JPS63104098A - 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] [Object of the Invention] (Industrial Application Field) The present invention is directed to D P (dynamic program
[Oing] The present invention relates to a speech recognition device that performs word recognition processing by matching.

(Prior art) In word speech recognition, conventionally, the phonological sequence of an input signal and a standard pattern are matched while absorbing fluctuations in the time direction so that the distance between the two is minimized.
P matching is often used as a method with a relatively high recognition rate. However, in this DP matching, for example, 2
There are two standard word patterns AB and A'CB', and both patterns have similar phonological sequences (A and A', B and B').
), when AB is pronounced, the cumulative score of the matching result often remains almost the same, and there is a problem in that the recognition rate decreases due to this cannot be ignored.

Therefore, attempts have been made to perform DP matching that takes phoneme duration into consideration, but according to this automaton control, the minimum distance is determined from multiple matching bus candidates in which the original duration length of each phoneme is shifted within the overall bus. Since the buses are selected one by one, the recognition accuracy is good compared to normal DP matching, but the number of matching bus candidates increases several times, and the recognition process takes time. Therefore, this method has the drawback that it is limited to word recognition for a few tens of words at most, and that it takes an extremely long time for recognition processing when the word size becomes hundreds or thousands.

(Problems to be Solved by the Invention) As described above, the conventional speech recognition device using DP matching has the drawback that it cannot perform highly accurate word recognition processing at high speed.

SUMMARY OF THE INVENTION The present invention has been made in view of these problems, and an object of the present invention is to provide a speech recognition device that can perform word recognition processing with high precision and at high speed.

[Structure of the Invention] (Means for Solving the Problems) The present invention includes a feature extraction section that analyzes an input audio signal at predetermined time intervals and extracts features, and a feature extraction section that analyzes input audio signals at predetermined time intervals to extract features. A phoneme recognition unit that compares feature information with a predetermined phoneme dictionary to obtain a phoneme sequence, and a word recognition unit that compares the phoneme sequence obtained by this phoneme recognition unit with a preset word dictionary to obtain word recognition results. The speech recognition device is characterized in that the word recognition section is divided into a preliminary Q selection section and a main selection section as described below.

That is, the preliminary selection unit matches the phoneme system obtained by the phoneme recognition unit with the standard phoneme sequence of a word dictionary set in advance by high-speed DP matching, and selects a case where the matching result is good and expected for each phoneme. A plurality of words having a small difference in duration length from each other are selected as word recognition candidates. Further, the main selection unit compares the standard phoneme sequence of the word recognition candidate selected by the preliminary selection unit with the phoneme sequence of the input voice by highly accurate DP matching that takes phoneme duration into consideration, and obtains a word recognition result. It becomes.

(Function) In the present invention, the preliminary selection section performs high-speed DP matching to obtain matching results for each word in approximately real time, and among those with good matching results, the preliminary selection section performs post-processing for each phoneme. Words with a small deviation from the original duration of the phoneme are selected as word recognition candidates, so the candidates in this selection section are candidates that take into account the phoneme duration to some extent, and the number of words is significantly It will be narrowed down. Therefore, even if highly accurate DP matching is performed in consideration of phoneme duration in this selection section, recognition processing will not take a significant amount of time due to the small number of word candidates.

(Example) Hereinafter, one example of the present invention will be described based on the drawings.

FIG. 1 is a diagram showing the configuration of a speech recognition processing device according to this embodiment. The A/D conversion unit 1 performs A/D conversion on a voice signal input from, for example, a voice input device or a public telephone line (not shown) at predetermined analysis time intervals (frames). The band-pass filter group (BPS) 2 is a feature extraction unit that spectrally analyzes the input speech, extracts the output value of each filter as a feature vector X, stores it in an internal pattern memory, and It is output to the recognition unit 3 and segment management unit 4. The phoneme recognition unit 3 calculates the similarity or distance between the pattern vector X obtained at each analysis time interval and the learning pattern of each phoneme category in the phoneme dictionary 5, and calculates the score for each phoneme label for each frame. seek. As a method for calculating the similarity, the well-known Euclidean distance, the well-known statistical distance as a pattern recognition technique, the composite similarity method, etc. can be used.

Note that here, 1 frame (usually 10~2゜lll5
Although a score for each phoneme can be obtained for each frame, it takes time to perform DP matching, which will be described later, for each frame. Therefore, the segment management unit 4 considers that the section in which the change in the feature vector input from the BPF 2, that is, the spectral change, is less than a certain threshold value belongs to the same phonetic section, and treats it as one segment, and performs segmentation. It has the function of automating the process.

The word recognition unit 6 provided in the latter stage inputs the obtained score series for each phoneme label as one phoneme series,
A word recognition result is obtained by comparing it with a standard phoneme sequence stored in a word dictionary, and this is output. This word recognition section 6 is composed of a preliminary selection section 8 and a main selection section 9. The preliminary selection section 6 further includes a preliminary recognition section 10 and a post-processing section 11. The preliminary recognition unit 10 obtains a score for each word in real time by referring to the word dictionary 7 through high-speed DP matching. The post-processing unit 11 limits the number of word recognition candidates by excluding words that have a large difference in duration from the expected duration of each phoneme based on the DP matching pass in the preliminary recognition unit 10. Through this post-processing, the subsequent main selection section 9 is provided with word recognition candidates in which the phoneme duration length is taken into consideration to some extent. The main selection unit 9 obtains a word recognition result from the input word recognition candidates by performing highly accurate DP matching in consideration of the phoneme duration length, and outputs the result.

In the speech recognition device configured as above, now,
The input word W is a phoneme sequence [P l r P2 .

..., PrLl, and the score of each phoneme for each segment output from the phoneme recognition unit 3 to the preliminary selection unit 10 is assumed to be Si (P). Note that P is a category name and i is a segment number.

The preliminary selection unit 10 selects the bus with the maximum score S of W between the standard pattern and the input cover turn for a certain word while absorbing temporal fluctuations by P matching as follows, and selects the bus with the maximum score S of W between the standard pattern and the input cover turn for a certain word, Score. In order to maximize the score S ofW for each word, the optimal path can be selected from the following recurrence formula.

Sof'W(t,j) With this recursion formula, there is only a bus that goes up a staircase one step at a time, but for example, if the standard pattern is a bus that uses phonemes included in a word as nodes and allows transitions between phonemes. If it is expressed as a network with edges, it is possible to form a bus that skips one stage (Fig. 3, 21).

The post-processing unit 11 calculates word recognition candidates by considering path distortion in the obtained score for each word (FIGS. 22 to 33). For example, if you match the standard pattern of the phoneme sequence "HAI" to the voice input "HAI", the result will be "H".

Each of the phonemes “A” and “I” has a relatively long phoneme duration and is easy to recognize, so as shown in Figure 2 (a), the bass that gives the maximum score (for example, 1+) is The path is long in the horizontal direction.On the other hand, the phonetic sequence "IAI" is given as the input pattern, and the standard pattern "HACI" is given as the input pattern.
If we consider the case of comparing with “C0,” there is no phoneme that corresponds to the introductory pattern, so as shown in Figure (b),
The path t2 in the "C" portion becomes shorter. As described above, the bus distortion dist is calculated, for example, as follows, taking into consideration the path lengths expected in advance for each phoneme. Note that here, dist-0 has the largest distortion, and dist-1 has the smallest distortion.

■ A, I, U, E, O, N, X (n) 2M.

Sounds such as C and S are relatively long phonemes that are easy to pronounce, so if these phonemes are included in the bus (23),
The file reads the shortest duration D1 (24), and if the duration d of the phoneme is even shorter than Dl, the shortest duration D1 is read (24).
5), the maximum distortion is dist-0 (26).

■ P, T, ni, etc. are relatively long phonemes that are difficult to pronounce, so if these phonemes are included in the path (27),
The file reads the maximum duration D2 (28), and if the duration d of the phoneme is longer than Dl, the maximum duration D2 is read (28).
9), dist −0 (26).

■ In one word, the mora (syllable) length is considered to be approximately constant. Therefore, the average time and variance of mora within a word are calculated from the path (30). If there are n moras and the length of each mora is ml, calculate dist-1-r〒π-n)''/n/m (3
1). However, since long vowels are generally twice as long as short vowels, an appropriate bias is applied to make the long vowel section longer than it actually is.

Furthermore, for a word with one mora, in the above description, the mora can be replaced with a phoneme, and the distortion can be calculated in the same way.

In this way, the distortion score (2
6, 31) and the DP matching score (32) are added together with appropriate weighting (w) (33), the obtained score is used as the score for each word, and the top few words are used for word recognition. It is output to the main selection section 9 as a candidate.

The main selection unit 9 reads standard patterns from the word dictionary 7 for only the input word recognition candidates, and matches these standard patterns with human cover patterns by DP matching. This matching is performed by taking phoneme duration into consideration. That is,
Assuming that the phoneme duration length of a certain phoneme category A is To, the optimal path at time t for the human kataan is determined by taking into account all the two paths before and after time t. Although the calculation time is several times that of DP matching in the preliminary selection section 10, the recognition rate is significantly improved.

In this device, the number of word candidates for the time-consuming DP matching in the main selection section 9 is greatly limited in the preliminary selection section, so that word recognition processing can be performed at high speed and with high precision.

[Effects of the Invention] As explained above, according to the present invention, by dividing the word recognition section into the preliminary selection section and the main selection section, the number of candidates in the main selection section, which has good recognition accuracy but takes time, can be reduced. It is possible to perform word recognition processing at high speed and with high accuracy.

[Brief explanation of the drawing]

FIG. 1 shows a speech recognition device 1 according to an embodiment of the present invention.
A/D converter, 2... band pass filter group, 3...
- Phonological recognition unit, 4... Segment management unit, 5... Phonological dictionary, 6... Word recognition unit, 7... Word dictionary, 8.
...Preliminary selection section, 9.. Main selection section, 10.. Preliminary recognition section, 11.. Post-processing section. Applicant's agent Patent attorney Takehiko Suzue AI Input

Claims (1)

[Claims] A feature extraction unit that analyzes an input audio signal at predetermined time intervals and extracts features, and a feature information extracted by the feature extraction unit that compares with a predetermined phonetic dictionary. A speech recognition device comprising: a phoneme recognition unit that obtains a phoneme sequence of the input voice; and a word recognition unit that compares the phoneme sequence found by the phoneme recognition unit with a preset word dictionary to obtain a word recognition result. In the above, the word recognition unit matches the phoneme system obtained by the phoneme recognition unit with a standard phoneme sequence in a word dictionary set in advance by high-speed DP matching, and if the matching result is good and the prediction is correct for each phoneme. a preliminary selection unit that selects a plurality of words having a small difference in duration length from the input speech as word recognition candidates; A speech recognition device comprising: a main selection unit that performs matching by highly accurate DP matching in consideration of duration and obtains a word recognition result.