JPS6315296A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to a speech recognition device that improves the recognition accuracy of words using continuous speech.

(Prior Art) In continuous speech recognition, recognition targets are input continuously in word units, and therefore a technique for extracting the phonological structure constituting the recognition target from input information is important.

Regarding this point, in conventional continuous speech recognition, the input speech signal is divided into fixed analysis time intervals called frames, and feature information such as spectrum is extracted for each frame, and phonological labels are attached to each frame. This was done to automatically segment tonal intervals while integrating frames with the same label. The frame length is, for example, 16 m5ec, 201 in order to identify phonemes with short durations such as plosive consonants.
It was set relatively short and fixed, such as sec.

However, if the frame length is set too short, the labels will change due to power fluctuations due to the influence of external noise, making it impossible to automate the segmentation described above. Since a template for the part is not prepared, a problem arises in that a frame including this "crossing °" cannot be recognized.

This problem can be solved by increasing the frame length to a certain extent, but if the frame length is increased,
The recognition rate of short-duration phonemes such as plosive consonants will decrease.

(Problems to be Solved by the Invention) As described above, in conventional continuous speech recognition, if the frame length is set too short, it may become difficult to automate segmentation due to fluctuations in speech power, and There was a problem in that the "crossing" part could not be recognized, and if the frame length was set long, it became impossible to recognize phonemes with short durations.

Therefore, it is an object of the present invention to suppress fluctuations in voice power and the influence on phoneme recognition of "crossing" parts without reducing the recognition rate of phonemes with short durations, thereby making it possible to suppress r! An object of the present invention is to provide a speech recognition device that has excellent recognition accuracy and is capable of automatic segmentation.

[Structure of the Invention] (Means for Solving the Problems) The audio signal is input to the feature extraction section. This feature extraction section analyzes the input audio signal for each predetermined frame and extracts features for each frame, and the extracted feature information for each frame is input to the phoneme recognition section. sound@
The recognition unit compares the input feature information with a predetermined phoneme dictionary and assigns a phoneme label to each frame based on the result of the check. In this production, the phoneme recognition section performs a plurality of phoneme recognition processes in parallel with different frame lengths on the same portion of the audio signal. The phoneme labels of each frame length obtained in parallel are introduced into the word 0 identification section.

The word recognition unit compares the phoneme label system assigned by the sound I recognition unit with a predetermined word dictionary to obtain word recognition results. Select a phonological label for the frame length.

(Operation) According to the present invention, a plurality of phoneme recognitions with different frame lengths are performed in parallel at each time.

Then, word recognition is performed using, for example, the phoneme label with the highest frame length and the highest score (similarity, distance, etc.) among the plurality of phoneme labels obtained. This means that the frame length is equivalently changed adaptively according to the phonological data, making it possible to significantly increase the word recognition rate compared to the conventional method that fixes the frame length to a constant length. .

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings.

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

The A/D conversion unit 1 converts an audio signal input from an audio input unit or a public telephone line (not shown) into a digital signal of approximately 12 bits at a sampling rate of, for example, approximately 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, which is fixed to about 16 m5ec, for example, and extracts its features. For example, when spectral analysis is used for feature extraction, the feature extraction unit 2 can be configured with a group of bandpass filters (filter punctures). The feature vector extracted here is input to the parallel conversion section 3.

The parallel conversion unit 3 divides the input feature vector for each unit frame into n systems, and includes an n-1 stage delay circuit 3.
n-1 adders 3b for adding the outputs of these delay circuits 3a and the input feature vectors, and n-1 adders 3b for dividing the outputs of these adders 3b by the number of frames to be added.
It is composed of one coefficient circuit 3C. The n types of feature vectors input to the phoneme recognition unit 4 via these n systems are obtained by taking a weighted average with different numbers of added frames. It is responsible for phonological feature information with a short IFA time,
The larger the number of frames added, the more the power fluctuations and the influence of the overlapping portion are suppressed by the past stationary phoneme feature vectors, and the phoneme information is carried. These feature vectors are input to the phoneme recognition unit 4 in parallel.

The phoneme recognition unit 4 compares the inputted n types of feature vectors with a phoneme dictionary and assigns a phoneme label to each feature vector. For example, a composite similarity method can be used to calculate the similarity with the phoneme dictionary. Further, instead of the similarity, the distance between the dictionary template and the template may be calculated.

These n types of phoneme labels are temporarily stored in the phoneme label buffer 6 along with their scores (similarities or distances).

The phoneme label buffer 76 also stores feature vectors generated in the past. sound m label buffer 6
From these feature vectors, as shown in Figure 2,
Centering on time tn, which is n-1 frames ago, n-1 types of weighted average feature vectors covering the past and future are selected and output to the forward sorting unit 7.

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

The word recognition unit 8 compares the input phoneme label sequence with the word dictionary 9. The verification may be performed using, for example, the known I) P
(dynamic progran+ming)
? done by tink. As shown in Figure 3, this DP matching takes the input phoneme sequence on the horizontal axis and the standard pattern represented by the network on the vertical axis, and performs 1 step from the starting point.
Frame by frame, the bus that has the maximum sum of similarities with each phoneme label constituting the standard pattern is selected. At time t, the optimal bus for all nodes of the standard pattern is calculated, and in the next frame, scores are calculated for every bus allowed by the network of the standard pattern, and the optimal bus at time t+i is calculated. Ru. Then, the standard pattern with the maximum similarity sum is output as the recognition result.

According to the speech recognition device according to the present embodiment configured as described above, since the phonetic label of the frame length with the highest degree of similarity among the n types of analysis frame lengths is adopted at each time, it is possible to equivalently The length can be adaptively changed according to phoneme data.

For example, in FIG. 2, if the audio data obtained in the unit frame at time t-n is of short duration, such as a plosive, then the score of the collegial label obtained with R short frame length is becomes the maximum, and this phonological label is adopted as the label for that frame.

On the other hand, if the speech data bound by time t-n is a part that is a transition point between different types of phonemes, then time t
-N is the center, and stationary phoneme data is collected before and after it. Therefore, if a weighted average of multiple past or future frames is taken with time t-n as the center, the influence of the "crossing part" is suppressed according to the number of weighted frames. The score of the phonetic label obtained by this increases, and this is adopted as the recognition result.

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

By the way, automation of segmentation is a major technical challenge in continuous speech recognition, but this device can absorb fluctuations in speech power by adaptively changing the analysis frame length according to phonetic data. Since it is possible to prevent changes in labels due to misrecognition of phonemes, there is a high probability that the same label will be assigned to the same phoneme interval.

Therefore, there is an advantage that segmentation can be easily automated by performing the same label integration process.

FIG. 4 is a diagram showing the configuration of speech recognition according to another embodiment of the present invention. For this reason, the same parts as in FIG. 1 are given the same reference numerals, and the explanation of the overlapping parts will be omitted. This device differs from the previous device in that the word recognition unit 11
It is in. This word recognition unit 11 determines the degree of similarity between the phoneme labels outputted from the sound 110 recognition unit 4 up to each time t-1, t-2, . . . Compute in parallel. Obtained time t-1, t
Is the result (score) of DP matching up to each time point of -2゜..., t-n a single IDP matching score buffer? 1
2.

DP matching is a method of calculating the optimal bus for all nodes at a certain time, but in this embodiment, as shown in FIG.
The result of DP matching at time t-2 and the phoneme label for one frame, the result of DP matching at time t-2 and the phoneme label of the pattern averaged over two frames (score is doubled),...
N types of matching are performed, such as the result of DP matching at time tn and the phoneme label of the average pattern of n frames (score is 0 times). As a result, it is possible to dynamically find an R-type bus while dynamically finding the optimal frame length up to time t. After repeating such matching, the pattern with the maximum similarity sum 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, there is no reduction in calculation speed.

FIG. 5 shows yet another embodiment. In this device, in the device shown in FIG. 3, the word! 8! A phoneme duration 111111 section 21 is newly provided in parallel with the identification section 11. This phoneme continuation control unit 21 is a unit that calculates phoneme duration based on the label obtained by analyzing the unit frame length in the sound II recognition unit 4, and performs DP matching of the frame length of the duration corresponding to each label. This is a system for sending the information to the Rvt recognition unit 11.

According to this device, when a phonetic label whose duration is expected to be short, such as p″, t II, is obtained by analyzing the unit frame length, the DP up to time t-1 is
The optimal bass is found using the matching results and phoneme labels for one frame. Conversely, for phonemes that are expected to have a long duration, such as vowels, the DP matching results up to time t-n and n The optimal bus is determined using the phoneme labels for the frames. That is, non-stationary phonemes with a short duration are recognized with a short frame length, and regular phonemes with a long duration are recognized with a long frame length. Note that the setting of the IIM time in the phoneme duration control section 21 may be determined based on the duration probability distribution for each phoneme that is determined empirically.

[Effects of the Invention] As explained above, according to the present invention, various lengths are prepared as phoneme analysis frame lengths, and for each of these frame lengths, the sound tm HI! After processing,
Since the frame length with the best VT recognition result is adopted, the analysis frame length can be adaptively changed according to the input phoneme. Therefore, without reducing the zero recognition rate of phonemes with short durations, it is possible to suppress fluctuations in vocal power and the influence on the recognition accuracy of the tone of the "crossing" part, resulting in excellent recognition accuracy and the possibility of automatic segmentation. We can provide audio ratio V& equipment.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a speech recognition device according to an embodiment of the present invention, FIG. 2 is a diagram for explaining the contents of a phoneme label buffer in the same device, and FIG. 3 is a word recognition device in the same device. FIG. 4 is a block diagram showing the configuration of a speech recognition device according to another embodiment of the present invention, and FIG. 6 are block diagrams showing the configuration of a speech recognition device according to still another example of the present invention. 1... A/D conversion section, 2... Feature extraction section, 3...
Parallel conversion unit, 4... Phoneme recognition unit, 5... Onsan dictionary,
6... Phonological label buffer, 7... Label sorting section, 8゜11... 11 word recognition section, 9... Word dictionary, 1
2...Word DP matching score buffer, 21...
・Phonological duration time system part. Applicant's representative Patent attorney Takehiko Suzue t-2n t-nt , 1, Figure 2 Figure 3 Figure 5

Claims (4)

[Claims] (1) A feature extraction unit that analyzes the input audio signal for each predetermined frame and extracts the features for each frame, and the feature information for each frame extracted by this feature extraction unit and a predetermined phonetic dictionary. and a phoneme recognition unit that assigns a phoneme label to each frame based on the comparison result, and a phoneme recognition unit that compares the system of phoneme labels assigned by this phoneme recognition unit with a preset word dictionary to obtain word recognition results. In the speech recognition device, the phoneme recognition unit performs a plurality of phoneme recognition processes with different frame lengths on the same speech signal in parallel, and the word recognition unit obtains a word recognition unit that obtains a word recognition unit. A speech recognition device characterized in that a phoneme label to be used for word recognition is selected from a plurality of phoneme labels having different frame lengths based on a result of phoneme recognition. (2) The word recognition unit performs word recognition using a phoneme label with a frame length that has the highest degree of similarity to a phoneme dictionary among the plurality of phoneme labels obtained in parallel by the phoneme recognition unit. A speech recognition device according to claim 1, characterized in that: (3) When the word recognition unit performs word recognition using these phoneme labels among the plurality of phoneme labels obtained in parallel by the phoneme recognition unit, the degree of similarity with the word dictionary is the highest. 2. The speech recognition device according to claim 1, wherein the speech recognition device outputs a word recognition result using a frame length phonetic label. (4) The word recognition unit selects a phoneme label with a frame length closest to the predicted phoneme duration based on the recognition result in the phoneme recognition unit, among the plurality of phoneme labels obtained in parallel by the phoneme recognition unit. 2. The speech recognition device according to claim 1, wherein the speech recognition device performs word recognition using.