WO2022153504A1 - 学習方法、学習システム及び学習プログラム - Google Patents
学習方法、学習システム及び学習プログラム Download PDFInfo
- Publication number
- WO2022153504A1 WO2022153504A1 PCT/JP2021/001354 JP2021001354W WO2022153504A1 WO 2022153504 A1 WO2022153504 A1 WO 2022153504A1 JP 2021001354 W JP2021001354 W JP 2021001354W WO 2022153504 A1 WO2022153504 A1 WO 2022153504A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- learning
- data
- noise
- acoustic model
- shuffle
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 41
- 238000012549 training Methods 0.000 claims description 9
- 230000006870 function Effects 0.000 description 16
- 238000007781 pre-processing Methods 0.000 description 14
- 238000012545 processing Methods 0.000 description 11
- 238000010586 diagram Methods 0.000 description 10
- 238000013527 convolutional neural network Methods 0.000 description 5
- 230000010365 information processing Effects 0.000 description 5
- 230000002093 peripheral effect Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000006866 deterioration Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000012946 outsourcing Methods 0.000 description 1
- 239000010454 slate Substances 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
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
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/063—Training
Definitions
- the present invention relates to a learning method, a learning system and a learning program.
- voice recognition that converts voice information into characters has been used in contact centers and conference proceedings systems in companies.
- the usage scene of voice recognition is very wide, and various forms are prepared for the provision method, and the voice recognition service is also provided in a form that anyone can use with cloud type services etc. ing.
- the acoustic model is a model that links the characteristics of the sound of voice with the information of phonemes.
- DNN Deep Neural Network
- a pair of voice feature series and phoneme series in the usage scene is prepared, and the relationship between the feature quantity and the phoneme type is learned for each frame.
- a Neural Network with various structures can be used for the DNN used for learning the voice model, but the most basic model is a network using a fully connected layer or a Convolutional Neural Network.
- the phonemes of a certain frame are determined by inputting several frames before and after the frame. As the number of frames before and after a certain frame, for example, a total of 11 frames, 5 frames before and after the frame, are used.
- the features of the voice are arranged in the original time series, and in the unlikely event that the learning data leaks out, the features are used for voice recognition. Confidential information included in the feature amount may be restored.
- the phoneme state numbers are also arranged in the original time series, the state numbers can be converted into phoneme strings, and there is a risk that confidential information will be restored.
- the present invention has been made in view of the above, and an object of the present invention is to improve the safety of model learning for speech recognition.
- the shuffle step of acquiring the learning data arranged in a time series and rearranging the learning data in an order different from the order of the time series includes a learning step of learning an acoustic model using the rearranged learning data.
- the safety of model learning for speech recognition can be improved.
- FIG. 1 is a block diagram of a learning system according to the first embodiment.
- FIG. 2 is a diagram showing an example of a learning database.
- FIG. 3 is a diagram showing an example of a shuffled learning database.
- FIG. 4 is a flowchart of the learning process of the acoustic model by the learning system according to the first embodiment.
- FIG. 5 is a block diagram of the learning system according to the second embodiment.
- FIG. 6 is a diagram showing an example of a computer that executes a learning program.
- FIG. 1 is a block diagram of a learning system according to the first embodiment.
- the learning system 1 includes a preprocessing device 10 and a learning device 20.
- the preprocessing device 10 and the learning device 20 are connected via a network such as the Internet.
- the learning device 20 is, for example, a computer that provides an acoustic recognition service located in the cloud.
- the preprocessing device 10 is a computer of a user who uses the sound recognition service provided by the learning device 20, provides learning data to the learning device 20, creates a trained sound model 24, and causes the learning device 20 to create a trained sound model 24.
- the sound recognition service is used by using the created learned sound model 24.
- the preprocessing device 10 can handle confidential information, it is preferable that the learning device 20 performs processing in a state where the confidential information is hidden.
- the pretreatment device 10 will be described. As shown in FIG. 1, the preprocessing device 10 has a learning database 11, a shuffle unit 12, and a noise addition unit 13.
- the learning database 11 is a database that holds learning data for performing machine learning on the acoustic model 24.
- FIG. 2 is a diagram showing an example of a learning database.
- the phoneme state number is an example of a phoneme label.
- the shuffle unit 12 assigns a new data number j that is uniquely determined regardless of the data number or the time to the feature amount and the phoneme number registered in the learning database 11. As a result, the shuffle unit 12 changes the order in which the learning data are arranged from the order registered in the learning database 11. That is, the shuffle unit 12 acquires the learning data arranged in the time series and rearranges the learning data in an order different from the order in the time series.
- the shuffle unit 12 reconstructs the database by assigning a new data number to the feature amount and the phoneme number as follows. Specifically, the shuffle unit 12 determines the feature amount X (j) for the new data number j and the phoneme state number Y (j) as represented by the following mathematical formula (1).
- L is an integer of 0 or more, and is a parameter indicating the peripheral time length used for determining the phoneme at the time.
- the shuffle unit 12 uses the mathematical formula (1) for the feature quantities X (j) and phonemes for all the training data of the data numbers i (1, 2, ..., I) registered in the learning database 11.
- the state number Y (j) of is determined.
- the shuffle unit 12 uses the mathematical formula (1) for the feature amount X (j) and the phonemes for all the times t (0, 1, 2, ..., T) for each learning data of the data number i.
- the determination of the state number Y (j) of is repeated. In that case, the shuffle unit 12 uniquely and randomly assigns a new data number j to the feature amount X around the time t and the phoneme state number at the time t in the learning database 11 for each time t. This will create a new database. After that, the shuffle unit 12 outputs the generated new database to the noise addition unit 13.
- the noise addition unit 13 receives input from the shuffle unit 12 of a new database in which learning data having a new data number j is registered.
- the noise addition unit 13 adds a minute amount of noise having an amount of information that cannot be restored by rearranging the data to each feature amount registered in the acquired new database. As a result, the noise addition unit 13 cannot be restored binary even if the information of the feature amount at each time t and its surroundings is used.
- the noise adding unit 13 adds a minute noise to the feature amount by using the following mathematical formula (2).
- ⁇ represents a minute noise
- is an absolute value of the feature amount X (j).
- ⁇ is expressed by the following mathematical formula (3) using ⁇ representing a minute value.
- 0, ⁇ ) represents a normal distribution having an average of 0 and a variance of ⁇ .
- ⁇ can be set based on the fluctuation range of voice within the same phoneme.
- the noise addition unit 13 repeats the load of noise on the feature amount X using the mathematical formula (2) for all the learning data of the data number j. In this way, the noise addition unit 13 adds noise to the feature amount to the extent that the data cannot be rearranged and restored, so that the binary can be created using the time corresponding to the learning data whose data number is j and the feature amount around it. It makes it difficult to find the exact matching features. This makes it difficult to restore the original data.
- the noise addition unit 13 transmits a database in which noise is added to the registered features to the learning device 20 and stores it as a shuffled learning database 21.
- the learning device 20 has a shuffled learning database 21, a learning data selection unit 22, a learning unit 23, and an acoustic model 24.
- FIG. 3 is a diagram showing an example of a shuffled learning database.
- the shuffled learning database 21 as the learning data of the new data number j determined by the shuffle unit 12, the data in which noise is added to the feature amount around the time t corresponding to the learning data and the phonetic elements at the time t are added. Status number is registered.
- the learning data selection unit 22 selects the data number j as the dam. Then, the learning data selection unit 22 acquires the learning data of the data number j from the learning data held by the shuffled learning database 21. Then, the learning data selection unit 22 outputs the acquired learning data to the learning unit 23. The learning data selection unit 22 repeats the selection of the learning data from the shuffled learning database 21 and the output to the learning unit 23 until the learning of the learning unit 23 is completed.
- the learning unit 23 receives input of learning data from the learning data selection unit 22.
- the learning unit 23 learns the acoustic model 24 by using the feature amount X (j) and the phoneme state number Y (j) included in the acquired learning data.
- the learning unit 23 can learn the acoustic model 16 by general learning.
- the learning unit 23 converts the state number Y (j) into a OneHot vector and updates the parameters of the acoustic model 24 based on the loss function shown in the following mathematical formula (4). ..
- ⁇ represents the acoustic model 24.
- Y (j) c is the c-th element when Y (j) is a OneHot vector.
- X (j), ⁇ ) c indicates the probability that the feature quantity X (j) is Y (j).
- the learning unit 23 acquires learning data from the learning data selection unit 22 and repeats learning of the acoustic model 24 using the acquired learning data until the update of the acoustic model 24 converges. That is, the learning unit 23 learns the acoustic model 24 using the learning data rearranged by the shuffle unit 12. After that, when the update of the acoustic model 24 converges, the learning unit 23 ends the learning of the acoustic model 24.
- the acoustic model 24 can use a fully connected layer or a CNN (Convolutional Neural Network).
- the trained acoustic model 24 is used for a voice recognition service or the like on the cloud.
- FIG. 4 is a flowchart of the learning process of the acoustic model by the learning system according to the first embodiment.
- the shuffle unit 12 randomly selects a data number from the learning database 11 and randomly determines the time in the learning data having the selected data number. Next, the shuffle unit 12 acquires the peripheral feature amount at the determined time of the learning data of the selected data number and the state number at that time. Then, the shuffle unit 12 assigns a new data number j that is uniquely determined regardless of the data number or the time to the acquired feature amount and phoneme number. By repeating the above processing, the shuffle unit 12 acquires the peripheral feature amount of each data number in each accident billion and the state number at each time from the learning database 11, and assigns a new data number to each. A new database is generated (step S1).
- the noise addition unit 13 receives input from the shuffle unit 12 of a new database in which the order of learning data is shuffled. Next, the noise adding unit 13 adds a minute noise to each feature amount registered in the acquired new database to generate the shuffled learning database 21 (step S2).
- the learning data selection unit 22 randomly selects a data number. Then, the learning data selection unit 22 acquires the learning data of the data number selected from the learning data held in the shuffled learning database 21 (step S3).
- the learning unit 23 receives input of learning data from the learning data selection unit 22. Then, the learning unit 23 executes learning of the acoustic model 24 using the feature amount X (j) and the phoneme state number Y (j) included in the acquired learning data (step S4).
- step S5 determines whether or not the update of the acoustic model 24 has converged. If the update of the acoustic model 24 has not converged (step S5: negation), the learning process of the acoustic model 24 returns to step S3.
- step S5 affirmative
- the learning unit 23 ends the learning process of the acoustic model 24.
- the learning process of the acoustic model by the learning system 1 is a shuffle step of acquiring the learning data arranged in a time series and rearranging the learning data in an order different from the order of the time series. And a learning process of learning the acoustic model 24 using the learning data rearranged by the shuffle process.
- shuffle process learning data having a feature amount and a phoneme label is acquired, and new learning data is generated by combining the feature amount in a certain period around a specific time and the phoneme label at a specific time, in chronological order. Arranges new training data in a different order. Further, in the learning process, the acoustic model 24 is learned using the new learning data.
- the learning device randomly adds data numbers and times from a database of learning data in which data numbers are added in chronological order and feature quantities and phonetic state numbers are arranged in chronological order. Is selected, and a new database is created by assigning a new data number as training data by collecting the feature quantities around the selected time in the training data of the data number and the state number of the phonetic element at that time. Further, the learning device creates a shuffled learning database by adding noise to each of the features included in each learning data of the new database. Then, the learning device learns the acoustic model using the shuffled learning database.
- the time-series information disappears by changing the order of the learning data, and it becomes difficult to restore the confidential information because it is difficult to match as a binary due to a minute noise. That is, even if the learning data for learning the speech model contains confidential information, the acoustic model can be learned by converting the acoustic feature quantity sequence and the phoneme sequence so that the confidential information cannot be restored. .. In this way, by using information whose confidential information cannot be restored, it is possible to learn the acoustic model while suppressing the risk of information leakage. Therefore, it is possible to improve the safety of model learning for speech recognition.
- FIG. 5 is a block diagram of the learning system according to the second embodiment.
- the preprocessing device 10 of the learning system 1 according to the present embodiment is different from the first embodiment in that it adds noise generated from an acoustic model prepared in advance.
- the pretreatment device 10 according to the present embodiment has a base acoustic model 14 in addition to each part of the first embodiment. In the following description, description of the functions of the same parts as in the first embodiment will be omitted.
- the base acoustic model 14 is an acoustic model that has been separately learned in advance.
- the base acoustic model 14 can be generated by having the learning selection unit 15 and the learning unit 23 learn using other learning data.
- the noise addition unit 13 uses the base acoustic model 14 to add noise to the feature amount in consideration of the fluctuation of the sound. Specifically, the noise addition unit 13 adds noise to the feature amount using the following mathematical formula (5) for all the training data to which the data number j is assigned.
- ⁇ represents a minute positive value noise
- the function in parentheses of ⁇ differentiates the loss function Rose by x and inputs X (j), Y (j) and ⁇ corresponding to the base acoustic model 14. It is the value that was set.
- the function in parentheses of ⁇ is a function that differentiates the loss function by the elements of each feature quantity, and when it is positive, the loss function becomes large, and when it is negative, the loss function becomes small.
- a large loss function means that the model deviates from the model that considers the fluctuation of the voice
- a small loss function means that the model approaches the model that considers the fluctuation of the voice.
- the noise addition unit 13 adds noise to the feature amount in the direction in which the loss function becomes smaller in the mathematical formula (5). As a result, the noise addition unit 13 can add noise in a direction that does not deviate from the fluctuation of the sound.
- the learning process of the acoustic model 24 by the learning system 1 according to the second embodiment generates minute noise having an amount of information that cannot be restored to the state of being arranged in chronological order even if the learning data is rearranged.
- the learning device uses an acoustic model prepared in advance to add noise to the feature amount so that the loss function becomes small.
- the learning device can add noise in a direction that does not deviate from the fluctuation of the voice, and it is possible to reduce the deterioration of the accuracy of the voice brocade due to the addition of the noise. That is, it is possible to learn an acoustic model that suppresses the risk of information leakage while suppressing a decrease in the accuracy of speech recognition by adding fluctuations that tend to occur in speech as noise. Therefore, it is possible to maintain the accuracy of speech recognition and improve the safety of model learning for speech recognition.
- each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific forms of distribution and integration of each device are not limited to those shown in the figure, and all or part of them may be functionally or physically dispersed or physically distributed in arbitrary units according to various loads and usage conditions. Can be integrated and configured.
- the preprocessing device 10 and the learning device 20 can be combined to form a single learning device.
- each processing function performed by each device is realized by a CPU (Central Processing Unit) and a program that is analyzed and executed by the CPU, or hardware by wired logic. Can be realized as.
- CPU Central Processing Unit
- the preprocessing device 10 and the learning device 20 can be implemented by installing a display control program that executes the above information processing as package software or online software on a desired computer.
- the information processing device can function as the preprocessing device 10 or the learning device 20.
- the information processing device referred to here includes a desktop type or notebook type personal computer.
- information processing devices include smartphones, mobile communication terminals such as mobile phones and PHS (Personal Handy-phone System), and slate terminals such as PDAs (Personal Digital Assistants). Is done.
- the preprocessing device 10 and the learning device 20 can be implemented as a management server device in which the terminal device used by the user is a client and the service related to the above management processing is provided to the client.
- the management server device is implemented as a server device that receives a config input request as an input and provides a management service for inputting a config.
- the management server device may be implemented as a Web server, or may be implemented as a cloud that provides services related to the above management processing by outsourcing.
- FIG. 6 is a diagram showing an example of a computer that executes a learning program.
- the computer 1000 has, for example, a memory 1010 and a CPU 1020.
- the computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. Each of these parts is connected by a bus 1080.
- the memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012.
- the ROM 1011 stores, for example, a boot program such as a BIOS (BASIC Input Output System).
- BIOS BASIC Input Output System
- the hard disk drive interface 1030 is connected to the hard disk drive 1090.
- the disk drive interface 1040 is connected to the disk drive 1100.
- a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100.
- the serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120.
- the video adapter 1060 is connected to, for example, the display 1130.
- the hard disk drive 1090 stores, for example, OS1091, application program 1092, program module 1093, and program data 1094. That is, the learning program that defines each process of the preprocessing device 10 or the learning device 20 having the same function as the preprocessing device 10 or the learning device 20 is implemented as a program module 1093 in which a code that can be executed by a computer is described. To.
- the program module 1093 is stored in, for example, the hard disk drive 1090.
- a program module 1093 for executing a process similar to the functional configuration in the preprocessing device 10 or the learning device 20 is stored in the hard disk drive 1090.
- the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).
- the setting data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, a memory 1010 or a hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 into the RAM 1012 as needed, and executes the processing of the above-described embodiment.
- the program module 1093 and the program data 1094 are not limited to those stored in the hard disk drive 1090, but may be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). Then, the program module 1093 and the program data 1094 may be read by the CPU 1020 from another computer via the network interface 1070.
- LAN Local Area Network
- WAN Wide Area Network
Landscapes
- Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Electrically Operated Instructional Devices (AREA)
Abstract
Description
[学習システム]
図1は、第1の実施形態に係る学習システムのブロック図である。図1に示すように、本実施形態に係る学習システム1は、前処理装置10及び学習装置20を有する。前処理装置10と学習装置20とは、例えば、インターネットなどのネットワークを介して接続される。学習装置20は、例えばクラウドに配置された音響認識サービスを提供するコンピュータである。前処理装置10は、学習装置20により提供される音響認識サービスを利用する利用者のコンピュータであり、学習装置20に対して学習データを提供し、学習済みの音響モデル24を作成させて、その作成された学習済みの音響モデル24を用いて音響認識サービスを利用する。前処理装置10では機密情報を扱えるが、学習装置20では機密情報を隠蔽した状態で処理が行われることが好ましい。
前処理装置10について説明する。前処理装置10は、図1に示すように、学習データベース11、シャッフル部12及びノイズ付加部13を有する。
次に、学習装置20について説明する。学習装置20は、図2に示すように、シャッフル済学習データベース21、学習データ選択部22、学習部23及び音響モデル24を有する。
次に、図4を参照して、本実施形態に係る学習システム1による音響モデル24の学習処理の流れについて説明する。図4は、第1の実施形態に係る学習システムによる音響モデルの学習処理のフローチャートである。
以上に説明したように、本実施形態に係る学習装置は、時系列順にデータ番号が付加され時系列順に特徴量と音素の状態番号とが並んだ学習データのデータベースから、ランダムにデータ番号及び時刻を選択して、そのデータ番号の学習データにおける選択時刻の周辺の特徴量とその時刻での音素の状態番号をまとめて学習データとして新たなデータ番号を付与して新たなデータベースを作成する。さらに、学習装置は、新たなデータベースの各学習データに含まれる特徴量のそれぞれにノイズを加えてシャッフル済学習データベースを作成する。そして、学習装置は、シャッフル済学習データベースを用いて音響モデルの学習を行う。
図5は、第2の実施形態に係る学習システムのブロック図である。本実施形態に係る学習システム1の前処理装置10は、予め用意された音響モデルから生成されたノイズを付加することが第1の実施形態と異なる。本実施形態に係る前処理装置10は、第1の実施形態の各部に加えてベース音響モデル14を有する。以下の説明では、第1の実施形態と同じ各部の機能については説明を省略する。
ベース音響モデル14は、予め別途学習しておいた音響モデルである。ベース音響モデル14は、他の学習データを用いて学習選択部15及び学習部23に学習させることで生成可能である。
以上に説明したように、本実施形態に係る学習装置は、予め用意された音響モデルを用いて損失関数が小さくなるようにノイズを特徴量に付加する。これにより、学習装置は、音声の揺らぎを逸脱しない方向にノイズを加えることができ、ノイズの付加による音声錦の精度劣化を軽減することが可能となる。すなわち、音声にありがちな揺らぎをノイズとして加えることで音声認識の精度の低下を抑えつつ、情報漏洩に関するリスクを抑えた音響モデルの学習を行うことができる。したがって、音声認識の精度の維持及び音声認識のためのモデル学習の安全性の向上を実現することが可能となる。
また、図示した各装置の各構成要素は機能概念的なものであり、必ずしも物理的に図示のように構成されていることを要しない。すなわち、各装置の分散及び統合の具体的形態は図示のものに限られず、その全部又は一部を、各種の負荷や使用状況等に応じて、任意の単位で機能的又は物理的に分散又は統合して構成することができる。特に、前処理装置10と学習装置20とを組み合わせて、一台の学習装置として構成することも可能である。さらに、各装置にて行われる各処理機能は、その全部又は任意の一部が、CPU(Central Processing Unit)及び当該CPUにて解析実行されるプログラムにて実現され、あるいは、ワイヤードロジックによるハードウェアとして実現され得る。
一実施形態として、前処理装置10及び学習装置20は、パッケージソフトウェアやオンラインソフトウェアとして上記の情報処理を実行する表示制御プログラムを所望のコンピュータにインストールさせることによって実装できる。例えば、上記の表示制御プログラムを情報処理装置に実行させることにより、情報処理装置を前処理装置10又は学習装置20として機能させることができる。ここで言う情報処理装置には、デスクトップ型又はノート型のパーソナルコンピュータが含まれる。また、その他にも、情報処理装置にはスマートフォン、携帯電話機やPHS(Personal Handy-phone System)等の移動体通信端末、さらには、PDA(Personal Digital Assistant)等のスレート端末等がその範疇に含まれる。
10 前処理装置
11 学習データベース
12 シャッフル部
13 ノイズ付加部
14 ベース音響モデル
20 学習装置
21 シャッフル済学習データベース
22 学習データ選択部
23 学習部
24 音響モデル
Claims (7)
- 時系列に並んだ学習データを取得して、前記時系列の順序とは異なる順序で前記学習データを並び替えるシャッフル工程と、
前記シャッフル工程により並び替えられえた前記学習データを用いて音響モデルの学習を行う学習工程と
を備えたことを特徴とする学習方法。 - 前記シャッフル工程は、特徴量及び音素ラベルを有する前記学習データを取得して、特定時刻の周辺の一定期間における前記特徴量と前記特定時刻の前記音素ラベルとを組み合わせて新たな学習データを生成し、前記時系列の順序とは異なる順序で前記新たな学習データを並べ、
前記学習工程は、前記新たな学習データを用いて前記音響モデルの学習を行う
ことを特徴とする請求項1に記載の学習方法。 - 前記学習データを並び替えても時系列で並んだ状態に復元できない情報量を有する微小なノイズを生成して、前記シャッフル工程において並び替えられた前記学習データに付加するノイズ付加工程をさらに備えたことを特徴とする請求項1又は2に記載の学習方法。
- 前記ノイズ付加工程は、予め作成されたベース音響モデルを基に、前記ノイズを生成することを特徴とする請求項3に記載の学習方法。
- 前記ノイズ付加工程は、音声の揺らぎを逸脱しない前記ノイズを生成することを特徴とする請求項3又は4に記載の学習方法。
- 時系列に並んだ学習データを取得して、前記時系列の順序とは異なる順序で前記学習データを並び替えるシャッフル部と、
前記シャッフル部により並び替えられえた学習データを用いて音響モデルの学習を行う学習部と
を備えたことを特徴とする学習システム。 - コンピュータに請求項1~5に記載の方法で処理を実行させるための学習プログラム。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US18/272,246 US20240078999A1 (en) | 2021-01-15 | 2021-01-15 | Learning method, learning system and learning program |
PCT/JP2021/001354 WO2022153504A1 (ja) | 2021-01-15 | 2021-01-15 | 学習方法、学習システム及び学習プログラム |
JP2022575008A JPWO2022153504A1 (ja) | 2021-01-15 | 2021-01-15 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/001354 WO2022153504A1 (ja) | 2021-01-15 | 2021-01-15 | 学習方法、学習システム及び学習プログラム |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022153504A1 true WO2022153504A1 (ja) | 2022-07-21 |
Family
ID=82448146
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2021/001354 WO2022153504A1 (ja) | 2021-01-15 | 2021-01-15 | 学習方法、学習システム及び学習プログラム |
Country Status (3)
Country | Link |
---|---|
US (1) | US20240078999A1 (ja) |
JP (1) | JPWO2022153504A1 (ja) |
WO (1) | WO2022153504A1 (ja) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2004252167A (ja) * | 2003-02-20 | 2004-09-09 | Nippon Telegr & Teleph Corp <Ntt> | 音素モデル学習用文リスト生成方法、生成装置、および生成プログラム |
WO2011052412A1 (ja) * | 2009-10-28 | 2011-05-05 | 日本電気株式会社 | 音声認識システム、音声認識要求装置、音声認識方法、音声認識用プログラムおよび記録媒体 |
JP2011248001A (ja) * | 2010-05-25 | 2011-12-08 | Nippon Telegr & Teleph Corp <Ntt> | 音響モデル学習用ラベル作成装置、その方法及びプログラム |
JP2015225296A (ja) * | 2014-05-29 | 2015-12-14 | 富士通株式会社 | 音響モデル調整装置及びプログラム |
JP2016212273A (ja) * | 2015-05-11 | 2016-12-15 | 国立研究開発法人情報通信研究機構 | リカレント型ニューラルネットワークの学習方法及びそのためのコンピュータプログラム、並びに音声認識装置 |
JP2021039293A (ja) * | 2019-09-05 | 2021-03-11 | Nttテクノクロス株式会社 | 情報処理装置、情報処理方法及びプログラム |
-
2021
- 2021-01-15 WO PCT/JP2021/001354 patent/WO2022153504A1/ja active Application Filing
- 2021-01-15 US US18/272,246 patent/US20240078999A1/en active Pending
- 2021-01-15 JP JP2022575008A patent/JPWO2022153504A1/ja active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2004252167A (ja) * | 2003-02-20 | 2004-09-09 | Nippon Telegr & Teleph Corp <Ntt> | 音素モデル学習用文リスト生成方法、生成装置、および生成プログラム |
WO2011052412A1 (ja) * | 2009-10-28 | 2011-05-05 | 日本電気株式会社 | 音声認識システム、音声認識要求装置、音声認識方法、音声認識用プログラムおよび記録媒体 |
JP2011248001A (ja) * | 2010-05-25 | 2011-12-08 | Nippon Telegr & Teleph Corp <Ntt> | 音響モデル学習用ラベル作成装置、その方法及びプログラム |
JP2015225296A (ja) * | 2014-05-29 | 2015-12-14 | 富士通株式会社 | 音響モデル調整装置及びプログラム |
JP2016212273A (ja) * | 2015-05-11 | 2016-12-15 | 国立研究開発法人情報通信研究機構 | リカレント型ニューラルネットワークの学習方法及びそのためのコンピュータプログラム、並びに音声認識装置 |
JP2021039293A (ja) * | 2019-09-05 | 2021-03-11 | Nttテクノクロス株式会社 | 情報処理装置、情報処理方法及びプログラム |
Also Published As
Publication number | Publication date |
---|---|
JPWO2022153504A1 (ja) | 2022-07-21 |
US20240078999A1 (en) | 2024-03-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Polino et al. | Model compression via distillation and quantization | |
US10515155B2 (en) | Conversational agent | |
US10025692B2 (en) | Scalable continuous integration and delivery systems and methods | |
US11418461B1 (en) | Architecture for dynamic management of dialog message templates | |
WO2022121257A1 (zh) | 模型训练方法、语音识别方法、装置、设备及存储介质 | |
KR20190117713A (ko) | 신경망 아키텍처 최적화 | |
US11948023B2 (en) | Automatic application program interface (API) selector for unsupervised natural language processing (NLP) intent classification | |
CN112328849A (zh) | 用户画像的构建方法、基于用户画像的对话方法及装置 | |
CN112686382B (zh) | 一种卷积模型轻量化方法及*** | |
US8019593B2 (en) | Method and apparatus for generating features through logical and functional operations | |
CN111767394A (zh) | 一种基于人工智能专家***的摘要提取方法及装置 | |
US20100153108A1 (en) | Method for dynamic learning of individual voice patterns | |
WO2022153504A1 (ja) | 学習方法、学習システム及び学習プログラム | |
CN111797220A (zh) | 对话生成方法、装置、计算机设备和存储介质 | |
CN112667791A (zh) | 潜在事件预测方法、装置、设备及存储介质 | |
CN109978145B (zh) | 一种处理方法及装置 | |
CN107766944B (zh) | 一种利用api分析进行***功能流优化的***和方法 | |
JP7099254B2 (ja) | 学習方法、学習プログラム及び学習装置 | |
JP6772115B2 (ja) | 音響モデル学習装置、音声認識装置、音響モデル学習方法、音声認識方法、及びプログラム | |
WO2023152914A1 (ja) | 埋め込み装置、埋め込み方法、および、埋め込みプログラム | |
WO2022070342A1 (ja) | 学習装置、学習方法及び学習プログラム | |
Lee et al. | Robustness of network centrality metrics in the context of digital communication data | |
WO2024023946A1 (ja) | 音声処理装置、音声処理方法及び音声処理プログラム | |
WO2022034675A1 (ja) | 信号処理装置、信号処理方法、信号処理プログラム、学習装置、学習方法、および、学習プログラム | |
US20230359904A1 (en) | Training device, training method and training program |
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: 21919397 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2022575008 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18272246 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21919397 Country of ref document: EP Kind code of ref document: A1 |