JP2853731B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device.

[0002]

2. Description of the Related Art Since a speech recognition apparatus requires a very large amount of calculation, attempts have been made to reduce the amount of calculation by beam search. As a method of setting the beam width for the candidate pruning by the beam search, at the time of each candidate pruning, a method of leaving a fixed number of candidates from those having a high likelihood,
A method of leaving a candidate having a likelihood within a certain range from the maximum likelihood is well known.

In a paper entitled "Beam Search in Continuous Speech Recognition" published by Ito et al. It is reported that the search efficiency is stable with respect to a change in the set value of the width. In addition, the number of candidates having a likelihood within a certain range from the maximum likelihood is considered to fluctuate under the influence of the surrounding noise environment at the time of utterance. There is no change in the number of candidates.

[0004]

However, in the above-described method of leaving a fixed number of candidates (assuming that the number is M), it is necessary to perform a rearrangement process for obtaining the M-th candidate at the time of candidate pruning. However, there is a disadvantage that the amount of processing is large.

[0005]

According to the first aspect of the present invention, the difference between the maximum likelihood and the likelihood of a certain number of candidates in the first several frames of the input speech is determined, and thereafter, The present invention provides a speech recognition apparatus characterized in that a threshold value for pruning candidates in a beam search is set using the difference.

According to the second aspect of the present invention, a feature extracting unit for analyzing a speech signal and outputting a feature vector time series,
A standard pattern storage unit that stores a standard pattern created in advance, a cumulative likelihood storage unit that holds a cumulative likelihood, and a time series of the cumulative likelihood and the feature vector stored in the cumulative likelihood storage unit. A recurrence formula calculation unit for obtaining a new cumulative likelihood from the standard pattern; and a certain number of the cumulative likelihoods obtained by the recurrence formula calculation unit for a certain partial sequence of the feature vector time series. Is output, and the difference between the maximum cumulative likelihood and the certain number of likelihoods is stored, and for subsequent subsequences, the difference is calculated using the difference between the stored likelihoods. A cumulative likelihood output unit that determines a cumulative likelihood to be output according to the threshold value, and a result output unit that obtains a recognition result for the speech signal from the cumulative likelihood output from the cumulative likelihood output unit. To obtain a speech recognition device .

According to the third aspect of the invention, the average value of the difference between the cumulative likelihood of the M-th candidate and the maximum cumulative likelihood of each of the arbitrary number of partial sequences of the input speech is determined. , Between the next partial series, a threshold setting method in the speech recognition apparatus, wherein a threshold for candidate pruning is set using the average value of the differences obtained in the previous partial series.

[0008]

Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention. Referring to FIG. 1, an embodiment of the present invention includes a feature extraction unit 101, a standard pattern storage unit 102, a cumulative likelihood storage unit 103, a recurrence formula calculation unit 104, a cumulative likelihood output unit 105, And an output unit 106.

The feature extraction unit 101 converts a speech input into a time series of feature vectors, and outputs the result to a recurrence formula calculation unit 104. The standard pattern storage unit 102 stores a standard pattern. The cumulative likelihood storage unit 103 stores the cumulative likelihood output from the cumulative likelihood output unit 105. Before the processing is started, the initial value of the cumulative likelihood for all recognition path candidates is set to 1.
Holds 0. The recurrence formula calculation unit 104 obtains the cumulative likelihood up to the i-th frame from the feature vector of the i-th frame, the standard pattern, and the cumulative likelihood up to the (i−1) -th frame. The cumulative likelihood output unit 105 selects, from the input set of cumulative likelihoods, one used for calculating the cumulative likelihood of the next frame, and outputs it to the cumulative likelihood storage unit 103. The result output unit 106 outputs a recognition result based on the accumulated likelihood up to the last frame.

Next, the operation of this embodiment will be described with reference to FIGS.

The input speech is converted into a feature vector representing a spectrum of the frequency of the speech at regular time intervals in a feature extraction unit 101, and a recurrence formula calculation unit 1 is provided.
04 is output. This fixed time interval is hereinafter referred to as a frame. In the i-th frame, the recurrence formula calculating unit 10
In 4, the reference pattern _{REF = {R 1, ...,} R N} , which is held in the standard pattern storage section 102, wherein _{R w = {r w (1} ), ..., r w (J w)} using Then, the local likelihood l _w (i, j) (w = 1,..., N, j = 1,..., J _w ) of the feature vector of the current frame with respect to each standard pattern is obtained. Here, N is the number of standard patterns, and J _w is the frame length of the w-th standard pattern. Next, the local likelihood and the i-th i-
1 cumulative likelihood set of frames _{G = {g 1 (i-} 1,1), ..., g 1 (i-1, J 1), ..., g N (i-1, 1), ..., g N ( From i−1, J _N ) 求 め る, a recognition path candidate of the current frame and its cumulative likelihood are obtained as Equation 1 below by maximization processing based on dynamic programming (Step 1 in FIG. 2).

[0013]

(Equation 1) The cumulative likelihood output unit 105 determines the M-th cumulative likelihood from the maximum value if i ≦ K in comparison with a predetermined K, and sets the Mth cumulative likelihood as a threshold TH for candidate pruning. The difference d between this and the maximum likelihood is determined. The accumulated value S _d of d is updated as S _d = S _d + d to obtain an average later (steps 2, 6, and 7).

Note that S _d is initialized to 0 before the first frame.

If i = K, an average D = S _d / K of the difference between the maximum likelihood between K frames and the candidate pruning threshold is determined (step 3).

When i> K, the threshold value TH for candidate pruning is set to TH = g _max -D. g _max is
This is the maximum value of the cumulative likelihood in the i-th frame (step 4).

In each frame, the cumulative likelihood output unit 10
5 outputs only the recognition path candidates having the likelihood larger than the threshold value TH of the cumulative likelihood to the cumulative likelihood storage unit (step 5).

If the current frame is the last frame,
The cumulative likelihood output unit 105 outputs to the result output unit 106 all recognition path candidates that have reached the end point of the standard pattern.
The result output unit 106 determines the recognition path candidate having the largest cumulative candidate and outputs the recognition result (steps 8 and 9).

As described above, in the present embodiment, the first K
In the frame, has been described by way of obtaining the average value of the difference between the maximum cumulative likelihood of accumulated likelihood of the M-position candidate, more generally, any L _max number of partial series l of the input speech
For each of i ₁ , li ₂ ,..., li _Lmax (L _max ≧ 1) (these subsequences are temporarily referred to as learning intervals),
The average value of the learning section li _k and the next learning section l
Between i _{k + 1} , a method of setting a threshold for candidate pruning using the average value of the differences obtained by _l i _k can be used.

[0020]

As described above, the speech recognition apparatus according to the present invention provides the maximum likelihood of the M-th candidate and the maximum likelihood corresponding to the fluctuation of the surrounding noise environment and the change of the set value of the beam width M. Even when the difference between the accumulated likelihoods fluctuates greatly, the average value of this difference is obtained using a part of the input, and the pruning threshold is determined using this. Thus, the pruning of candidates according to the method of leaving candidates up to the M-th cumulative likelihood is performed, and there is an effect that the processing amount for determining the pruning threshold does not increase while maintaining stable search efficiency.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of a speech recognition device of the present invention.

FIG. 2 is a flowchart showing a processing flow of an embodiment of the voice recognition device shown in FIG. 1;

[Explanation of symbols]

 Reference Signs List 101 Feature extraction unit 102 Standard pattern storage unit 103 Cumulative likelihood storage unit 104 Recurrence formula calculation unit 105 Cumulative likelihood output unit 106 Result output unit

Continuation of the front page (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 561 G10L 5/06 JICST file (JOIS)

Claims (3)

(57) [Claims] In the first several frames of input speech, the difference between the maximum likelihood and the likelihood of a certain number of candidates is determined,
Thereafter, using the difference, a threshold value for pruning candidates in a beam search is set. 2. A feature extraction unit that analyzes a speech signal and outputs a feature vector time series, a standard pattern storage unit that stores a standard pattern created in advance, and a cumulative likelihood storage unit that stores cumulative likelihood. A recurrence formula calculating unit for calculating a new cumulative likelihood from the cumulative likelihood stored in the cumulative likelihood storage unit, the time series of the feature vector, and the standard pattern; and For the subsequence, a constant number of cumulative likelihoods determined by the recurrence formula calculation unit is output, and the difference between the maximum cumulative likelihood and the constant number likelihood is accumulated. For the subsequent subsequences, a cumulative likelihood output unit that determines the cumulative likelihood to be output, based on a threshold obtained using the accumulated likelihood difference, and an output from the cumulative likelihood output unit. From the accumulated likelihood That the recognition result speech recognition apparatus characterized by having a a determined result output unit. 3. An arbitrary number of threshold learning sections having an input voice.
Of the maximum likelihood and the likelihood of a certain number of candidates
The average value of the differences is calculated, and the threshold learning section and the next threshold learning section are calculated.
A threshold setting method in a speech recognition device, wherein a threshold value of a candidate pruning is set using an average value of likelihood differences obtained in the threshold learning section between the intervals .