JPH0715638B2 - Syllable pattern cutting device - Google Patents

Description

The present invention relates to an improvement of a syllable pattern cutout device in a voice input device.

<Prior Art> Conventionally, in a device that converts voice into a character string, such as a voice word processor, a syllable pattern is cut out from an input voice pattern in advance and registered as a syllable standard pattern, and the syllable standard pattern and the input voice are cut out. There is a method of recognizing a syllable based on the distance between the syllable and the test pattern. At that time, when the syllable pattern is cut out (segmentation), the syllable section is cut out from the voiced section by using the acoustic feature parameters such as the power of the input voice and the spectrum change. However, if the syllable section is cut out using only the above-mentioned characteristic parameters, the syllable pattern cutout may be inaccurate due to the influence of noise and articulation. In that case, an incorrect syllable standard pattern is created, which greatly affects the syllable recognition performance.

Therefore, there has been proposed a voice word processor which prevents an erroneous syllable standard pattern from being created by echoing back a voice waveform corresponding to a syllabic segment to be cut out, outputting the voice, and having the operator confirm the voice waveform.

<Problems to be Solved by the Invention> As described above, in the conventional syllable pattern cutout method, the voice waveform corresponding to the syllable pattern cutout section is echoed back to confirm the syllable section (that is, the syllable boundary position) to be cut out. This prevents the operator from making an incorrect syllable standard pattern.

However, in the above syllable pattern cutout method, there is a problem that if the operator is not careful, the echo back corresponding to the syllable pattern cutout section may be missed. Further, there is a problem that a time is required for echoing back after each utterance, and a registration work takes time.

Therefore, an object of the present invention is to provide a "syllable pattern cutout device" which can easily and accurately determine a syllable boundary position without requiring output of echo back voice.

<Means for Solving the Problem> In order to achieve the above object, the syllable pattern cutout device of the present invention includes a speech analysis unit that extracts a characteristic parameter of an input speech and detects a section length of a phoneme. A syllable boundary position detection unit for detecting a syllable boundary position candidate of the input speech based on the above, and an average syllable length estimation for obtaining an estimated average syllable length based on the speech section length of a word whose utterance content is known and the number of syllables forming the word. Part, a vowel consonant vowel sequence generating part that generates a vowel consonant vowel sequence forming the known word, and the vowel within a predetermined range determined based on the estimated average syllable length for the syllable boundary position candidate. A syllable boundary position candidate evaluation unit that evaluates the existence ratio of the phoneme section length corresponding to a consonant vowel string, and an evaluation result of the existence ratio for each pair of the syllable boundary position candidate and the vowel consonant vowel string. The result is used as the DP matching grid point value, and the DP matching is performed using the difference between the length between the syllable boundary position candidates and the estimated average syllable length as the weight of the path connecting the grid points. It is characterized by further comprising a syllable boundary determining unit that determines a syllable boundary from the syllable boundary position candidate.

<Operation> The speech analysis unit extracts the characteristic parameter of the input speech and detects the section length of the phoneme, and the syllable boundary position detection unit detects the syllable boundary position candidate of the input speech based on this characteristic parameter. On the other hand, the estimated average syllable length is obtained by the average syllable length estimation unit on the basis of the voice section length of a word whose utterance content is known and the number of syllables forming the word. A vowel consonant vowel sequence forming a known word is generated by a vowel consonant vowel sequence generator, and a phoneme sequence extracted by matching with a phoneme standard pattern for an input voice is referred to, and the syllable boundary position candidate is described above. The existence ratio of the phoneme section length corresponding to the above vowel consonant vowel sequence within the predetermined range determined based on the estimated average syllable length is evaluated by the syllable boundary position candidate evaluation unit, and the result is taken as the reliability. At the time of DP matching of correspondence between the syllable boundary position candidate and the vowel consonant vowel sequence, the evaluation result of the existence ratio is set as a grid point value of DP matching, and the length between the syllable boundary position candidates and the estimated average syllable length. Is used as the weight of the path connecting the grid points, and the syllable boundary is determined from the syllable boundary position candidate from the DP matching result. The syllable boundary section thus determined is cut out as a syllable pattern.

That is, the operator does not need to confirm the cut out syllable section by echo back or the like, and can automatically and correctly determine the syllable boundary.

<Example> Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments.

FIG. 1 shows a block diagram of a syllable clipping device according to the present invention. This syllable extraction device inputs a voice and the Roman alphabet of a word (syllable) of this voice, and outputs a syllable pattern (time series of cepstrum coefficients) from the input voice, which is performed by the following procedure. .

That is, when the syllable standard pattern is registered by the syllable segmentation device, information that cannot be used in normal speech recognition is top-downed because the utterance content is known in advance by the Roman alphabet input. Can be used. First, since the number of syllables included in the input voice is known in advance from the Roman alphabet input, the estimated average syllable length can be obtained by dividing the voice section length by the number of syllables. Next, a syllable boundary is determined based on the estimated average syllable length from among syllable boundary position candidates detected by spectrum change, power change, and the like. In that case, as the reliability used in the determination, the vowel of the preceding syllable, the consonant of the succeeding syllable, and the phoneme type corresponding to the vowel of the succeeding syllable within a certain range (estimated average syllable length × constant) before and after the syllable boundary position are used. And the length between syllable boundary position candidates (syllable length)
And a value based on the difference between the average syllable length and the average syllable length.

Next, an outline of the syllable cutting device will be described with reference to FIG.

A frame (cycle of 8 m
For each s), characteristic parameters such as powers and cepstrum coefficients, and phoneme classification symbols and phoneme symbols which will be described in detail later are obtained. Then, the syllable boundary detection unit 2 obtains candidates for syllable boundary positions from the characteristic parameters, phoneme symbols, phoneme symbols, and the like. On the other hand, a VCV sequence (vowel-consonant-vowel sequence) that constitutes the word is obtained by the VCV generation unit 3 from the input Romanization of the word. Then, the VCV spotter 4 calculates the reliability for each pair of the syllable boundary position candidate and the VCV sequence, and the DP (dynamic programming) unit 5 calculates the syllable boundary position candidate and the above-mentioned reliability.
The pair having the highest reliability among the pairs with the VCV string is searched for, and the syllable boundary determining unit 6 cuts out the corresponding section as a syllable pattern.

Hereinafter, each part of the syllable cutting device will be described in detail.

(1) Speech analysis unit 1 Obtains characteristic parameters such as linear prediction (LPC) cepstrum, power, and differential power from the speech input from the microphone under the conditions shown in Table 1.

Using this feature parameter, the rough phonological features of each frame are symbolized into six types of symbols (hereinafter referred to as phonological classification symbols) as shown in Table 2 and output. Furthermore, phonemes automatically extracted from isolated single syllables (5 vowels and / n /, / s /)
By matching each frame with the standard pattern, the fourth
The phoneme symbol string as shown in the figure is output. Further, the following two types of segmentation parameters are calculated from the above characteristic parameters.

・ Power dip ... Primary differential coefficient of power.

Spectral change: Cepstrum coefficient between frames separated by 8 frames (or 4 frames).

(2) Syllable boundary detecting unit 2 A syllable boundary position candidate is obtained using the phonological classification symbol, power dip and spectrum change obtained by the speech analyzing unit 1 (see Table 3).

Boundary symbols “(” and “)” mean silence → voiced, voiced → voiceless, voiceless → voiced and voiced →
It is attached to the frame corresponding to each change point of unvoiced sound. Further, the boundary symbol "P" is attached to a frame where the power dip is large, and the boundary symbol "s" is attached to a frame where the spectrum change is large.

(3) Average Mora Length Estimator 7 From the number of moras M obtained from the Roman alphabet input and the speech section length LT between the boundary symbols “(” and “)” detected by the syllable boundary detector 2, Obtain the estimated average mora length LM.

The initial value of the estimated average mora length LM is 21 frames (speech rate 6
(Corresponding to mora / second), and is updated for each word utterance. If the speech section length at the time of uttering a word consisting of M-mora is LT frame, it can be expressed as L = LT / M when the average mora length is L frame. If the range of the average mora length L is 16 ≦ L <31, LM = L. Further, in consideration of the case where the voice section length LT is erroneously detected due to noise or bad utterance, if L <16 or 31 ≦ L, LM = (21 + L) / 2.

(4) VCV generation unit 3 When the romanized notation of the word input by voice is input, the VCV sequence forming the word is obtained. VCV above
Is composed of 3 pairs of vowel V2 of the preceding syllable, consonant C1 of the succeeding syllable and vowel V1 of the succeeding syllable for each syllable boundary. For a word consisting of n syllables, n + 1 VCV strings including the boundary with the silent interval are generated. (See FIG. 2).

However, the part corresponding to C1 of the syllable consisting only of vowels is V1.
It is represented by the same symbol as in, but the symbol for the syllable part in the case of a syllable consisting of consonant + syllable + vowel is omitted.

(5) VCV spotter 4 For each pair of the syllable boundary position candidate detected by the syllable boundary detecting unit 2 and the V2C1V1 generated by the VCV generating unit 3, the reliability D as a syllable boundary is as follows. And ask. However, the smaller the value of the reliability D described here, the higher the possibility that it is a syllable boundary position.

First, according to Table 4, the above-mentioned phoneme symbol string (phoneme 1 and phoneme 2: written in 9 uppercase letters) from the speech analysis unit 1 to be searched for is calculated based on the lowercase letters of the alphabet constituting V2C1V1. Here, phoneme 1 and phoneme 2 mean that there are two phonemes to be searched.

The position of the syllable boundary position candidate is t frames, and the interval (t-1M /
2) to t, the number of phoneme frames corresponding to V2 is c (V2), and the number of phoneme frames corresponding to C1 from section t to (t + lM / 2) is c (C1), section (t + LM / Let c (V1) be the number of phoneme frames corresponding to V1 from 2) to (t + LM). Then, the reliability D is calculated as follows with reference to the phonemes 1 and 2 calculated from Table 4 as described above.

When C1 does not output "B" D ← LM-c (V2) -c (C1) When C1 is "B" D ← LM-c (V2) -c (V1) C1 is "B" or When "S", D ← LM-c (V2) -MAX (c (C1 = "S"), c (V1)). However, for MAX (,), select the larger value among the values in (,).

Further, if the kind of syllable boundary position candidate is “(”, D ← D-LM / 2 “(”, D ← D-LM × 2/3 If “p”, D ← D-LM / 3 .

(6) DP unit 5 The reliability D (i,
j) using the i-th syllable boundary position candidate and the j-th V2C
DP matching is performed for each pair with 1V1 to find the syllable boundary position taking the path with the smallest cumulative distance G (i, j) (see FIG. 3). However, both end points are free to remove noise such as breath that may be present at the beginning and end of the word. Here, the cumulative distance G (i, j) at the point (i, j) is expressed by the equation (1).

Here, S (i, k) represents the deviation of the frame length (corresponding to the syllable length) between the syllable boundary position candidates from the estimated average mora length LM, and the i-th syllable boundary position candidate is t (i) frames. Then, S (i, k) is expressed as in equation (2).

S (i, k) = || t (i) -t (k) | -LM | (2) That is, based on the assumption that each syllable length does not vary much within a word (syllable) speech, the syllable length is The difference between and the estimated average mora length is used as a weight in DP matching.

(7) Syllable boundary determination unit 6 The section between the syllable boundary position candidates obtained by the DP unit 5 is cut out as a syllable pattern (see FIG. 4). At this time,
When the following conditions are satisfied, the syllable cutout from the word voice is rejected. Therefore, in the case of registration of the syllable standard pattern, it is necessary to re-voice the rejected word and input it by voice.

-Includes short syllables whose syllable length is LM / 4 or less.

-Includes long syllables whose syllable length is 2LM or more.

-Includes an uncertain syllable boundary whose reliability D is LM / 2 or more.

Here, the phoneme first candidate and the phoneme second candidate in FIG. 4 are candidates selected in ascending order of matching distance in the matching with the standard pattern based on the isolated monosyllable performed in the speech analysis unit 1.

The above is the configuration and basic algorithm of the syllable extraction device. However, in order to improve performance so that it can deal with phonetic rules and vocalization fluctuations, some exceptional rules are added as follows.

Unvoiced In the above V2C1V1, if C1 is an unvoiced consonant and V1 is “I” and “U”, the phoneme corresponding to V1 at the syllable boundary position of the V2C1V1 of the input speech is a silent “.”. I'll decide to go.

Vowel chain In V2C1V1 above, if V2 is "A" and C1 is "I", the syllable boundary position of the V2C1V1 pair of the input voice
The phoneme corresponding to C1 may be "E".

When the initial syllables / i /, / hi /, / hu / and / tsu / are at the beginning of a word, the frame length between candidate syllable boundary positions tends to be shorter.
Therefore, the deviation S (i, k) of the frame length of the syllable boundary position candidate section used when the cumulative distance G (i, k) is calculated in the DP unit 5 from the average mora length LM is expressed by the following equation. change.

S (i, k) = || t (i) -t (k) | -LM / 2 | Suffix When the syllables / i / and / N / are at the end, the frame length between syllable boundary positions becomes shorter. Tend. Therefore, the shift S (i, k) is changed as in the following equation.

S (i, k) = || t (i) -t (k) | -LM / 2 | Thus, according to the present invention, the characteristic parameter is obtained from the input speech, and the syllable boundary of the input speech is based on this characteristic parameter. While obtaining the position candidate, obtain the estimated average syllable length from the input speech, and based on this estimated average syllable length and the number of frames of V2, C1 and V1, obtain the reliability of the above syllable boundary position candidate as a syllable boundary. The syllable boundary position is determined based on the reliability and the syllable pattern is cut out.
Therefore, the syllable pattern can be accurately extracted regardless of the difference in the vocalization speed of the input voice due to the characteristics of the operator.

Further, according to the present invention, since it is not necessary to echo back the voice waveform corresponding to the syllable cut-out section and have the operator confirm it, the burden on the operator and the time required for the registration work can be reduced. Therefore, the syllable boundary position can be determined easily and accurately.

In the above embodiment, the romanization is input when inputting the utterance content of the voice input word, but the present invention is not limited to this, and a kana may be input.

In the above embodiment, the present invention is used when the standard pattern is registered. However, the present invention is not limited to this, and can be used, for example, when recognizing an input voice. At that time, contrary to the case of the above-described embodiment, the VCV generation unit 3 uses the phoneme strings of the plurality of obtained recognition candidates as the utterance content input of the word that is voice input.
, And a recognition candidate matching the syllable boundary of the input voice may be output as a recognition result.

<Effects of the Invention> As is clear from the above, the syllable pattern cutout device of the present invention extracts a characteristic parameter of an input speech and detects a phoneme section length, and an input speech based on the characteristic parameter. A syllable boundary position detecting unit for detecting a candidate syllable boundary position, an average syllable length estimating unit for obtaining an estimated average syllable length based on a voice section length of a word whose utterance content is known and the number of syllables forming the word, A vowel consonant vowel sequence generating unit that generates a vowel consonant vowel sequence that constitutes a known word, and a vowel consonant vowel sequence within a predetermined range determined based on the estimated average syllable length for the syllable boundary position candidate. For each pair of the syllable boundary position candidate evaluation unit that evaluates the existence ratio of the corresponding phoneme section length, and the evaluation result of the existence ratio for each pair of the syllable boundary position candidate and the vowel consonant vowel sequence, a DP match And the difference between the length between the syllable boundary position candidates and the estimated average syllable length is used as the weight of the path connecting the grid points, and DP matching is performed. Since the syllable boundary determining unit for determining the syllable boundary from the candidate is provided, the syllable boundary position candidate having the highest reliability as the syllable boundary position candidate among the detected syllable boundary position candidates is automatically determined as the syllable boundary. be able to. Therefore, according to the present invention, it is possible to easily and accurately determine the syllable boundary without outputting the echo back voice.

[Brief description of drawings]

FIG. 1 is a block diagram of a syllable segmentation device according to the present invention, FIG. 2 is a diagram showing an example of generating a VCV sequence, FIG. 3 is a diagram showing an example of a DP matching path in syllable boundary determination, and FIG.
The figure is a diagram showing an example of syllable segmentation. 1 ... Voice analysis unit, 2 ... syllable boundary detection unit, 3 ... VCV
Generation unit, 4 ... VCV spotter, 5 ... DP unit, 6 ... syllable boundary determination unit, 7 ... average mora length estimation unit.

Claims (1)

[Claims] 1. A speech analysis unit for extracting a characteristic parameter of an input speech and detecting a section length of a phoneme; a syllable boundary position detection unit for detecting a syllable boundary position candidate of the input speech based on the characteristic parameter; An average syllable length estimation unit that obtains an estimated average syllable length based on the speech interval length of a word whose content is known and the number of syllables that make up the word, and a vowel consonant vowel that generates a vowel consonant vowel sequence that makes up the known word. A string generation unit and a syllable boundary position candidate that evaluates the existence ratio of the section length of the phoneme corresponding to the vowel consonant vowel string within a predetermined range determined based on the estimated average syllable length candidate for the syllable boundary position candidate. An evaluation unit, for each pair of the syllable boundary position candidate and the vowel consonant vowel sequence,
DP in which the evaluation result of the existence ratio is a DP matching grid point value, and the difference between the length between the syllable boundary position candidates and the estimated average syllable length is the weight of the path connecting the grid points.
A syllable pattern cutout device comprising a syllable boundary determining unit that performs matching and determines a syllable boundary from the syllable boundary position candidate based on the DP matching result.