JPS6344699A - Voice recognition equipment - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION "Field of Industrial Application" The present invention relates to a speech recognition device that recognizes input speech in a top-down manner using phonemes or syllables as recognition units.

"Prior Art" In this type of speech recognition device, input speech signals are stored in memory as a time series of characteristic parameters such as power and sictor, while words to be recognized are stored in dictionary memory as a series of phonemes or syllables. During recognition, a word in the dictionary is assumed, and the length, level, etc. are determined according to predetermined rules for each phoneme or syllable according to the phoneme or syllable sequence of the word. Determine whether the input audio signal can be divided correspondingly, and if the corresponding phoneme or syllable can be extracted, the degree of correspondence (similarity)
The next phoneme or syllable is extracted by assigning a value indicating , segmentation and scoring are performed, and if the input speech signal can be segmented by the assumed word, the scores are totaled, and this is the predetermined value. If the above is the case, the input audio signal is recognized as the assumed word.

In conventional top-down speech recognition devices that use phonemes or syllables as recognition units, when the phoneme is a vowel, the continuation length of a phoneme is dealt with by simply giving a fixed duration (Acoustical Society of Japan Study group material S-82-
62 (DEC, 20, 1982,)), if the content or state of the utterance by the speaker is not limited, there is a drawback that the difference from the fixed duration becomes large, leading to a decrease in the recognition rate.

In addition, there are large differences in the duration of consonants between consonant groups such as voiceless plosives and voiceless fricatives (edited by IEICE: "Hearing and Speech", 3rd edition, 1982). There is also the drawback that the recognition rate decreases when the duration is used.

In order to solve these drawbacks, conventional speech recognition devices have relaxed restrictions on the duration of phonemes or syllables without using specific determination means. For this reason, even if phonemes or syllables with clearly different duration lengths are assumed for the speech signal, the assumed words are difficult to remove and the amount of processing increases.

Furthermore, since processing is performed sequentially starting from the first phoneme or syllable, even if the number of phonemes or syllables in a word assumed to be in a speech signal is clearly smaller or larger than the speech length of the input speech signal, the problem may occur. Since processing is performed without considering appropriateness, there is a drawback that the amount of processing increases.

This invention has the disadvantage of a decrease in the recognition rate due to giving fixed duration lengths to vowels and consonants, and the disadvantage of an increase in the amount of processing due to loosening restrictions on duration lengths without using specific determination means. In order to eliminate this problem, we compare the estimated speech length from a hypothetical word with the actual speech length for the input speech signal based on statistical data on the duration of individual phonemes or syllables or groups. The purpose of this invention is to provide a speech recognition device that actively utilizes time-related features by performing statistical determination.

``Means for Solving the Problems'' This invention utilizes the temporal characteristics of speech, which have conventionally been used auxiliary or avoided, as the duration of phonemes or syllables, in a speech recognition device that uses phonemes or syllables as recognition units. The most important feature is that

That is, statistical data on the duration of phonemes or syllables is obtained for each phoneme or syllable individually or for each group and stored in a duration memory, and statistical data on the duration of each phoneme or syllable is calculated for each phoneme or syllable according to the phoneme or syllable sequence of a hypothetical word. When segmentation is completed, the input speech length from then on and the estimated speech length obtained from the statistical data of the duration of each unprocessed phoneme or syllable in the hypothetical word are calculated, and these speech lengths and the estimated speech length are calculated. If the difference is greater than a predetermined value, for example, the estimation error expected based on the variance of the statistical data of the duration of each phoneme or syllable, subsequent phoneme or syllable recognition processing (segmentation) to reduce unnecessary processing.

"Example" FIG. 1 shows an embodiment of the invention.

An audio signal inputted from an input terminal 11 is converted into a sidigital signal by a φ converter 12, and the digital signal is converted into a feature/cerameter time series by a feature extraction section 13. The output of this feature extractor 13 is stored in a time series memory 14.

The recognition processing unit 15 reads the phoneme or syllable sequence of the word assumed for the input speech signal from the dictionary memory 16, and performs recognition processing on the input speech signal in units of phonemes or syllables.
In other words, segmentation and scoring are performed (second
Figure step si). If the recognition process for one phoneme or syllable is completed and the hypothetical word is not renoected (
Step S2), the audio length of the unprocessed part of the input audio signal (
L-Lp) is calculated (Step S3), and the statistical data of duration lengths from the duration memory 17 are assumed for subsequent phonemes or syllables, and the estimated speech length ΣLk given by the sum of them is calculated (step S3). 7'S,).

The speech length of this unprocessed speech signal portion is compared with the estimated speech length, and if the difference is greater than the estimation error E expected from the variance of the statistical data of phoneme or syllable duration length, subsequent phonemes or terminate the recognition process for the syllable and
In the following cases, the process returns to step S1 to continue the recognition process (step SS).

FIG. 3 shows a conceptual diagram when processing is performed from the beginning to the p-th phoneme or syllable and the processing is discontinued at this point.

The discontinuation determination at this time is performed when the following formula is satisfied.

L: Audio length of the audio signal Lp: Audio length up to the point when the p-th phoneme or syllable is processed Lk: Statistical data of the duration of the k-th phoneme or syllable (k>p) Nl: Audio signal Using the assumed number of phonemes or parts of the word E: estimation error σl: variance of the i-th phoneme), FIG.
From dictionary memory 16/aizuwakam
It shows how the process is aborted when assuming atu, /.

That is, the input speech signal /abashiri/ is shown by the curve 21, and the hypothetical word /aizawakamatu
There are 5 segmentation candidates: the first part of / is two types of /a/■■ and three types of /a/ combination ■■(. Because /a/ at the beginning of the hypothetical word is a vowel. A recognition process is performed to determine whether the vowel rise* exists in the input audio signal, and then, first, regarding the first candidate ■/,
/ recognition processing is performed, and the input audio signal is segmented by /a/. At this time, the speech length of the remaining part of the input speech /b -i / and the estimated speech length of the remaining part of the hypothetical word /i-u/ are compared, and since the latter is longer than the estimation error E, , the recognition process for this candidate ■ is aborted, and the process moves on to the recognition process for the next candidate ■. Similarly, for candidate ■, the recognition process is terminated after segmentation for /a/. For candidate ■■, /a1/ is tentatively associated with /a/ or /a b / in the input audio signal, and then the remaining speech length is compared with the estimated remaining speech length, and the recognition process is aborted. Indicates the condition. Also, candidate ■ba /a1/ is /a in the input audio signal.
/a i/ is short, so
The difference between the speech length of the remaining part /b '' i / of the input speech signal at that time and the estimated speech length of the remaining υ part /z−u/ of the hypothetical word is less than the estimation error E, and therefore the next / Moving on to the recognition process for Z/, this is recognized as /b/ in the input audio signal.
By comparing the remaining speech length with the estimated speech length, it is determined that the difference is greater than the estimation error E, and this recognition process is terminated.

The part marked with a thick line in the figure is the part where the recognition process is discontinued according to the present invention, but in the conventional apparatus, the recognition process is performed on this thick line part, and the recognition process is discontinued at the subsequent X mark. In other words, according to the present invention, the processed noodles are less than the conventional method by the amount of processing indicated by each thick line.

As is clear from the above results, the conventional technology? In comparison, the device of the present invention can reduce the amount of processing and allow for estimation errors in consideration of statistical variations in the duration of each phoneme or syllable.1. ), it is possible to reduce the risk of inadvertently aborting the process, and it is possible to reduce the reduction in recognition rate due to the use of phoneme or syllable duration.

"Effects of the Invention" As explained above, by applying the present invention, the temporal characteristics of phonemes or syllables can be determined by 1. Based on this, it is possible to make a judgment by comparing the estimated speech length of the word assumed in the speech signal with the input speech length of the actual speech signal.k) , it has the advantage of reducing the amount of processing required for hypothetical words consisting of sequences of phonemes or syllables that clearly differ in number or type.

As a result of performing recognition using the present invention on the above-mentioned phonemes for each phoneme, it is approximately 20% when using a 100 word dictionary memory capable of recognizing 100 words, and 1% when using a 100 word dictionary memory capable of recognizing 1000 words.
When using a 000 word dictionary memory, the processing amount could be reduced by about 30%.

Furthermore, in the comparison between the estimated speech length and the input speech length according to the present invention, an estimation error is given that takes estimation errors into consideration.
It also has the advantage of suppressing the decline in recognition rate due to the use of statistical data on the duration of phonemes or syllables.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is a flowchart showing the operation of the main part of the recognition processing operation, Fig. 3 is a conceptual diagram when determining whether to discontinue processing by estimation, and Fig. 4
The figure is a diagram showing how processing is actually terminated according to the present invention. Patent Applicant Nippon Telegraph and Telephone Corporation Representative Taku Kusano 2 = 1 7 Earth Power Irregularity 37 Adhesion R1 272 Figure 4

Claims (1)

[Claims] (1) The input speech signal is stored in a time series memory as a time series of its feature parameters, a word is assumed from a dictionary memory that stores words to be recognized, and the phoneme or syllable sequence of that word is followed. In a speech recognition device that performs segmentation of the parameter time series of an input speech signal for each phoneme or syllable and performs scoring for each segmentation for recognition, statistical data of the duration length of individual phonemes or syllables or groups thereof is stored. speech length calculation means for calculating the remaining speech length of the input speech signal that has not yet been segmented for each segmentation; and statistics of duration lengths corresponding to the phonemes or syllables in the assumed word. Estimated speech length calculation means for reading data from the continuous length memory and calculating the estimated speech length of a sequence of phonemes or syllables remaining after the phoneme or syllable for which segmentation has been completed in the assumed word; and a determination means for comparing the estimated voice length calculated by the voice calculation means with the voice length calculated by the voice calculation means and determining whether or not to terminate recognition processing for the hypothetical word. Device.