504A (zh) 语音识别方法和装置
JP2020115206A (ja) システム及び方法
WO2023040523A1 (zh) 音频信号的处理方法、装置、电子设备及存储介质
CN112185410B (zh) 音频处理方法及装置
CN114338623B (zh) 音频的处理方法、装置、设备及介质
CN117294985B (zh) 一种tws蓝牙耳机控制方法
US20240194220A1 (en) Position detection method, apparatus, electronic device and computer readable storage medium
CN111722696B (zh) 用于低功耗设备的语音数据处理方法和装置
CN117059068A (zh) 语音处理方法、装置、存储介质及计算机设备
JP6268916B2 (ja) 異常会話検出装置、異常会話検出方法及び異常会話検出用コンピュータプログラム
CN116612778B (zh) 回声及噪声抑制方法、相关装置和介质
WO2024017110A1 (zh) 语音降噪方法、模型训练方法、装置、设备、介质及产品
US20230186943A1 (en) Voice activity detection method and apparatus, and storage medium
CN109920433A (zh) 嘈杂环境下电子设备的语音唤醒方法
Zhu et al. A robust and lightweight voice activity detection algorithm for speech enhancement at low signal-to-noise ratio
CN112562717A (zh) 啸叫检测方法、装置、存储介质、计算机设备
US20230223033A1 (en) Method of Noise Reduction for Intelligent Network Communication

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20911413

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020911413

Country of ref document: EP

Effective date: 20220209

NENP Non-entry into the national phase

Ref country code: DE