JP5089655B2 - Acoustic model creation device, method and program thereof - Google Patents

Description

  The present invention relates to an acoustic model creation device used for speech recognition, a method thereof, and a program.

New creation / update of an acoustic model is usually done by first using a posterior random variable based on forward and backward calculations from an observed signal sequence for learning, a corresponding state sequence, and an acoustic model (any initial model in the case of new creation). γ _t (i) (for example, forward-backward algorithm: see non-patent document 1) or forward probability δ _t (i) based on forward calculation (for example, viterbi algorithm: refer to non-patent document 2) is obtained. Note that the posterior random variable γ _t (i) exists in the state i (1 ≦ i ≦ J, J is the number of states in the model) at time t (1 ≦ t ≦ T, T is the length of the observed signal sequence). The forward probability δ _t (i) represents the probability of the state sequence that gives the highest probability among the state sequences that reach state i at time t. It is necessary to obtain the most likely state sequence (maximum likelihood path) based on the posterior random variable γ _t (i) and forward probability δ _t (i), and to update the acoustic model while following the maximum likelihood path. The various statistics (hereinafter referred to as “sufficient statistics”) are extracted, and the acoustic model is updated accordingly.

  FIG. 1 shows a configuration example of a conventional acoustic model creation apparatus 100 when using the Viterbi algorithm, and FIG. 2 shows a processing flow example. The acoustic model creation apparatus 100 includes a learning data acquisition unit 101, a feature amount analysis unit 102, a state series conversion unit 103, an acoustic model storage unit 104, a forward calculation unit 105, a sufficient statistics accumulation unit 106, and a model update unit 107. .

  The learning data acquisition unit 101 receives learning data composed of a speech signal and a learning label representing the speech in kana characters, for example, and outputs a speech signal and a learning label (S1). When the acquisition of learning data is completed, the data acquisition completion information is transmitted to the sufficient statistic accumulation unit 106.

The feature amount analyzing unit 102 receives the speech signal, extracts the speech feature amount series O = (o ₁ , o ₂ ,..., O _t ,..., O _T ) and outputs the extracted speech features (S2). As the feature quantity to be extracted, for example, dynamic parameters such as 1 to 12 dimensions of MFCC (Mel-Frequency Cepstrum Coefficient) and ΔMFCC which is a change amount thereof, power, Δ power, and the like are used. Also, CMN (cepstrum average normalization) processing may be performed. Note that the feature amount is not limited to MFCC or power, and any parameter used for speech recognition may be used.

  The state series conversion unit 103 receives the learning label and the acoustic model, decomposes the learning label into a phoneme series, further converts it into a state series, and outputs it (S3). For example, if the learning label is "Hello", "pause ko ng ni ch iha pause" decomposed phoneme sequence as further which "* -pause + k [1] * -pause + k [2 ] * -pause + k [3] pause-ko [1] ... " [1] [2] [3] are state numbers of the phoneme model.

The acoustic model storage unit 104 includes an initial state distribution π _i , a state transition probability a _ij from the state i to the state j (1 ≦ j ≦ J, J is the number of states in the model), and the output probability of the state j at time t.

(M is the number of basis normal distributions (mixed number) belonging to state j, N is the basis normal distribution, c _jk is the mixture weight coefficient for the kth basis normal distribution of state j, μ _jk is the average vector, and U _jk is the common vector. An acoustic model λ including a dispersion matrix is stored. FIG. 3 conceptually illustrates the acoustic model λ. The acoustic model λ is HMM (Hidden Markov Models), and includes a plurality of phoneme models. The phoneme model may be a monophone (for example, * -a + *) or a triphone (for example, k-a + i). Each phoneme model is composed of, for example, three states, state transition probabilities are defined within and between each state, and the output probability of each state is a mixed normal distribution composed of a plurality of (three in FIG. 3) basis normal distributions. Is defined.

The forward calculation unit 105 receives the speech feature series O and the state series, performs forward calculation using the acoustic model λ corresponding to the state series, and has a forward probability δ _t (i) and a back pointer ψ _t (i). A forward calculation history consisting of is output (S4). The back pointer ψ _t (i) is used to store the state of the transition source for the state sequence that gives the highest probability among the state sequences that reach state i at time t. For example, the forward calculation can be performed as follows.

(i) Initialization δ ₁ (i) = π _i · b _i (o ₁ ) (1 ≦ i ≦ J, J is the number of states in the model) (2)
ψ ₁ (i) = 0 (3)

(ii) Forward calculation

(iii) End

The sufficient statistic accumulation unit 106 obtains the maximum likelihood path q _t ^* from the input forward calculation history, and stores the sufficient statistic until the processing of all learning data acquired while following the maximum likelihood path q _t ^* is completed. Accumulation is performed (S5). The maximum likelihood path q _t ^* can be obtained by the following equation using a back pointer.

q _t ^* = ψ _{t + 1} (q _{t + 1} ^* ) (t = T−1, T−2,..., 1) (8)
The sufficient statistic includes a statistic used for model update, for example, a weighting factor, an average vector, a covariance vector, and the like of each mixed element of a mixed normal distribution of output probabilities, which can be calculated from the following equations, respectively.

In the Viterbi algorithm, the posterior random variable γ _t (j) in equation (12) is

である。

It is.

The process of obtaining the maximum likelihood path using the Viterbi algorithm by the forward calculation unit 105 and the sufficient statistic accumulation unit 106 will be specifically described with reference to FIGS. 4 (a) and 4 (b), the horizontal axis represents time (corresponding to the speech feature amount series), the vertical axis represents the state (corresponding to the phoneme series), and each circle (●, ◎, ○) represents a certain time. represents the state i at t. First, as shown in FIG. 4A, the forward calculation unit 105 starts from a certain state at t = 1 (state 1 in FIG. 4A) as a starting point, and the forward probability δ _t in a plurality of states at t = 2. Calculate In FIG. 4 (a), the forward probability δ _t has been calculated is ◎, the one that has not calculated δ _t is ◯, the one on the maximum likelihood path is ● (start state of t = 1 or t = T The end state is always on the maximum likelihood path). Also, a back pointer ψ _t is held for each calculated item. Subsequently, starting from the state in which the forward probability has been calculated at t = 2 (state 1 and state 2 in FIG. 4A), the forward probability δ _t in a plurality of states at t = 3 is calculated, and the back Holds the pointer ψ _t . Thereafter, the forward calculation history is accumulated in the same manner until t = T (the length of the speech feature amount sequence). Using the forward calculation history accumulated in this manner, the sufficient statistic accumulation unit 106 traces from time T to T-1, T-2,..., And t = 1 as shown in FIG. To obtain the maximum likelihood path q _t ^* . In equation (13), the posterior random variable γ _t of the Viterbi algorithm is 1 on the maximum likelihood path because the probability of passing through the maximum likelihood path is 1 (100%), and other than the maximum likelihood path is 0 (0%). .

  The model update unit 107 receives a sufficient statistic, constructs and outputs a post-learning acoustic model, and thereby updates the acoustic model in the acoustic model storage unit 104. Note that the post-learning acoustic model storage unit 108 may be provided, for example, without being updated as it is by the output, and may be written in there for the time being.

Kiyohiro Shikano, 4 others, "Voice Recognition System", Ohmsha, May 2001, p.18-31 L.R.Rabiner, "A tutorial on hidden Markov models and selected applications in speech recognition", PROCEEDINGS OF THE IEEE, February 1989, VOL.77, No.2, p.257-284

When using the Forward-Backward algorithm, the posterior random variable γ _t (i) is calculated, and the maximum likelihood path q _t ^*

Thus, since it is obtained by narrowing down from a certain range, a highly accurate learning effect can be expected, but since the backward calculation is performed, the processing speed is difficult.

In addition, the Viterbi algorithm is capable of high-speed processing because backward calculation is not necessary, but as can be seen from equation (13), the maximum likelihood path is believed and γ _t (i) = 1 is set to a sufficiently statistical equation. Since the calculations from (9) to (12) are performed, the learning accuracy may be degraded when there is an error in the learning data or when an ambiguous utterance or noise is superimposed.

  An object of the present invention is to realize an acoustic model creation device and the like that have higher learning accuracy than the Viterbi algorithm while ensuring high speed by the Viterbi algorithm.

  The acoustic model creation device of the present invention includes a learning data acquisition unit, a feature amount analysis unit, a state series conversion unit, an acoustic model storage unit, a forward calculation unit, a sufficient statistic accumulation unit, and a model update unit.

  The learning data acquisition unit receives learning data including a speech signal and a learning label that represents the speech in characters, and outputs the speech signal and the learning label separately.

Feature amount analyzing unit, the audio signal entered voice feature amount sequence _{O = (o 1, o 2} , ···, o t, ···, o T) and outputs the extracted.

  The acoustic model storage unit stores an acoustic model including an initial state distribution, a state transition probability from one state to another state, an output probability in a certain state, and the like.

  The state series changing unit receives the learning label and the acoustic model, decomposes the learning label into a phoneme series, further converts this into a state series, and outputs the state series.

  The forward calculation unit receives the speech feature series O and the state series, performs forward calculation using the acoustic model corresponding to the state series, and outputs a forward calculation history including a forward probability and a back pointer To do.

The sufficient statistics accumulation unit receives the forward calculation history, obtains the maximum likelihood path, and accumulates sufficient statistics while following the maximum likelihood path. Note that the value of the posterior random variable used for accumulating sufficient statistics is the state appearance reliability f _i when the maximum likelihood path passes through the state i (1 ≦ i ≦ J, where J is the number of states) at time t. (o _t ) (0 ≦ f _i (o _t ) ≦ 1) and 0 elsewhere.

  The model update unit receives the sufficient statistics, constructs and outputs a learned acoustic model, and thereby updates the acoustic model in the acoustic model storage unit.

  According to the acoustic model creation device of the present invention, it is possible to improve learning accuracy while ensuring high speed by the Viterbi algorithm.

The figure which shows the function structural example of the acoustic model production apparatuses 100 and 200. FIG. The figure which shows the example of a processing flow of the acoustic model production apparatuses 100 and 200. FIG. The figure which illustrated the acoustic model notionally. The figure which illustrated notionally the processing which obtains the maximum likelihood path by Viterbi algorithm.

  FIG. 1 shows an example of the functional configuration of an acoustic model creation apparatus 200 according to the present invention. The only difference from the acoustic model creation apparatus 100 is that the forward calculation unit 105 is replaced with the forward calculation unit 205, and the sufficient statistic accumulation unit 106 is replaced with the sufficient statistic accumulation unit 206, respectively. Since it is the same, the same code | symbol is attached | subjected about the same part and description is abbreviate | omitted.

The sufficient statistic accumulation unit 206 basically performs the same processing as the sufficient statistic accumulation unit 106. However, in the Viterbi algorithm for the equation (12), the posterior probability when the maximum likelihood path passes through the state j at time t. The variable γ _t (j) = 1 and γ _t (j) = 0 in other cases, but in the present invention, when the maximum likelihood path passes through the state j at time t, γ _t (j) = F _j (o _t ). Here, f _j (o _t ) normally takes a range of 0 ≦ f _j (o _t ) ≦ 1 as the state appearance reliability. As confidence in the state appearing reliability f _j (o _t) of time t through the state j in the present invention, may be used the S / N of a frame according to the time t in the speech for the training data, the entire utterance It is good also as S / N of. The likelihood for the learning label for the learning speech data may be used. However, the value of the state appearance reliability f _j (o _t) is used normalized values to fit the minimum 0 or up to 1 or less. For example, normalization may be performed based on the maximum value or the minimum value obtained from the S / N or likelihood distribution for the learning data. Alternatively, the maximum S / N may be 30 dB, the minimum may be 0 dB, 30 may be 1 and 0 may be 0. In the Viterbi algorithm, the posterior probability γ _t (j) on the maximum likelihood path is set to 1, whereas in the present invention, the learning speech is obtained by setting the reliability based on the S / N or likelihood of the learning speech data. Since the posterior probabilities take into account the data noise contamination and learning label errors, the accuracy of acoustic model learning can be improved.

Further, in addition to performing the same processing as the forward calculation unit 105 in the forward calculation unit 205, the output probabilities calculated by the equations (2) and (4) are stored, and f _j (o _t ) is set to the time t. the ratio of the value of the Viterbi path output probability b _j on (maximum likelihood path) (o _t) to the sum of all the output probabilities calculated in, that may be used a value determined by the equation (15).

Since the output probability b _k (o _t ) that has been calculated by the forward calculation unit 205 (the portion marked by ◎ in FIG. 4) is diverted, the learning accuracy can be improved without increasing the amount of calculation. Here, since the output probability b _j (o _t ) of the state j on the maximum likelihood path at the time t of the numerator is always included in the denominator, the state appearance reliability f _j (o _t ) is always a value of 1 or less and is normal It becomes a converted value. Etc. If the learning label is incorrect, is not consistent state of the state j in the feature quantity o _t and maximum likelihood path at time t, the value of b _j (o _t) for the output probability of the state j is reduced . In this case, the difference between the output probability of the non-state j is smaller becomes state appearance reliability f _j (o _t) is a small value, it is possible to suppress the influence of the learning data has an error in the learning label.

Further, equation (16) may be applied instead of equation (15).

F _j (o _t) of formula (16) means a value of the ratio of the Viterbi path (maximum likelihood path) on the output probability b _j (o _t) to the sum of all the output probabilities belonging to the monophone at time t. In this case, unlike the equation (15), it is necessary to calculate a part that has not been calculated (a circle in FIG. 4). However, since the number of monophone states is smaller than the total number of states, the amount of calculation does not increase greatly, and the learning accuracy can be improved effectively. Note that the likelihood of a monophone corresponding to a state on the Viterbi path may be used (ex. K-a + i [1] → * -a + * [1]). At this time, calculation of state appearing reliability f _j (o _t) is calculated from all monophones becomes 1 or less stable value because it contains the value of the molecule in the denominator. Here, since normalization is performed with the output probability of the state already calculated in the forward calculation in Equation (15), the denominator is a value based only on the state depending on the learning label, but in Equation (16), it is always a monophone. belongs because it is calculated by the sum of the output probabilities of all states, it is possible to reduce the dependence on learning label, therefore, is the value of the steady state appeared reliability f _j (o _t) to improve the learning accuracy be able to.

  As described above, according to the acoustic model creation device of the present invention, it is possible to improve learning accuracy while ensuring high speed by the Viterbi algorithm.

  When the configuration of the utterance direction estimation device of each of the above embodiments is realized by a computer, the processing contents of the functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer. In this case, at least a part of the processing content may be realized by hardware.

  In addition, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.

Claims (7)

Learning data consisting of a speech signal and a learning label that represents the speech in characters are input, a learning data acquisition unit that separates and outputs the speech signal and the learning label,
A feature amount analysis unit that receives the speech signal and extracts and outputs a speech feature amount sequence O = (o ₁ , o ₂ ,..., O _t ,..., O _T );
An acoustic model storage unit that stores an acoustic model including an initial state distribution, a state transition probability from one state to another state, an output probability in a certain state, and the like;
A state sequence conversion unit that receives the learning label and the acoustic model, decomposes the learning label into a phoneme sequence, further converts this into a state sequence, and outputs the state sequence;
A forward calculation unit that receives the speech feature amount sequence O and the state sequence, performs forward calculation using the acoustic model corresponding to the state sequence, and outputs a forward calculation history including a forward probability and a back pointer; ,
A sufficient statistics accumulation unit that receives the forward calculation history, obtains a maximum likelihood path, and accumulates sufficient statistics while following the maximum likelihood path; and
The sufficient statistics are input, a model update unit that builds and outputs an acoustic model after learning, and thereby updates the acoustic model in the acoustic model storage unit, and
An acoustic model creation device comprising:
The value of the posterior random variable used when the sufficient statistics amount is accumulated in the sufficient statistics accumulation unit is as follows when the maximum likelihood path passes through the state i (1 ≦ i ≦ J, where J is the number of states) at time t. An acoustic model creation device characterized by a state appearance probability f _i (o _t ) (0 ≦ f _i (o _t ) ≦ 1) and 0 in other places. The acoustic model creation device according to claim 1,
The forward calculation unit calculates and stores the output probability b _j (o _t ) of the state j (1 ≦ j ≦ J, where J is the number of states) at time t in the course of the forward calculation,
The state appearance probability f _i (o _t ) is

(B _i (o _t ) is an output probability on the maximum likelihood path). The acoustic model creation device according to claim 1,
The forward calculation unit calculates and stores an output probability b _j (o _t ) of the state j (1 ≦ j ≦ J, where J is the number of states) belonging to the monophone at time t in the forward calculation process,
The state appearance probability f _i (o _t ) is

(B _i (o _t ) is an output probability on the maximum likelihood path), and the state i belongs to a monophone. A learning data acquisition step for separating learning data composed of a speech signal and a learning label that represents the speech in characters;
A feature amount analyzing step of extracting a speech feature amount series O = (o ₁ , o ₂ ,..., O _t ,..., O _T ) from the speech signal;
A state sequence conversion step of decomposing the learning label into a phoneme sequence and further converting it into a state sequence;
A forward calculation step of performing forward calculation using the acoustic model corresponding to the state series from the voice feature amount series O and the state series, and outputting a forward calculation history including a forward probability and a back pointer;
A sufficient statistics accumulation step for obtaining a maximum likelihood path from the forward calculation history and accumulating sufficient statistics while following the maximum likelihood path;
A model update step of constructing and outputting an acoustic model after learning from the sufficient statistics and thereby updating the acoustic model;
An acoustic model creation method for executing
The value of the posterior random variable used when accumulating the sufficient statistics in the sufficient statistics accumulation step is as follows when the maximum likelihood path passes through the state i (1 ≦ i ≦ J, where J is the number of states) at time t. A method of creating an acoustic model, characterized by state appearance probability f _i (o _t ) (0 ≦ f _i (o _t ) ≦ 1) and 0 in other locations. The acoustic model creation method according to claim 4,
The forward calculation step calculates and stores the output probability b _j (o _t ) of the state j (1 ≦ j ≦ J, where J is the number of states) at time t in the course of the forward calculation.
The state appearance probability f _i (o _t ) is

(B _i (o _t ) is an output probability on the maximum likelihood path). The acoustic model creation method according to claim 4,
The forward calculation step calculates and stores the output probability b _j (o _t ) of the state j (1 ≦ j ≦ J, where J is the number of states) belonging to the monophone at time t in the forward calculation process,
The state appearance probability f _i (o _t ) is

(B _i (o _t ) is an output probability on the maximum likelihood path), and the state i belongs to a monophone.   A program for causing a computer to function as the apparatus according to claim 1.

Images