JPH0756597B2 - Voice recognizer - Google Patents

Description

Description: [Object of the Invention] (Industrial field of application) The present invention relates to a speech recognition apparatus for improving the recognition accuracy of words by continuous speech.

(Prior Art) In recognition of continuous speech, a recognition target is continuously input word by word, and thus a technique of extracting a phoneme structure constituting the recognition information from input information is important.

In this regard, in the conventional continuous speech recognition, the input speech signal is divided into fixed analysis time intervals called frames, feature information such as spectrum is extracted for each frame, and phoneme is labeled for each frame. In this way, phoneme segmentation is automatically performed while integrating frames with the same label. Then, the frame length is set to be relatively short and fixed, for example, 16 msec and 20 msec, in order to identify phonemes such as plosive consonants having a short duration.

However, if the frame length is set too short, the label will change due to power fluctuations due to the influence of external noise, etc., making it impossible to automate the above-mentioned segmentation, and the so-called “crossover”, which is the point of phoneme change.
Since the template for the part (1) is not prepared, there arises a problem that the frame including the "crossover" cannot be recognized.

This problem can be solved by making the frame length longer, but if the frame length is increased,
The recognition rate of phonemes with short duration such as plosive consonants will be reduced.

(Problems to be Solved by the Invention) As described above, in the conventional continuous speech recognition, if the frame length is set to be too short, it becomes difficult to automate the segmentation due to the fluctuation of the speech power, or the phonological change point becomes difficult. There was a problem that the "crossover" part could not be recognized, and if the length of the frame was set long, it became impossible to recognize the phoneme with a short duration.

Therefore, an object of the present invention is to suppress fluctuations in voice power and effects on phoneme recognition of "overflow" portions without lowering the recognition rate of phonemes with short durations, and thus excellent in recognition accuracy and automatic segmentation. To provide a voice recognition device capable of

[Configuration of the Invention] (Means for Solving Problems) A voice recognition device according to the present invention analyzes a voice signal input for each frame of a certain length and extracts feature information for each frame. From the plurality of consecutive frames having the same information amount as the feature information, which is obtained by inputting the feature information of each frame from the feature extracting means and calculated based on the respective feature information of the plurality of consecutive frames. Parallel conversion means for generating feature information about a plurality of frame groups for a plurality of frame groups having different numbers of frames, and feature information for each frame extracted by the feature extraction means of a frame group consisting of one frame In addition to inputting each as information, the feature information of each frame group generated by the parallel conversion means is input respectively, and the feature information for each frame group and The phoneme recognition means for performing parallel matching with each phoneme dictionary created in parallel and assigning a phoneme level to each frame group based on the respective matching results, and a sequence of phoneme labels provided by the phoneme recognition means. And a word recognition result obtained by collating with a preset word dictionary, and a phonological label provided for word recognition among the plurality of phoneme labels is based on the collation result by the phonological unit or the result of the word recognition. It is characterized by comprising a word recognition means for selecting by selecting.

Preferably, the word recognition means performs word recognition using a phoneme label of a frame group having a highest similarity to a phoneme dictionary among a plurality of phoneme labels obtained in parallel by the phoneme recognition means. It is characterized by being.

Preferably, the word recognition means has the highest similarity to the word dictionary when word recognition is performed using these phoneme labels among the plurality of phoneme labels obtained in parallel by the phoneme recognition means. It is characterized in that it outputs a word recognition result using a phoneme label of a frame group.

Preferably, the word recognition means, among the plurality of phoneme labels obtained in parallel by the phoneme recognition means, the phoneme label of the frame group closest to the phoneme duration predicted based on the recognition result by the phoneme recognition means. It is characterized by performing word recognition using.

(Operation) According to the present invention, a plurality of phoneme recognitions having different frame lengths are performed in parallel at each time. Then, word recognition is performed using the phoneme label of the frame group having the highest score (similarity, distance, etc.) among the obtained phoneme labels. For this reason, the frame length is equivalently changed adaptively according to the phoneme data, and the word recognition rate can be significantly increased compared to the conventional method in which the frame length is fixed to a fixed length. .

Embodiments Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a diagram showing the configuration of a continuous speech recognition apparatus according to an embodiment of the present invention.

The A / D conversion unit 1 converts a voice signal input from a voice input unit (not shown) or a public telephone line into a digital signal of about 12 bits at a sampling rate of about 12 KHz and outputs the digital signal to the feature extraction unit 2. .

The feature extraction unit 2 analyzes the input digital audio signal for each unit frame length fixed to, for example, about 16 msec and extracts the feature. The feature extracting unit 2 can be configured by a band pass filter group (filter bank) when, for example, spectrum analysis is used for feature extraction. The feature vector extracted here is the parallel conversion unit 3
Entered in.

The parallel conversion unit 3 divides the input feature vector for each unit frame into n systems, and has n−1 stages of delay circuits 3a.
And n−1 adders 3b for adding the output of each delay circuit 3a and the input feature vector, and these adders 3b
Is divided by the number of added frames, and n-1 coefficient circuits 3c. The n types of feature vectors input to the phoneme recognition unit 4 via these n systems are weighted averages with different numbers of added frames. The smaller the number of added frames, the more explosive consonants, etc. It carries phonological feature information with a short duration, and as the number of added frames increases, it carries phonological information in which the effects of power fluctuations and “crossovers” are suppressed by past steady phonological feature vectors. . These feature vectors are input to the phoneme recognition unit 4 in parallel.

The phoneme recognition unit 4 collates the input n types of feature vectors with the phoneme dictionary, and assigns a phoneme label to each feature vector. Similarity calculation with the phonological dictionary,
For example, the composite similarity method can be used. Further, instead of the similarity, the distance to the template of the dictionary may be obtained. These n kinds of phoneme labels are stored in the phoneme label buffer 6 together with their scores (the similarity range is distance).
Is temporarily stored in.

In addition to this, the phoneme label buffer 6 also stores previously generated feature vectors. Phonological label buffer 6
From these feature vectors, as shown in FIG.
A total of 2n-1 types of weighted average feature vectors over the past and future are selected and output to the label sorting unit 7 centering on time t-n before n-1 frames.

The label sorting unit 7 sorts the scores of these 2n-1 types of feature vectors (sorts in descending order).
Thereby, for example, the highest phoneme label having the highest degree of similarity is output to the word recognition unit 8 together with the score as a phoneme label at time t-n.

The word recognition unit 8 collates the input phoneme label sequence with the word dictionary 9. For example, the known DP (dyna
mic programming) Matching is done. This D
As shown in FIG. 3, the P matching takes an input phoneme sequence on the horizontal axis and a network-represented standard pattern on the vertical axis, and one frame at a time from the start point, the sum of the degrees of similarity with each phoneme label forming the standard pattern. The path that maximizes is selected. At time t, the optimal paths for all nodes of the standard pattern are calculated, and in the next frame, the scores are calculated for all paths permitted by the network of the standard pattern, and the optimal path at time t + 1 is calculated. It Then, the standard pattern having the maximum sum of similarities is output as a recognition result.

According to the speech recognition apparatus of the present embodiment having the above-described configuration, the phonological label having the frame length with the highest degree of similarity among the n types of analysis frame lengths is adopted at each time. The length can be adaptively changed according to the phoneme data.

For example, in FIG. 2, when the voice data obtained in the unit frame at time t-n has a short duration such as a plosive sound, the score of the phoneme label obtained in the shortest frame length is At the maximum, this phoneme label is adopted as the label for that frame.

On the other hand, when the voice data obtained at the time t-n is the "crossover" portion which is the change point between different phonemes, the time t
Steady phoneme data is collected before and after -n. Therefore, if the weighted average of a plurality of past or future frames centered on the time t-n is taken, the influence of the "overflow" portion is suppressed according to the number of weighted frames.
Therefore, in this case, the score of the phoneme label obtained by the long frame becomes large, and this is adopted as the recognition result.

Further, even when there is fluctuation in the voice power, the influence of this fluctuation can be suppressed by taking the weighted average with other frames, so that the score of the phoneme label in a relatively long frame becomes high.

By the way, in the recognition of continuous speech, the automation of segmentation is a major technical issue, but this device absorbs fluctuations in speech power by adaptively changing the analysis frame length according to the phoneme data. Since it is possible to prevent the label from changing due to erroneous recognition of the phoneme, it is highly likely that the same label is given in the same phoneme section. Therefore, there is an advantage that the automation of the segmentation is facilitated by performing the integration processing of the same label.

FIG. 4 is a diagram showing a configuration of voice recognition according to another embodiment of the present invention. In this figure, the same parts as those in FIG. 1 are designated by the same reference numerals, and the description of the overlapping parts will be omitted. This device is different from the previous device in the word recognition unit 11. The word recognition unit 11 parallelizes the degree of similarity with the standard pattern in the word dictionary 9 for the phoneme labels up to the times t−1, t−2, ..., T−n output from the phoneme recognition unit 4. Calculate to. The obtained DP matching results (scores) up to time points t-1, t-2, ..., Tn are held in the word DP matching score buffer 12.

DP matching is a method of calculating the optimum paths for all the nodes at a certain time, but in this embodiment, as shown in FIG.
The result of DP matching and the phoneme label for one frame, the result of the DP matching at time t-2 and the phoneme label of the pattern averaged over two frames (score is twice), ..., At time t-n
N types of matching are performed, such as the result of DP matching and the phoneme label of the averaged pattern of n frames (score is n times). As a result, the optimum path can be dynamically obtained while dynamically obtaining the optimum frame length until time t. After repeating such matching, the pattern having the maximum sum of similarities at the end point becomes the recognition result and is sent to the display unit 7.

In this method, DP matching is performed n times for each frame, but since this is performed in parallel, the calculation speed does not decrease.

FIG. 5 shows still another embodiment. In this device, a phoneme duration control unit 21 is newly provided in parallel with the word recognition unit 11 in the device of FIG. The phoneme continuation control unit 21 obtains the phoneme duration based on the label obtained by the analysis of the unit frame length in the phoneme recognition unit 4, and performs the word recognition for performing the DP matching of the frame length of the duration corresponding to each label. This is a system for sending to section 11.

According to this device, when a phoneme label whose duration is expected to be short, such as “p” or “t”, is obtained by the analysis of the unit frame length, the DP matching result up to time t−1 is obtained. The optimal path is obtained using the phoneme label for one frame, and conversely, for a phoneme that is expected to have a long duration such as a vowel, the DP matching result up to time t-n and the phoneme label for n frames. The optimal path is obtained using and. That is, a non-stationary phoneme with a short duration is recognized with a short frame length, and a stationary phoneme with a long duration is recognized with a long frame length. The duration setting in the phoneme duration control unit 21 may be set based on the duration probability distribution for each phoneme empirically obtained.

[Effects of the Invention] As described above, according to the present invention, various lengths as phoneme analysis frame lengths are prepared, and parallel phoneme recognition processing is performed for each of these frame lengths for recognition. Since the best frame length of the result is adopted, the analysis frame length can be adaptively changed according to the input phoneme. For this reason, it is possible to suppress fluctuations in the speech power and the effects on the phoneme recognition in the “overflow” part without lowering the recognition rate of the phoneme with a short duration. Therefore, the speech recognition is excellent in the recognition accuracy and capable of automatic segmentation. A device can be provided.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a speech recognition apparatus according to an embodiment of the present invention, FIG. 2 is a diagram for explaining the contents of a phoneme label buffer in the apparatus, and FIG. 3 is word recognition in the apparatus. FIG. 4 is a block diagram showing the configuration of a speech recognition apparatus according to another embodiment of the present invention, and FIG. 5 is a diagram for explaining the operation of the word recognition section in the apparatus. FIG. 6 is a block diagram showing the configuration of a voice recognition device according to still another embodiment of the present invention. 1 ... A / D converter, 2 ... Feature extractor, 3 ... Parallel converter, 4
... Phoneme recognition unit, 5 ... Phoneme dictionary, 6 ... Phoneme label buffer, 7 ... Label sort unit, 8, 11 ... Word recognition unit, 9 ... Word dictionary, 12 ... Word DP matching score buffer, 21 ... Phoneme duration control unit. .

Claims (4)

[Claims] 1. A feature extraction unit that analyzes an input audio signal for each frame of a certain length and extracts feature information for each frame, and the feature information for each frame is input from the feature extraction unit and continues. A plurality of frame groups in which the number of frames is different from the feature information regarding the frame group composed of the plurality of consecutive frames having the same amount of information as the feature information calculated based on the respective feature information of the plurality of frames. And parallel input means for generating the parallel conversion means for inputting the feature information for each frame extracted by the feature extraction means as feature information of a frame group consisting of one frame, and for each frame group generated by the parallel conversion means. The feature information is input respectively, and the feature information of each frame group and the predetermined phoneme dictionary are collated in parallel, and the respective collation results are obtained. A phoneme recognition means for giving a phoneme label to each frame group based on the above, a sequence of phoneme labels given by this phoneme recognition means and a preset word dictionary are collated to obtain word recognition results, respectively, and A speech recognition device, comprising: a word recognition unit that selects a phoneme label to be used for word recognition from among a plurality of phoneme labels based on a matching result by the phoneme recognition unit or a result of the word recognition. 2. The word recognition means performs word recognition using a phoneme label of a frame group having a highest similarity to a phoneme dictionary among a plurality of phoneme labels obtained in parallel by the phoneme recognition means. The speech recognition apparatus according to claim 1, wherein the speech recognition apparatus is a speech recognition apparatus. 3. The word recognizing means has the highest similarity to a word dictionary when word recognition is performed using these phonological labels among a plurality of phonological labels obtained in parallel by the phonological recognizing means. The speech recognition apparatus according to claim 1, wherein the speech recognition device outputs a word recognition result using a phoneme label of a higher frame group. 4. The word recognition means selects a frame group closest to a phoneme duration estimated based on a recognition result of the phoneme recognition means from among a plurality of phoneme labels obtained in parallel by the phoneme recognition means. The speech recognition apparatus according to claim 1, wherein word recognition is performed using a phoneme label.