JPH11175087A - Character string matching method for word speech recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make the storage capacity and calculation quantity small by deciding similarity by recognizing matches between a phoneme series of respective words and phoneme symbols of the phoneme series of an input voice sequentially from the beginning of a word. SOLUTION: Word registration in a word dictionary 8 is performed with a vowel, a fricative, and a voiceless part so that vowels appearing in words are traced and a spectrum analysis of the input speech is taken; and conversion to three kinds of phoneme series is performed (12) and voiceless sounds in the beginning and ending of the word are removed. Further, discontinuous symbols are replaced with last symbols to generate a symbol sequence (13, 14), each word dictionary series, e.g. 'a*uaei' is compared with an input series 'a*u*aeie', character by character, from the beginning of the word, and when a match between, for example, 'a' and 'a' is obtained, an advance to a comparison of a next character is made, but when a mismatch between, for example, 4th 'a' and '*' is obtained, the penality is increased by one, only the input series is advanced by one character, and 4th 'a'and 5th 'a' are compared with each other.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a method for matching a dictionary with input speech when recognizing word speech on a computer or a smaller system.

[0002]

2. Description of the Related Art Speech recognition using phoneme information, that is, a dictionary of phoneme sequences is provided, a speech input signal is analyzed for each fixed frame, and each symbol of a phoneme sequence for each frame and a symbol of a word phoneme sequence in the dictionary are analyzed. There is a recognition method for calculating the similarity of the Also, the number of phoneme standard patterns called SPLIT method is
There are recognition techniques that have been increased to a number sufficient to represent spectral variations.

In the above-described speech recognition method using phoneme information, a standard phoneme pattern based on a spectrum is required. When calculating the similarity between the input voice and the standard pattern,
Since the distance between the spectra is calculated, the amount of calculation increases. Further, word speech is recognized based on the similarity sum obtained by comparing the obtained similarity with the word dictionary. As a result, the recognition rate is high with few misses, but the storage capacity and calculation amount required for recognition also increase significantly. If the word dictionary used in this method does not know about phonemes, it is difficult to create a new dictionary because the dictionary itself is a sequence of phonemes.

[0004] The SPLIT method can be said to be a combination of three conventional methods: a phoneme-based recognition system, a word-based recognition system, and a vector quantization method in speech coding. Where S
The standard pattern of the PLIT method is called a pseudophoneme standard pattern, and is obtained by clustering short-time spectral patterns of a large number of voice samples for each speaker.
Since it is based only on the distribution of the spectral pattern, there is no clear association with phonemes. Therefore, orthography cannot be used to generate a word dictionary. In addition, the storage capacity and the amount of calculation required for recognition are about 1/10 of the case of using a standard pattern in word units.

[0005] Since the SPLIT method is usually used for speech recognition of a specific speaker, matching between a vowel sequence and a word dictionary in the word speech recognition method for each phoneme will be described with reference to FIG. . The input voice from the input terminal 1 passes through the A / D converter 2 and is converted into a digital signal. Thereafter, the spectrum is analyzed by the analysis unit 3. The analyzed spectrum is compared with the standard pattern 5 (spectrum) of each phoneme by the similarity calculation unit 4 to obtain
Get the similarity with each phoneme. The similarity between each phoneme and the passage of time is stored in the similarity matrix 6. The similarity sum calculator 7 performs DP matching with the phoneme sequence of each word in the word dictionary 8 to obtain a similarity sum.

On the basis of the sum of similarities, the word judging unit 9
Is determined to be the word having the highest similarity sum. The result is passed to the output terminal 10.

[0007]

In the conventional method described above, the number of phonemes alone is about 40, and if DP matching using a word dictionary is performed there, a considerable amount of calculation is required. In addition, the user needs to know phonemes in order to newly register words in the dictionary.

[0008]

According to the present invention, the phonemic sequence of each word in the word dictionary and the phonemic sequence of the input speech are checked sequentially from the beginning of the phonemic symbol, and if they do not match,
The penalty is increased by +1 and the matching of the phoneme of the word phoneme sequence of the dictionary with the next succeeding phoneme symbol in the phoneme sequence of the input speech is advanced. It is determined that a word having a smaller penalty has a higher similarity to the word.

In particular, each word is expressed by a total of seven symbols including vowels, fricatives, and silence, and the input speech is also converted into these seven symbol sequences, and the matching method is applied.

[0010]

FIG. 1A shows a processing procedure of a recognition apparatus to which an embodiment of the present invention is applied. Although the processing contents are different from those in FIG. 4, the same reference numerals are given to the parts where the same processing is performed from the overall flow. The input voice from the input terminal 1 passes through the A / D converter 2 and is converted into a digital signal. Thereafter, the analysis unit 3 performs an LPC analysis to extract, for example, a cepstrum.

A sequence generation unit (for each frame) 12 obtains a time sequence of phonemes by comparison with a standard pattern (cepstrum vector) 5. In this embodiment, each word is represented by vowels, fricatives, and silences. The sequence generator 12 also makes the input speech into a sequence composed of vowels, fricatives, and silences. For example, when a voice of the word “cup” is input, a time sequence of phonemes shown in FIG. 2A is obtained. * Indicates silence, a indicates a vowel of "ka", "tsu"
Is detected as silence, and u is a vowel of “P”. There is a silence * at the beginning and end of the word, and a discontinuous vowel o is detected between "a" and "tsu", and a discontinuous vowel i is detected between "tsu" and "pu".

In the time sequence of the phoneme that appears earlier in the sequence processing (A) 13, the silent part * appearing at the beginning and the end of the phoneme is deleted, and the discontinuous phoneme symbol is replaced with the previous phoneme symbol. . In the example of FIG. 2A, as shown in FIG. 2B, the silence * at the beginning and the silence * at the end are deleted, the discontinuous phonological symbol o is replaced by the symbol a immediately before it, and the discontinuous phonological symbol i is changed. Is replaced by the symbol * immediately before.

Next, in this example, in the sequence processing (B) 14, the portion where the same phoneme symbol is continuous in the sequence generated in the sequence process (A) is replaced with one character of the continuous phoneme symbol. In the example shown in FIG. 2, as shown in FIG. 2C, eight consecutive phoneme symbols a are one a, eighteen consecutive silence symbols * are one *, and eight consecutive u are One u
Is replaced by

The similarity sum calculator 7 performs matching between the series output from the series processing (B) 14 and the word dictionary 8. Here, the word determination unit 9 is performed based on the obtained similarity sum.
To find a word that seems to be the correct answer and output it from the output terminal 10. Here, the similarity sum calculating unit 7, that is, the matching unit will be described in detail. First, the word dictionary 8 is written with only the vowel parts of the word emphasized. That is, in the symbols used in the word dictionary, “a” is “A”, “i” is “Yes”, “u” is “U”, “e” is “E”, and “o” is “ “O” and “*” are “silence” and “S”
Represents "frictional sound", respectively. To give an example written in a word dictionary, as shown in FIG.
* U ”, and gowns are written in character strings such as“ Cau ”. Although “b, d, g” and “m, n” are not used in actual recognition, they are generally denoted by “C” representing consonants for convenience. For example, if the game is "CeC
u ”. If many people utter a word in the word dictionary 8, such a character string will be taken.
Think about it and register.

Further, the similarity sum calculating section 7 does not perform the DP matching as conventionally performed, but uses a simpler method which is a feature of the present invention. As understood from the word notation in the word dictionary 8, the series processing (B) 14
The series that has come out is a series that can be matched with each word character string in the word dictionary 8. In the matching method of the present invention, for example, as shown in FIG.
And the phoneme series “a *” of the input voice processed in the series processing (B) 14.
u * aei ”will be described as an example.

Step 1: The first "a" in the dictionary series is compared with the first "a" in the input series. Since the symbols are the same, both characters are advanced by one character without any operation. Step 2: The second "*" in the dictionary series is compared with the second "*" in the input series, and both strings are advanced by one character because they are the same symbol. Step 3: The third "u" and "u" of each of the two sequences are compared, and both characters are advanced by one character with the same symbol.

Step 4: Each fourth "a" of both streams is compared with "*", and since they do not match, the penalty is increased by +1.
Then, advance the input sequence by one character. Step 5: The fourth “a” in the dictionary series is compared with the fifth “a” in the input series, and both strings are advanced by one character because they match. Step 6: The fifth “e” in the dictionary series is compared with the sixth “e” in the input series, and both strings are advanced by one character because they match.

Step 7: The sixth "i" in the dictionary series is compared with the seventh "i" in the input series, and both strings are advanced by one character by matching. Step 8: The dictionary sequence ends at the sixth, and the input sequence has one character "e" remaining, so the penalty is increased by 1 due to mismatch. As described above, the comparison is sequentially performed from the beginning of each symbol of both sequences, and if both are the same, both are advanced by one character, and if they are different, the penalty is incremented by one and only the input sequence is advanced by one character. This continues until the end of either series is reached. At the end of one, if there is a series that has not been compared to the other series, add a few minutes to the penalty.

In this way, matching with all the words in the word dictionary 8 is performed, and words having a penalty of 0 or less than a predetermined value are output as recognition results. Such a character string matching method is not limited to the word notation in the word dictionary, but is applied to the notation as shown in FIG. 2B, and is applied to matching between the dictionary and the processing sequence from the word sequence processing (A) 13. Alternatively, the word dictionary may be a notation including consonants, the input speech sequence may be a phoneme sequence including consonants, and the matching method as shown in FIG. 1B may be applied to these.

[0020]

As described above, according to the present invention, since the matching method is basically only comparison and addition, the load on the CPU is small and the system can operate even with a small system. When the sequence written in the word dictionary is limited to seven types of symbols as in the previous embodiment, the dictionary can be easily updated because the word dictionary is created by tracing the vowels appearing in the word. Also, since the sequence is centered on vowels, there is no need to have special knowledge about phonemes. To basic knowledge,
If there is knowledge such as devoicing of vowels and slight mixing of "e" when transitioning from "a" to "i", the recognition rate can be further increased.

Using this matching method, one-chip DS
P, 30 subjects (15 men, 15 women)
A), the recognition rate was 90.3% for men, 94.2% for women, and 9
2.3%.

[Brief description of the drawings]

FIG. 1A is a block diagram showing a functional configuration example of a word speech recognition apparatus to which the method of the present invention is applied, and FIG. 1B is a diagram showing an example of a matching method thereof.

FIG. 2 is a view showing a processing example of series processing (A) 13 in FIG. 1A.

FIG. 3 is a view showing an example of the contents of a word dictionary 8;

FIG. 4 is a block diagram showing a functional configuration of a conventional phoneme-based word speech recognition apparatus.

Continued on the front page (72) Inventor Yoshio Nakadai 3-192-2 Nishi-Shinjuku, Shinjuku-ku, Tokyo Japan Telegraph and Telephone Corporation (72) Inventor Yoshitake Suzuki 3-192-2, Nishi-Shinjuku, Shinjuku-ku, Tokyo Japan Inside Telegraph and Telephone Corporation (72) Inventor Toshihide Kurokawa 1-3-1 Gotenyama, Musashino-shi, Tokyo NTT Advanced Technology Corporation (72) Inventor Yamato Sato 1-chome, Gotenyama, Musashino-shi, Tokyo No. 1-3 NTT Advanced Technology Co., Ltd.

Claims (3)

[Claims] 1. A character string matching method for word speech recognition that performs recognition by matching a word dictionary represented by a sequence of phonemic symbols with a time sequence of phonemes obtained by a short-time analysis of input speech. The matching of phonemic symbols between the dictionary phoneme sequence and the phoneme sequence of the input speech is sequentially confirmed from the beginning of the word. If they do not match, the penalty is increased by +1. A character string matching method for word speech recognition, characterized in that the matching with the following phonological symbols is advanced. 2. The character string matching for word speech recognition according to claim 1, wherein the input phoneme sequence and the word phoneme sequence of the word dictionary are each composed of one or more of a vowel, a fricative, and a silent part. Law. 3. A silent part at the beginning and end of an input phoneme sequence is removed, and a phoneme of a transient discontinuous part is replaced with a phoneme immediately before.
3. A character string matching method for word speech recognition according to claim 2, wherein said matching is performed with each word of the word dictionary after replacing a portion where the same phoneme is continuous with the one phoneme.