JPH0895595A - Learning method for hidden markov model - Google Patents

Learning method for hidden markov model

Info

Publication number
JPH0895595A
JPH0895595A JP6228534A JP22853494A JPH0895595A JP H0895595 A JPH0895595 A JP H0895595A JP 6228534 A JP6228534 A JP 6228534A JP 22853494 A JP22853494 A JP 22853494A JP H0895595 A JPH0895595 A JP H0895595A
Authority
JP
Japan
Prior art keywords
covariance
variance
hmm
processing
hidden markov
Prior art date
Legal status (The legal status 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 status listed.)
Granted
Application number
JP6228534A
Other languages
Japanese (ja)
Other versions
JP3091648B2 (en
Inventor
Takashi Miki
敬 三木
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Oki Electric Industry Co Ltd
Original Assignee
Oki Electric Industry Co Ltd
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 Oki Electric Industry Co Ltd filed Critical Oki Electric Industry Co Ltd
Priority to JP06228534A priority Critical patent/JP3091648B2/en
Publication of JPH0895595A publication Critical patent/JPH0895595A/en
Application granted granted Critical
Publication of JP3091648B2 publication Critical patent/JP3091648B2/en
Anticipated expiration legal-status Critical
Expired - Fee Related legal-status Critical Current

Links

Abstract

PURPOSE: To reduce a voice dictionary learning procedure for recognition device. CONSTITUTION: Input voice data as the learning object are analyzed for every fixed time and transformed to a set of acoustic feature variables representing the feature of the input voice data in voice analyzing processing S11. The variance and the covariance of a set of acoustic feature variables are obtained in variance/covariance setting processing S12. The variance and the covariance of the function of an outputted probability density represented by the multi- dimensional normal distribution of HMM maximizing the likelihood of HMM for a set of acoustic feature variables are obtained in HMM parameter estimating processing S13. The variance and covariance obtained in the HMM parameter estimating processing is modified on the basis of the variance and covariance obtained in the variance/covariance setting processing in HMM parameter modifying processing S14. Whether or not the estimation of parameter in the HMM parameter estimating processing S13 is converged is decided in parameter convergence deciding processing S15.

Description

【発明の詳細な説明】Detailed Description of the Invention

【0001】[0001]

【産業上の利用分野】本発明は、音声認識における隠れ
マルコフモデル(Hidden Markov Model 、以下、HMM
という)の学習方法に関するものである。
BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to Hidden Markov Model (HMM) in speech recognition.
That is) about the learning method.

【0002】[0002]

【従来の技術】従来、このような分野の技術としては、
例えば、次のような文献に記載されるものがあった。 文献1;日本音響学会誌、48巻[1号]、(1992)、嵯峨
山茂樹、“数理統計モデルによる音声認識の現状と将
来”、P.431-437 文献2;中川聖一著、“確率モデルによる音声認識”、
(1988)、電子情報通信学会、P.29-73 HMMは、その統計的性質から、音声パタンのような発
声速度に伴う時間変動、発声の個人差、及び調音結合等
の揺らぎを含むパタンを適切に表現できることから、音
声認識の分野において近年広く用いられている。HMM
を用いて音声を認識する方法、即ち、HMM音声認識方
法とは、音声を確率的な遷移ネットワークと見做して認
識処理を行うものである。図2は、従来の音声認識方法
に用いられる単語HMMの構造例を示す図である。図2
において、Si(i;0,1,2,3)はHMMにおけ
る音声の特徴等の状態を表す。Aklは状態遷移確率、
bkl(x)は遷移の際にある音声スペクトルxを出力
する確率である出力確率密度関数を表す。音声スペクト
ルxは入力音声データの特徴を表すものである。HMM
は、図2に示すように、いくつかの状態Si(i;0,
1,2,3)と、それらの状態間の遷移を表す弧で表現
される。各弧には、その弧を遷移する状態遷移確率Ak
l及び出力確率密度関数bkl(x)がパラメータとし
て与えられている。
2. Description of the Related Art Conventionally, techniques in such a field include:
For example, some documents were described in the following documents. Reference 1: Journal of Acoustical Society of Japan, Volume 48 [No. 1], (1992), Shigeki Sagayama, "Present and Future of Speech Recognition by Mathematical Statistical Model," P.431-437 Reference 2: Seiichi Nakagawa, " Speech recognition by probabilistic model ”,
(1988), The Institute of Electronics, Information and Communication Engineers, P.29-73 HMM, based on its statistical properties, analyzes patterns such as voice patterns that include time fluctuations associated with vocalization speed, individual differences in vocalization, and fluctuations such as articulatory coupling. Since it can be expressed appropriately, it has been widely used in the field of voice recognition in recent years. HMM
The method of recognizing a voice by using, that is, the HMM voice recognition method is a method of recognizing a voice as a stochastic transition network. FIG. 2 is a diagram showing a structural example of a word HMM used in a conventional speech recognition method. Figure 2
In, Si (i; 0, 1, 2, 3) represents a state such as a voice feature in the HMM. Akl is the state transition probability,
bkl (x) represents an output probability density function which is a probability of outputting a certain speech spectrum x at the time of transition. The voice spectrum x represents the characteristics of the input voice data. HMM
, Some states Si (i; 0,
1, 2, 3) and arcs that represent the transitions between these states. For each arc, the state transition probability Ak that transits the arc
l and the output probability density function bkl (x) are given as parameters.

【0003】HMMは、初期状態S0から遷移を開始し
て、最終状態S3に到達するまで弧を通って遷移を繰り
返し、スペクトル系列を出力する確率が得られる。出力
確率密度関数bkl(x)をスペクトルxの関数とした
ものは、連続出力確率分布HMMと呼ばれている。通
常、出力確率密度関数bkl(x)の真の関数型を求め
ることは不可能なので、通例、数学的取扱いが簡単で、
かつ表現能力が高い多次元正規分布がよく用いられる。
出力確率密度関数bkl(x)の正味のパラメータは、
平均ベクトルμと分散及び共分散ρである。図3は、従
来のHMMの学習方法を示すフローチャートである。連
続出力確率分布HMMで音声認識を行うにあたっては、
認識対象となる音声を最もよく表すHMMのパラメータ
を求めることが必要である。この処理をHMMの学習と
いう。以下、図3に従ってHMMの学習手順S1〜S3
を説明する。音声分析処理S1において、学習対象とな
る音声データが入力され、音響特徴量に変換される。H
MMパラメータ推定処理S2において、前記文献2に記
載されたB−W(Baum-Welch)アルゴリズム等により、
HMMパラメータの推定を行う。パラメータ収束判定処
理S3において、パラメータ推定が収束したか否かを判
定し、未収束であれば、HMMパラメータ推定処理S2
へ戻り、更に推定計算を行う。パラメータ推定が収束し
た場合、HMM学習は終了する。
The HMM has a probability of starting a transition from the initial state S0 and repeating the transition through an arc until the final state S3 is reached to output a spectrum sequence. A function in which the output probability density function bkl (x) is a function of the spectrum x is called a continuous output probability distribution HMM. Usually, it is impossible to find the true functional form of the output probability density function bkl (x), so that the mathematical treatment is usually simple.
A multidimensional normal distribution with high expressive power is often used.
The net parameter of the output probability density function bkl (x) is
Mean vector μ and variance and covariance ρ. FIG. 3 is a flowchart showing a conventional HMM learning method. When performing speech recognition with the continuous output probability distribution HMM,
It is necessary to find the HMM parameters that best represent the speech to be recognized. This process is called HMM learning. Hereinafter, the learning procedure S1 to S3 of the HMM according to FIG.
Will be explained. In the voice analysis process S1, voice data to be learned is input and converted into acoustic feature quantities. H
In the MM parameter estimation processing S2, by the BW (Baum-Welch) algorithm described in the above-mentioned Document 2,
Estimate the HMM parameters. In the parameter convergence determination process S3, it is determined whether or not the parameter estimation has converged. If not, the HMM parameter estimation process S2
Then, the estimation calculation is performed again. When the parameter estimation converges, the HMM learning ends.

【0004】[0004]

【発明が解決しようとする課題】しかしながら、従来の
HMMの学習方法では、次のような課題があった。即
ち、HMMパラメータの推定、特に、多次元正規分布に
おけるパラメータである出力確率密度関数bkl(x)
中の分散及び共分散ρの推定には、大量の学習音声デー
タが必要であった。学習音声データ数が不足している
と、分散及び共分散ρの値が推定できず、音声認識の性
能が大幅に低下する欠点があった。本発明は、学習音声
データが少ない場合でも、適切な出力確率密度関数bk
l(x)を設定できるHMMの学習方法を提供するもの
である。
However, the conventional HMM learning method has the following problems. That is, the estimation of the HMM parameter, particularly the output probability density function bkl (x) which is a parameter in the multidimensional normal distribution.
A large amount of training speech data was required to estimate the variance and covariance ρ inside. If the number of training speech data is insufficient, the values of variance and covariance ρ cannot be estimated, and there is a drawback that the performance of speech recognition is significantly reduced. According to the present invention, an appropriate output probability density function bk can be used even when the learning voice data is small.
It is intended to provide an HMM learning method capable of setting l (x).

【0005】[0005]

【課題を解決するための手段】本発明では、前記課題を
解決するために、学習の対象となる入力音声データを分
析して該入力音声データの特徴を表す音響特徴量に変換
し、該音響特徴量から前記入力音声データのHMMのパ
ラメータを推定するHMMの学習方法において、次のよ
うな手段を講じている。即ち、学習の対象となる入力音
声データを一定時間毎に分析して入力音声データの特徴
を表す音響特徴量の組に変換する音声分析処理と、音響
特徴量の組の分散及び共分散を求める分散・共分散設定
処理とを行う。次に、音響特徴量の組に対するHMMの
尢度を最大にするHMMのパラメータを求めるHMMパ
ラメータ推定処理と、HMMパラメータ推定処理におい
て求めた出力確率密度関数のパラメータである分散及び
共分散を、分散・共分散設定処理において求めた分散及
び共分散を基準にして修正するHMMパラメータ修正処
理とを、行うようにしている。
According to the present invention, in order to solve the above-mentioned problems, input voice data to be learned is analyzed and converted into an acoustic feature amount representing the feature of the input voice data, In the HMM learning method of estimating the HMM parameters of the input speech data from the feature amount, the following means are taken. That is, the input voice data to be learned is analyzed at regular intervals and converted into a set of acoustic feature quantities representing the features of the input voice data, and the variance and covariance of the set of acoustic feature quantities are obtained. Performs dispersion / covariance setting processing. Next, the HMM parameter estimation process for obtaining the HMM parameter that maximizes the HMM strength for the set of acoustic feature amounts, and the variance and covariance that are the parameters of the output probability density function obtained in the HMM parameter estimation process are The HMM parameter correction process for correcting the variance and the covariance obtained in the covariance setting process is used as a reference.

【0006】[0006]

【作用】本発明によれば、以上のようにHMMの学習方
法を構成したので、音声分析処理において、学習の対象
となる入力音声データを一定時間毎に分析して入力音声
データの特徴を表す音響特徴量の組に変換する。分散・
共分散設定処理において、前記音響特徴量の組の分散及
び共分散を求める。HMMパラメータ推定処理におい
て、前記音響特徴量の組に対するHMMの尢度を最大に
するHMMのパラメータを求める。HMMパラメータ修
正処理において、前記HMMパラメータ推定処理におい
て求めた出力確率密度関数のパラメータである分散及び
共分散を前記分散・共分散設定処理において求めた分散
及び共分散を基準にして修正する。そのため、学習の対
象となる入力音声データが少ない場合でも、音響特徴量
全体のデータの統計量を用いて、推定精度の不十分なパ
ラメータを修正することにより、HMM認識精度が入力
音声データが多いときと同様に高く保たれる。その結
果、認識装置の音声辞書学習手続きが大幅に軽減され
る。従って、前記課題を解決できるのである。
According to the present invention, since the learning method of the HMM is configured as described above, in the voice analysis process, the input voice data to be learned is analyzed at regular intervals to show the characteristics of the input voice data. Convert to a set of acoustic features. dispersion·
In the covariance setting process, the variance and covariance of the acoustic feature set are obtained. In the HMM parameter estimation processing, an HMM parameter that maximizes the HMM quality for the set of acoustic feature quantities is obtained. In the HMM parameter correction processing, the variance and covariance, which are the parameters of the output probability density function obtained in the HMM parameter estimation processing, are revised based on the variance and covariance obtained in the variance / covariance setting processing. Therefore, even when the input speech data to be learned is small, the HMM recognition accuracy is high in the input speech data by correcting the parameters whose estimation accuracy is insufficient by using the statistics of the data of the entire acoustic feature amount. As high as it is, it is kept high. As a result, the voice dictionary learning procedure of the recognition device is greatly reduced. Therefore, the above problem can be solved.

【0007】[0007]

【実施例】図1は、本発明の実施例のHMMの学習方法
を説明するためのフローチャートである。本実施例のH
MMの学習方法では、例えば、プログラム制御されるコ
ンピュータを用いて図1に示す処理S11〜S15が実
行される。次に、この図1を用いて本発明のHMMの学
習方法を説明する。音声分析処理S11において、学習
対象となる音声データが入力され、フレーム毎に音響特
徴量の組である特徴ベクトルx(t)iに変換される。
特徴ベクトルx(t)iは、例えば、前記文献2に示さ
れているLPCケプストラムを用いる。ここで、tはフ
レーム番号、iはLPCケプストラムの番号PはLPC
ケプストラムの次数である。(i=1,…,P) 音声データは、通常複数の特徴ベクトル系列(以下、特
徴パタンという)からなる。ここで、特徴パタンの数を
PNとし、特徴ベクトルの総数をFNとする。分散・共
分散設定処理S12において、特徴ベクトルx(t)i
の分散及び共分散αijを、次の(1)式を用いて求め
る。
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a flowchart for explaining an HMM learning method according to an embodiment of the present invention. H of this embodiment
In the MM learning method, for example, the processes S11 to S15 shown in FIG. 1 are executed using a program-controlled computer. Next, the HMM learning method of the present invention will be described with reference to FIG. In the voice analysis process S11, the voice data to be learned is input and converted into a feature vector x (t) i which is a set of acoustic feature amounts for each frame.
As the feature vector x (t) i, for example, the LPC cepstrum shown in the above-mentioned Document 2 is used. Where t is the frame number, i is the LPC cepstrum number P is the LPC
The order of the cepstrum. (I = 1, ..., P) The audio data is usually composed of a plurality of feature vector series (hereinafter referred to as feature patterns). Here, the number of feature patterns is PN, and the total number of feature vectors is FN. In the variance / covariance setting process S12, the feature vector x (t) i
The variance and covariance α ij of are calculated using the following equation (1).

【0008】[0008]

【数1】 ここで、βiは特徴ベクトルx(t)iの平均値であ
り、次の(2)式で表される。
[Equation 1] Here, βi is the average value of the feature vector x (t) i and is represented by the following equation (2).

【0009】[0009]

【数2】 ここでは、全ての学習データに基づいて分散及び共分散
αij、及び平均値βiを計算したが、何らかの先験的
な値或いは学習データの一部分から求めてもよい。更
に、分散及び共分散αij、及び平均値βiは、1つだ
けでなく、何らかのカテゴリ別に求め、適宜使い分けて
もよい。ここで設定された分散及び共分散αijを基準
分散・共分散という。
[Equation 2] Here, the variance and covariance αij and the average value βi are calculated based on all the learning data, but they may be obtained from some a priori values or a part of the learning data. Further, the variance and covariance αij and the average value βi are not limited to one, but may be obtained for some category and appropriately used. The variance and covariance αij set here are referred to as a reference variance / covariance.

【0010】HMMパラメータ推定処理S13におい
て、B−Wアルゴリズム等により、先の特徴ベクトルx
(t)iを用いてHMMパラメータの推定を行う。HM
Mパラメータは、初期状態確率πk、遷移確率Akl、
多次元正規分布における出力現確率密度関数中の平均ベ
クトルμkli及び分散・共分散ρklijで表され、
これらのパラメータの集合をθとする。HMMパラメー
タの推定とは、特徴ベクトルxに対するHMMの尢度L
(θ,x)を最大にする最尢推定値θ´を求めることで
ある。この推定されたHMMのパラメータを初期状態確
率πk´、遷移確率Akl´、平均ベクトルμkli´
及び分散・共分散ρklij´とする。又、その尢度の
最大値をL´とする。HMMパラメータ修正処理S14
において、HMMパラメータ推定処理S13で推定され
た分散・共分散ρklij´を、特徴パタン数PNに応
じて、次の(3)式を用いて修正する。αijは、分散
・共分散設定処理S12において設定された基準分散・
共分散である。
In the HMM parameter estimation processing S13, the feature vector x is calculated by the BW algorithm or the like.
(T) The HMM parameters are estimated using i. HM
The M parameters are initial state probability πk, transition probability Akl,
It is represented by the mean vector μkli and the variance / covariance ρklij in the output current probability density function in the multidimensional normal distribution,
Let θ be the set of these parameters. The estimation of the HMM parameter is the degree L of the HMM for the feature vector x.
This is to find the minimum estimated value θ ′ that maximizes (θ, x). The estimated HMM parameters are the initial state probability πk ′, the transition probability Akl ′, and the average vector μkli ′.
And the variance / covariance ρklij ′. Further, the maximum value of the degree of elasticity is L '. HMM parameter correction processing S14
In, the variance / covariance ρklij ′ estimated in the HMM parameter estimation process S13 is corrected using the following equation (3) according to the number of characteristic patterns PN. αij is the reference variance / covariance set in the variance / covariance setting process S12.
It is covariance.

【0011】[0011]

【数3】 但し、 M;特徴パタン数PNの大小判断を行うための閾値 即ち、特徴パタン数PNが閾値M以下である場合、上限
2 αijと下限C1αijを設定することである。但
し、C1 及びC2 は分散・共分散ρklij´の範囲を
定める定数である。
[Equation 3] However, M: a threshold value for judging the magnitude of the characteristic pattern number PN, that is, when the characteristic pattern number PN is equal to or less than the threshold value M, an upper limit C 2 αij and a lower limit C 1 αij are set. However, C 1 and C 2 are constants that define the range of the variance / covariance ρklij ′.

【0012】又、別の修正方法として、信頼度fが特徴
パタン数PNの関数(0≦f(PN)≦1)であると考
えて、次の(4)式のように修正する方法もある。 ρklij´=f(PN)ρklij´+{1−f(PN)}αij ・・・(4) 関数f(PN)については、実際のデータに基づいて決
めていくのがよいが、簡易な関数として、次の(5)式
でも十分に妥当な修正ができる。 f(PN)=1−1/(1+0.2PN) ・・・(5) パラメータ収束判定処理S15において、先の尢度最大
値L´の増加の度合により、HMMパラメータ推定処理
S13でのパラメータ推定が収束したか否かを判定す
る。その判定方法としては、例えば、尢度最大値L´
と、1回前の尢度最大値L´´との差(L´−L´´)
が、或る閾値D以下であれば収束したと判断する方法が
ある。即ち、L´−L´´>Dならば、HMMパラメー
タ推定処理S3へ戻り、L´−L´´≦Dならば、パラ
メータ推定が終了する。以上のように、本実施例では、
音声分析処理S11における学習対象となる入力音声デ
ータが少ない場合、HMMパラメータ修正処理S14に
おいて、特徴ベクトル全体のデータの基準分散・共分散
αijを用いて推定精度の不十分なパラメータを修正す
ることにより、少量学習時でのHMM認識精度を高く保
つことができる。その結果、認識装置の音声辞書学習
(登録)手続きという使用者の負担となる作業を、大幅
に軽減できる。尚、本発明のHMMの学習方法は、多次
元正規分布を出力確率密度関数にもつHMMを用いた全
ての音声認識装置や、その他HMMを利用したパタン認
識に容易に利用できる。
As another correction method, it is considered that the reliability f is a function (0 ≦ f (PN) ≦ 1) of the number PN of characteristic patterns, and the correction is performed by the following expression (4). is there. ρklij ′ = f (PN) ρklij ′ + {1-f (PN)} αij (4) The function f (PN) should be decided based on actual data, but a simple function As a result, the following equation (5) can be used to make a sufficiently appropriate correction. f (PN) = 1−1 / (1 + 0.2PN) (5) In the parameter convergence determination process S15, the parameter estimation in the HMM parameter estimation process S13 is performed according to the degree of increase of the previous maximum power L ′. It is determined whether or not has converged. As the determination method, for example, the maximum value L ′
And the difference (L'-L '') from the maximum value L ″ of the previous degree
However, if there is a certain threshold value D or less, there is a method of determining that it has converged. That is, if L'-L ''> D, the process returns to the HMM parameter estimation processing S3, and if L'-L''≤D, the parameter estimation ends. As described above, in this embodiment,
When the input speech data to be learned in the speech analysis processing S11 is small, the parameters of insufficient estimation accuracy are corrected in the HMM parameter correction processing S14 by using the reference variance / covariance αij of the data of the entire feature vector. It is possible to keep the HMM recognition accuracy high when learning a small amount. As a result, it is possible to significantly reduce the work that is a burden on the user for the voice dictionary learning (registration) procedure of the recognition device. The HMM learning method of the present invention can be easily used for all speech recognition devices using an HMM having a multidimensional normal distribution as an output probability density function, and for pattern recognition using other HMMs.

【0013】[0013]

【発明の効果】以上詳細に説明したように、本発明によ
れば、学習の対象となる入力音声データが少ない場合で
も、HMMパラメータ修正処理において、入力音声デー
タの特徴を表す音響特徴量全体の統計量を用いて、推定
精度の不十分なパラメータを修正することにより、少量
学習時でのHMM認識精度を高く保つことができる。そ
のため、認識装置に対する音声辞書学習(登録)手続き
という使用者の負担となる作業を大幅に軽減できる。
As described above in detail, according to the present invention, even when the input voice data to be learned is small, in the HMM parameter correction processing, all the acoustic feature quantities representing the features of the input voice data are processed. By correcting the parameter of which the estimation accuracy is insufficient by using the statistic, it is possible to keep the HMM recognition accuracy high at the time of learning a small amount. Therefore, it is possible to significantly reduce the work of the user, which is a voice dictionary learning (registration) procedure for the recognition device.

【図面の簡単な説明】[Brief description of drawings]

【図1】本発明の実施例のHMMの学習方法のフローチ
ャートである。
FIG. 1 is a flowchart of an HMM learning method according to an embodiment of the present invention.

【図2】単語HMMの構造例を示す図である。FIG. 2 is a diagram showing a structural example of a word HMM.

【図3】従来のHMMの学習方法のフローチャートであ
る。
FIG. 3 is a flowchart of a conventional HMM learning method.

【符号の説明】[Explanation of symbols]

S11 音声分
析処理 S12 分散・
共分散設定処理 S13 HMM
パラメータ推定処理 S14 HMM
パラメータ修正処理
S11 Speech analysis processing S12 Dispersion /
Covariance setting process S13 HMM
Parameter estimation process S14 HMM
Parameter modification process

Claims (1)

【特許請求の範囲】[Claims] 【請求項1】 学習の対象となる入力音声データを分析
して該入力音声データの特徴を表す音響特徴量に変換
し、該音響特徴量から前記入力音声データの隠れマルコ
フモデルのパラメータを推定する多次元正規分布で表わ
される出力確率密度関数を持つ隠れマルコフモデルの学
習方法において、 学習の対象となる入力音声データを一定時間毎に分析し
て該入力音声データの特徴を表す音響特徴量の組に変換
する音声分析処理と、 前記音響特徴量の組の分散及び共分散を求める分散・共
分散設定処理と、 前記音響特徴量の組に対する隠れマルコフモデルの尢度
を最大にする隠れマルコフモデルのパラメータを求める
隠れマルコフモデルパラメータ推定処理と、 前記隠れマルコフモデルパラメータ推定処理において求
めたパラメータ中の分散及び共分散を前記分散・共分散
設定処理において求めた分散及び共分散を基準にして修
正する隠れマルコフモデルパラメータ修正処理とを、 行うことを特徴とする隠れマルコフモデルの学習方法。
1. An input speech data to be learned is analyzed and converted into an acoustic feature quantity representing a characteristic of the input speech data, and a parameter of a hidden Markov model of the input speech data is estimated from the acoustic feature quantity. In a learning method of a hidden Markov model having an output probability density function represented by a multidimensional normal distribution, input speech data to be learned is analyzed at regular time intervals, and a set of acoustic feature quantities representing the characteristics of the input speech data. A voice analysis process for converting to, a variance / covariance setting process for obtaining the variance and covariance of the acoustic feature set, and a hidden Markov model for maximizing the degree of hiding of the hidden Markov model for the acoustic feature set. Hidden Markov model parameter estimation processing for obtaining parameters, and variance and distribution in the parameters obtained in the hidden Markov model parameter estimation processing. Learning method of hidden Markov models and hidden Markov model parameter correction processing, and carrying out the modified covariance based on the variance and covariance calculated in the covariance setting processing.
JP06228534A 1994-09-26 1994-09-26 Learning Hidden Markov Model Expired - Fee Related JP3091648B2 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP06228534A JP3091648B2 (en) 1994-09-26 1994-09-26 Learning Hidden Markov Model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
JP06228534A JP3091648B2 (en) 1994-09-26 1994-09-26 Learning Hidden Markov Model

Publications (2)

Publication Number Publication Date
JPH0895595A true JPH0895595A (en) 1996-04-12
JP3091648B2 JP3091648B2 (en) 2000-09-25

Family

ID=16877916

Family Applications (1)

Application Number Title Priority Date Filing Date
JP06228534A Expired - Fee Related JP3091648B2 (en) 1994-09-26 1994-09-26 Learning Hidden Markov Model

Country Status (1)

Country Link
JP (1) JP3091648B2 (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100404852B1 (en) * 1996-08-03 2004-02-25 엘지전자 주식회사 Speech recognition apparatus having language model adaptive function and method for controlling the same
CN110059392A (en) * 2019-04-11 2019-07-26 桂林电子科技大学 A kind of landslide deformation prediction method
CN112513563A (en) * 2018-08-31 2021-03-16 株式会社小松制作所 Work machine transported object specifying device, work machine transported object specifying method, completion model production method, and learning dataset

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100404852B1 (en) * 1996-08-03 2004-02-25 엘지전자 주식회사 Speech recognition apparatus having language model adaptive function and method for controlling the same
CN112513563A (en) * 2018-08-31 2021-03-16 株式会社小松制作所 Work machine transported object specifying device, work machine transported object specifying method, completion model production method, and learning dataset
CN110059392A (en) * 2019-04-11 2019-07-26 桂林电子科技大学 A kind of landslide deformation prediction method

Also Published As

Publication number Publication date
JP3091648B2 (en) 2000-09-25

Similar Documents

Publication Publication Date Title
US5812975A (en) State transition model design method and voice recognition method and apparatus using same
JP2690027B2 (en) Pattern recognition method and apparatus
CN111816165A (en) Voice recognition method and device and electronic equipment
JP4336865B2 (en) Voice recognition device
JP4531166B2 (en) Speech recognition method using reliability measure evaluation
JPS62231996A (en) Allowance evaluation of word corresponding to voice input
EP1465154B1 (en) Method of speech recognition using variational inference with switching state space models
US6240389B1 (en) Pattern matching method and apparatus
EP1863014B1 (en) Apparatuses and methods for learning and using a distance transition model
WO2007105409A1 (en) Reference pattern adapter, reference pattern adapting method, and reference pattern adapting program
JP3632529B2 (en) Voice recognition apparatus and method, and recording medium
KR100574769B1 (en) Speaker and environment adaptation based on eigenvoices imcluding maximum likelihood method
US6226610B1 (en) DP Pattern matching which determines current path propagation using the amount of path overlap to the subsequent time point
JP4703648B2 (en) Vector codebook generation method, data compression method and apparatus, and distributed speech recognition system
JP2852298B2 (en) Standard pattern adaptation method
JPH0895595A (en) Learning method for hidden markov model
JP3536471B2 (en) Identification device and identification method, and speech recognition device and speech recognition method
JPWO2008126254A1 (en) Speaker recognition device, acoustic model update method, and acoustic model update processing program
JP3176210B2 (en) Voice recognition method and voice recognition device
JP3029803B2 (en) Word model generation device for speech recognition and speech recognition device
JPH06266384A (en) Acousting model adaption system
JP2000259198A (en) Device and method for recognizing pattern and providing medium
JP2000099078A (en) Standard pattern learning device
JP2734828B2 (en) Probability calculation device and probability calculation method
JPH11212591A (en) Pattern recognition method, device therefor and recording medium recorded with pattern recognizing program

Legal Events

Date Code Title Description
A01 Written decision to grant a patent or to grant a registration (utility model)

Free format text: JAPANESE INTERMEDIATE CODE: A01

Effective date: 20000711

FPAY Renewal fee payment (event date is renewal date of database)

Free format text: PAYMENT UNTIL: 20080721

Year of fee payment: 8

FPAY Renewal fee payment (event date is renewal date of database)

Free format text: PAYMENT UNTIL: 20090721

Year of fee payment: 9

LAPS Cancellation because of no payment of annual fees