JPS5849998A - Voice recognition unit - Google Patents

Abstract

(57)【要約】本公報は電子出願前の出願データであるた
め要約のデータは記録されません。(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

【発明の詳細な説明】 本発明は音声を認識する事のできる音声線′ｍ装置に関
する。DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech line device capable of recognizing speech.

第１図に従来の首声認識装置の構成を示Ｔ、同図に於い
て、（１）はマイクロフォンであシ、入力された酋−が
電気的な信号波形に変換される。（２）は該マイクロフ
ォン（１）からの音声の信号波形を周波　−数分析して
得られる周波数成分値な音声の特徴パラメータとして抽
出するパラメータ抽出回路であ〕、音声帯域（１００１
ｇ〜４１馬）を８分割した８つの周波数成分値が時系列
的に１６ナンプル抽出され′る。（３）は入力音声パタ
ーンメモリであ〕、該パラメータ抽出回路（２）から得
られる入力音声の１６ナンプルの８成分からなる周波数
成分値が入力音声パターンとして貯えられる。（４バ４
）・・・は複数の参照音声パターンメモリで４）、入力
音声パターンと同ｒｖ、ＩＩＩの１６チンプルの８成分
からなる参照音声パターンが予じめ複数の認識音声毎に
貯見られている。（５）は上記入力音声パターンノ％５
Ｃ３）の入力音声パターンと上記膠原音声パターンメモ
リ＜４）１４）−の各認識音声の参照音声パターンとの
誤差を求める誤差算出回路であｐ１両音声パターンの対
応Ｔるナンプルの対雇Ｔる周波数成分値間の誤差の総和
が両音声パターン間の距離として算出され、この距離が
距離メモリ（６）（６）・−に各認識音声毎に貯えられ
る・（７）は音声決定回路であ〕、上記距離メモ！ｊ（
６３（句・・・に貯えられ九各認識音声毎の距離の内、
最小のものを検出して、Ｃの最小の距離を示す認識音声
がこの時の入力音声として決定される・ 新様な従来の音声認識装置は、その誤差算出回路（５Ｊ
　Ｋ　ｉいて、８行１６列にパターン化された入力及び
参照同音声パターン間の各成分の誤差の総和、即ち距離
を全ての認識音声について計算しなければならなかりた
。Ｃの為、多数の認識音声に対応しようとＴれば、上記
誤差算出回路（５）での計算量が増し、この結果入力音
声に対する認識応答を実時間で実行する事ができなくな
る不都合があうた。FIG. 1 shows the configuration of a conventional speech recognition device. In the figure, (1) is a microphone, which converts the input voice into an electrical signal waveform. (2) is a parameter extraction circuit for extracting voice characteristic parameters such as frequency component values obtained by frequency analysis of the signal waveform of the voice from the microphone (1).
16 frequency component values obtained by dividing the frequency component (g~41 horse) into 8 are extracted in chronological order. (3) is an input voice pattern memory] in which frequency component values consisting of 8 components of 16 samples of the input voice obtained from the parameter extraction circuit (2) are stored as an input voice pattern. (4ba4
)... is a plurality of reference speech pattern memories 4). A reference speech pattern consisting of 8 components of 16 chimples of the same rv and III as the input speech pattern is stored in advance for each of a plurality of recognized speeches. (5) is the above input voice pattern %5
An error calculation circuit that calculates the error between the input voice pattern of C3) and the reference voice pattern of each recognized voice of <4) 14) - p1 is used to calculate the correspondence T of both voice patterns and the pair of numbers T. The sum of the errors between the frequency component values is calculated as the distance between both voice patterns, and this distance is stored in the distance memory (6) (6) - for each recognized voice. (7) is the voice determination circuit. ], Distance memo above! j(
63 (phrases...), of the distances for each recognized voice,
The minimum distance is detected, and the recognized voice indicating the minimum distance of C is determined as the input voice at this time.The new conventional voice recognition device uses its error calculation circuit (5J
K i , the sum of the errors of each component, that is, the distance, between the input and reference speech patterns patterned in 8 rows and 16 columns had to be calculated for all recognized speech. Because of C, when trying to deal with a large number of recognized voices, the amount of calculation in the error calculation circuit (5) increases, resulting in the inconvenience that recognition responses to input voices cannot be executed in real time. Ta.

本発明は祈る不都合を解消する目的で為されたものでる
）、以下に詳述する。The present invention has been made for the purpose of solving the above mentioned inconveniences), and will be described in detail below.

第２図に本発明の音声認識装置の一実施例を示す。同図
に於いて％（１）〜（５）は第１図に示した従来装置と
同様にマイクロフォン〜誤差算出回路を示しているが１
本発明装置の入力音声パターンメモリ（３）及びｗｌ−
一照音声パターンメモリ１４Ｊ−・・が、夫々適数ｓｗ
：、に分割されている点で従来装置と異なうてお）１本
実施例の場合は、第１乃至第４饋域Ｇｌ１国町肌包Ｄ（
社）簡（財）・−に分割されている。即ち。FIG. 2 shows an embodiment of the speech recognition device of the present invention. In the figure, %(1) to (5) indicate the microphone to error calculation circuit as in the conventional device shown in Figure 1.
Input voice pattern memory (3) of the device of the present invention and wl-
Issho voice pattern memory 14J-... has an appropriate number of sw
This device differs from the conventional device in that it is divided into : , 1) In this embodiment, the 1st to 4th area Gl1 Kunimachi skin area D(
The company is divided into the following: That is.

この場合、８成分からなる周波数成分値が時系列的な１
６チンプルで構成される入力及び参照音声パターンを用
るなら、始めの４チンプルが第１領域Ｇも＠〃・−に貯
えられ１次の４チンプルが第２領域（至）、（４３・・
・に貯えられ、更に次の４チンプル、及び最後の４ナン
プルが同様に第ｓｕｍ職（至）、囮・・・及び第４頷斌
−１（財）・・・に犬々貯えられる。そして誤差１出回
％（５）Ｋ於ては、汽力及び参照両音声パターンメモリ
（３Ｊ、圃・軸の各Ｉｌｉ域毎に各音声パターン間の４
ナンプル分の誤差が夫々算出されるｅｓ　（８１（８）
・・・は誤差、累積メモリであシ、上記誤差算出回ｗｌ
（５）から得られる４ナンプル分の誤！が、６認識音声
毎に累積して貯えられｉ・そしてこの各誤差累積メモリ
（８）（８）−・・に１！ｔンプル分の累積誤差、即ち
。In this case, the frequency component values consisting of 8 components are 1
If we use an input and reference speech pattern consisting of 6 chimples, the first 4 chimples will be stored in the first area G as well, and the first 4 chimples will be stored in the second area (to) (43...
The next 4 digits and the last 4 digits are stored in the same way as the sum position (to), the decoy, and the 4th position -1 (wealth). And error 1 occurrence % (5) In K, both steam power and reference voice pattern memory (3J, 4 between each voice pattern for each Ili area of field/axis)
es (81(8)
... is an error, it is the cumulative memory, and the above error calculation time wl
An error equal to 4 numbers obtained from (5)! is accumulated and stored for every 6 recognized voices i, and each error accumulation memory (8) (8) - 1! Cumulative error for t samples, ie.

両パターン間の距離が蓄積された時点で、Ｈ声決定回路
（７）に依〕、この時の入力音声が決定される。When the distance between both patterns is accumulated, the input voice at this time is determined by the H voice determination circuit (7).

（９Ｊは比較回路であシ、各誤差累積メモリ（８）（８
）・−に貯えられクラある各累積誤差の内、これ等の累
積誤差の平、均値よ如小なる認識音声が認識音声候補・
として比較検出される。圓は上記参脈パタ、ンメモ９（
４バ４）−・・の各ｓｔ＠ｎ廁■（財）・−を読み出す
為の読み出し制御回路であ〕、上記比較回路（９３に依
りて検出された認識音声候補に対する参照音声パターン
メモ！Ｐ（４の次のＩＩＩＭの４ナンプル分が読み出さ
れ、上記比較回路（９）に依りて検出されなかりた認識
音声、即ち両音声パターン間の前半部分に於いてずでに
それ等の誤差が大きいと判定され、入力音声と同じとみ
なＴ事が不可能な認識音声、はこの時点で認識錫塩から
除外される。(9J is a comparison circuit, each error accumulation memory (8) (8
)・- Among the various cumulative errors stored in , the recognized voice that is smaller than the average value of these cumulative errors is the recognized voice candidate.
Detected by comparison. En is the reference pattern mentioned above, Nmemo 9 (
This is a readout control circuit for reading out each st@n廁■(corporate)・- of 4) and the reference speech pattern memo for the recognized speech candidates detected by the comparison circuit (93). The 4 numbers of IIIM next to P(4) are read out, and the recognized speech that was not detected by the comparison circuit (9), that is, the first half between the two speech patterns, is already recognized. Recognized speech that is determined to have a large error and cannot be regarded as the same as the input speech is excluded from the recognition at this point.

次に、祈る構成の音声認識装置の動作を説明する。マイ
クロフォン（１）Ｋ［−声が入力され、パラメータ抽出
回ｉ１　（２）から８成分１６ｆンプルの周波数成分値
が入力音声パターンとして、抽出される。Next, the operation of the voice recognition device having a prayer configuration will be explained. Microphone (1) K[- voice is input, and frequency component values of 8 components and 16f samples are extracted as an input voice pattern from the parameter extraction step i1 (2).

そして、この入力音声パターンが４ナンプル分づつ分割
されて入力音声パターンメモリ（３Ｊの第１乃至第４１
１１ｍ（３１）〜（ロ）に貯えられる。一方、複数の参
照音声パターンメモ５（４）・軸の夫々には各認識音声
の参照音声パターンが４？ンプル分づつ第１乃至第４ｓ
域（４υ〜（財）に４分割されて格納されている。Then, this input voice pattern is divided into 4 numbers each in the input voice pattern memory (3J 1st to 41st numbers).
It is stored in 11m (31) to (b). On the other hand, each of the multiple reference voice pattern memo 5 (4) axes has 4 reference voice patterns for each recognized voice. 1st to 4th s for each sample
area (4υ~(goods) and is stored in four parts.

そして売上、認識電通の為の第１ステツプでは、読み出
し制御回路−に依うて、全ての参ｗＡ音声パターンメモ
リ（４ハ４）・−の第１　ａ＊（４１）Ｇｉυ・・・の
参照音声パターンの４ナンプル分が、順次読み出され、
これ等と、入力音声パターンメモ９（３）の第１＊ｆｉ
、Ｃ（υの入力音声パターンの４ナンプル分と１、苓の
誤差が誤差算出回ｌｉ！ｒｔ５）に依りて算出される。In the first step for sales and recognition, the readout control circuit reads the first a * (41) Giυ... of all references wA voice pattern memory (4c4). The four number numbers of the voice pattern are read out sequentially,
In addition to these, the first *fi of input voice pattern memo 9 (3)
, C (the error of 4 numbers of the input voice pattern of υ and 1, 蓓 is calculated according to the error calculation cycle li!rt5).

祈して得られた両パターン間の始めの４ナンブル分に限
うての誤差が全ての認識音声毎に各誤差累積メモリ＋８
１（８）−’に貯えられ、比較回ｆｉ％（９）に依９て
、これ等の誤差の内、これ等の誤差の平均値よ〕小なる
誤Ｍ’に示す認識音声、即ち、認識音声の約半数となる
第１の認識音声候補群が選出される。そして。The error limited to the first 4 numbers between the two patterns obtained by praying is the error cumulative memory + 8 for each recognized voice.
1(8)-', and depending on the comparison times fi%(9), among these errors, the average value of these errors] is the smaller error M', i.e., A first group of recognized speech candidates, which is about half of the recognized speech, is selected. and.

次の第２ステツプでは、読み出し制御回路（２）に於い
て、ｎ述の第１ステツプで比較回路（９）に依９で選出
された第１の認識音声候補に限って、その番９１　音声
Ａ　！　−ン／　モ５　（４）（４）・”　（Ｄ第２　
ｆａｌｌ（４２）’Ａ３＝　Ｋ貯えられた膠原音声パタ
ーンの４１ンプル分が順次読み出され、これ等と入力音
声パターンメモリ（３）の第２領域−に貯えられた入力
音声パターンの４＋ンプル分と、の誤差が誤差算出回路
（５）に依りて算出される。所して得られた両パターン
間の第の ２番目の４ナンブル分に限フＳＷＡ差が第１の認識音声
候補毎罠各誤差累積メモリ（８１（８１−・・に累積蓄
積され、第１の認識音声候補群の各誤差累積メモリ（８
）（８）・−の誤差の内、これ等の累積誤差の平均値よ
）小なる累積誤！！を示す第２の認識音声候補群（全て
の認識音声の約−の数）が選出される。次の第３ステツ
プに於−ても、第２ステツプと同様に第２の認識音声候
補群にりいての両音声パターンの第５＠域での誤差算出
め理が為され、第６の認識音声候補群（全ての認識音声
の約−の数）が選出される。そして最後の第４ステツプ
では、読み出し制御回路−に依りて読み出さｒ、た第６
の認識音声候補群についての両音声パターン間の最後の
４ナンブルの誤差が誤差算出回路（５）に依って算出さ
れ、夫々に対応した誤差累積メモリＴ８）（８）に累積
蓄積される。所して得られた第６の認識音声候補群の累
積誤差、即ち両音声パターン間の距離の内最小となるも
のを、音声決定回路（７）に依りて検出すると共に、仁
の距離が最小となる認識音声がこの時の入力音声と決定
される。In the next second step, the readout control circuit (2) selects only the first recognized speech candidate selected in step 9 by the comparison circuit (9) in the first step described above, and selects the number 91 speech. A! -n/Mo5 (4)(4)・” (D 2nd
fall (42)'A3=K 41 samples of the stored collagen speech patterns are read out in sequence, and these and 4+ samples of the input speech pattern stored in the second area - of the input speech pattern memory (3) The error between and is calculated by the error calculation circuit (5). The SWA difference is limited to the second 4 numbers between the two patterns obtained in this way, and is cumulatively accumulated in each trap error accumulation memory (81 (81-...) for each first recognized speech candidate, and Each error cumulative memory (8
)(8)・- The average value of these cumulative errors) is a small cumulative error! ! A second group of recognized speech candidates (approximately - number of all recognized speech) indicating the following is selected. In the next third step, similarly to the second step, the error calculation principle in the fifth region of both speech patterns in the second recognition speech candidate group is performed, and the sixth recognition A group of voice candidates (approximately - number of all recognized voices) is selected. In the fourth and final step, the readout control circuit reads r, and the sixth
The error calculation circuit (5) calculates the last four numbers of errors between the two speech patterns for the recognized speech candidate group, and accumulates them in the corresponding error accumulation memories T8 (8). The cumulative error of the sixth recognized speech candidate group obtained in this way, that is, the minimum distance between both speech patterns is detected by the speech determination circuit (7), and the distance between the two speech patterns is detected as the minimum. The recognized speech that becomes is determined as the input speech at this time.

本発明に係る音声認識装置は、上述の如く、入力及び各
参照音声パターンを犬５４ｓＩｊｆｃに分割した場合、
その誤差算出回Ｍ　（５Ｊか行なう誤差計算量は、第１
図の従来例に比べて、第１ステツプに於あ夛、全体内−
は釣上の誤差計算量で済む事にな■ る・ 本発明の音声認識装置は１以上の説明から明らかな如く
、入力及び各膠原音声パターンの入力及び参照音声パタ
ーンを複数１ｍＫに分割する手段な設け、入力及び各参
照音声パターン間のＩＩＪｉ域部分に於いて１丁でにＣ
れ等の誤差が大き−と判定され、入力音声であるとみと
められな一適数個の認識音声＃ｔｇｉ１１謙音声から除
外し、これ等除外され九認識歯声に対Ｔる両参照音声パ
ターン間の残シの領域での誤差算出を省略したものであ
るので、全ての認識音声に対して入力及び各＃照パター
ン間の誤！！を完全に計算していた従来装置に比べ、こ
の誤差計算獣な大巾に低減せしめる事ができ、多数の認
Ｍ音声に対しても°−Ａ時間でのａ声応答が可能となる
。The speech recognition device according to the present invention, as described above, when the input and each reference speech pattern is divided into 54sIjfc,
The number of error calculations M (5J) is the first
Compared to the conventional example shown in the figure, there is an increase in the number of steps in the first step.
As is clear from the above description, the speech recognition device of the present invention has means for dividing the input and each collagen speech pattern into a plurality of 1 mK segments. In the IIJi area between the input and each reference voice pattern, one
A suitable number of recognized voices #tgi11 which are determined to have large errors and cannot be considered to be input voices are excluded from the voice, and these are excluded and both reference voice patterns for the nine recognized dental voices are created. This method omits error calculation in the region of residuals in between, so there are no errors between input and each pattern for all recognized speech! ! Compared to conventional devices which completely calculate the error, this calculation error can be greatly reduced, and it is possible to respond to a-voice in °-A time even to a large number of recognized M-voices.

【図面の簡単な説明】[Brief explanation of the drawing]

Claims (1)

【特許請求の範囲】 １）入力音声の信号波形からその音声の特徴パラメータ
系列を抽出Ｔるパラメータ抽出回路と。 該パラメータ抽出回路から得られる特徴パラメータ系列
を入力音声パターンとして貯える入力音声パターンメモ
リと、予じめ多数の認識音声にり一て、これ等の特徴パ
ラメータ系列を参照音声パターンとして貯えた＃原音声
パターンメモリと、ム参照音声パターンメモリの参照音
声パターン、及び上記入力音声パターンメモリの各入力
音声パターンを夫々間Ｓに複数ＷＡ域に分割Ｔるパター
ン分割手段と、該パターン分割手段に依フて分割された
ｆＥ４斌毎に、入力音声パターンと各＃＊雷産声パター
ンとの誤！！を算−する誤差算出回路と、かうなル、音
声パターンの分割ｌｌ域毎に上記娯差算ｍ１路から得ら
れ４各認識音声に対応する誤Ｚａ累積が大きなものｔ為
ら順に適数の認識音声を除外し。 これに依うて残された認識音声候補の参ＷＩＡ雪声パタ
ーンを用−て１次の音声パターンの分割＠＊に於−て誤
差算出回路から得られる誤差の累積が大きなものから順
に適数の認識音声を除外する処理なく〕返し、最後の音
声パターンの分割ｉＩ域での誤差の累積が最大となる認
識音声をこの時の入力音声と決定する事を特徴とした音
声認識装置。[Scope of Claims] 1) A parameter extraction circuit that extracts a characteristic parameter series of input speech from a signal waveform of the input speech. an input speech pattern memory that stores the feature parameter series obtained from the parameter extraction circuit as an input speech pattern, and #original speech that stores these feature parameter series as reference speech patterns based on a large number of recognized speeches in advance. A pattern memory, a reference voice pattern in the reference voice pattern memory, and a pattern dividing means for dividing each input voice pattern in the input voice pattern memory into a plurality of WA areas in intervals S, and depending on the pattern dividing means. For every divided fE4 bin, there is a mistake between the input voice pattern and each #* thunder voice pattern! ! An error calculation circuit that calculates the error Za obtained from the above-mentioned error calculation m1 path for each divided region of the speech pattern, and 4. Exclude recognized voice. Using the reference WIA snow voice patterns of the remaining recognition voice candidates, the primary voice pattern is divided into appropriate numbers in descending order of the cumulative error obtained from the error calculation circuit. The speech recognition device is characterized in that the recognized speech is determined to be the input speech at this time, without processing to exclude the recognized speech of the last speech pattern.