JPH0289099A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To correctly recognize syllables without being affected by the change of the speaker' speaking speed by calculating a syllable section estimating parameter of each input voice based on a feature quantity extracted from the input voice and estimating an average syllable length of each input voice with the syllable section estimating parameter and segmenting a syllable section based on the average syllable length and the feature quantity. CONSTITUTION:When a voice is inputted, a feature quantity is extracted from this input voice by a feature extracting part 3. The syllable section estimating parameter of each input voice is calculated in an average syllable length estimating part 4 based on this extracted feature quantity, and the average syllable length of each input voice is estimated based on this calculated syllable section estimating parameter. A syllable section is segmented in a syllable section extracting part 5 based on the estimated average syllable length and the feature quantity. Consequently, the syllable section is segmented based on the average syllable length estimated for each input voice. Thus, syllables are correctly recognized without being affected by the speaker's speaking speed.

Description

DETAILED DESCRIPTION OF THE INVENTION <Industrial Application Field> The present invention relates to an improvement in a speech recognition device that recognizes input speech in units of syllables.

<Conventional technology> Conventionally, there are the following types of speech recognition devices.

In other words, from the input audio? Extract the lI amount, cut out a syllable section based on the extracted feature amount and average syllable length, and calculate the similarity between the feature amount of the extracted syllable section and the feature amount of the standard pattern stored in advance in the storage unit. The results of speech recognition are determined on a syllable-by-syllable basis. At this time, the average syllable length is determined by using a fixed value set and stored in advance, or by calculation from the speech interval length and the number of syllables of a plurality of words inputted in the past.

<Problems to be Solved by the Invention> As described above, in the above-mentioned conventional speech recognition device, the average syllable length used when recognizing syllables is either a fixed value set in advance and stored, or a past value is used. Since the calculation is performed based on the length of the speech interval and the number of syllables of a plurality of words input by voice, the following problem arises. That is, in the case of a speaker who is accustomed to using a speech recognition device, there is no problem because the speaking speed is constant, but in the case of a speaker who is not accustomed to using a speech recognition device, the speaking speed varies with each utterance. Therefore, there is a problem in that the average syllable length used as described above and the syllable length of the actual human-generated speech differ greatly, making it impossible to correctly determine syllable intervals.

Therefore, an object of the present invention is to provide a speech recognition device that can correctly recognize syllables without being affected by the speaking speed of the speaker by estimating the average syllable length for each input speech. be.

<Means for Solving the Problems> In order to achieve the above object, the present invention extracts features from input speech, cuts out syllables based on the extracted features, and calculates similarity on a syllable basis. In a speech recognition device that recognizes syllables, a parameter for estimating a syllable interval is calculated for each input voice based on the feature amount extracted from the human-generated speech, and the parameter for estimating a syllable interval is calculated based on the parameter for estimating a syllable interval. It is characterized by comprising a syllable length estimating section that estimates the average syllable length for each voice, and a syllable section extracting section that extracts a syllable section based on the estimated average syllable length and the feature m.

Further, in the speech recognition device, the syllable length estimation unit calculates a parameter for estimating a syllable interval, which is a combination of a power change and a spectrum change of the feature amount, for each human voice.

Further, the syllable interval estimation parameter is a parameter whose value becomes larger in the transition from a consonant to a vowel than in other parts.

<Operation> When audio is input, feature amounts are extracted from the input audio. Then, based on this extracted feature 1, the syllable length estimator calculates the syllable interval estimation parameter for each human voice, and the average syllable length for each input voice is calculated based on the calculated syllable interval estimation parameter. Presumed. Then, based on this estimated average syllable length and the feature amount, a syllable segment is extracted by a syllable segment extraction unit. Therefore, a syllable section is cut out based on the average syllable length estimated for each input voice.

In addition, in this speech recognition device, the syllable length estimating unit calculates a syllable interval estimation parameter that combines the power change and spectrum change of the feature amount for each input voice, and calculates the average syllable length based on this parameter. Presumed. Therefore, the average syllable length can be accurately estimated.

In addition, this speech recognition device calculates a syllable interval estimation parameter for each input voice such that the value is larger in the transition from a consonant to a vowel than in other parts, and a syllable interval is extracted based on this. Therefore, the average syllable length is estimated based on the transition point from the consonant to the vowel, and the syllable is extracted more accurately.

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

FIG. 1 is a block diagram of the speech recognition device of the present invention,
Reference numeral 1 is a microphone for capturing the uttered voice, and reference numeral 2 is an amplifier that amplifies only the audio band of the human-generated voice.

The feature extractor 3 uses a feature mff1 (for example, a short section (5+s
s~20m5) and the output of a 16-band filter bank, etc.). The average syllable length estimating unit 4 calculates a syllable interval estimation parameter based on the feature extracted by the feature extracting unit 3, as will be described in detail later, and calculates the average syllable length based on the syllable interval estimation parameter. Estimate. The syllable section extractor 5 extracts a syllable section based on the feature amount extracted by the feature extractor 3 and the average syllable length estimated by the average syllable length estimator 4.

The similarity calculation section 6 calculates the similarity between the feature amount in the syllable section extracted by the syllable section extraction section 5 and the feature amount of the standard pattern registered in advance in the standard pattern memory 7, and calculates the similarity calculation result. Perform syllable recognition based on. The interface section 9 is used when communicating with an external device (not shown) such as a word processor. The CPU 8 includes a feature extraction unit 3. Average node length estimator 4. The syllable segment extraction unit 5 controls the similarity calculation unit 6 and the interface unit 9 to execute speech recognition operations.

Here, the syllable interval estimation parameter calculated by the average syllable length estimating section 4 is a parameter (PWSP -f (Pot, 5PEC
): POW=Pa'7-. Use 5PEC=spectral change). Equation (1) shows a specific equation for the syllable interval estimation parameter PWSP used in this embodiment.

PWSP= (S l/S 2)X ((P 2- P
1)/(P 1+ K))...(1) Here, Sl: Spectral change in the 8-frame window at time t.

S2: Spectral change of a 4-frame window at time (t-4).

Pl: Power at time (t-3) P2: Power at time t K: Constant.

The syllable interval estimation parameter PWSP captures the rise of the power and the point where the spectrum change is large. Therefore, it often shows a large value at the transition from a consonant to a vowel in a syllable.

FIG. 2 shows changes in power and syllable interval estimation parameter pvsp when the speaker utters /tana/.

From this diagram, when uttering /tana/, the beginning of the syllable /ta/ (A) and the transition from the consonant to the vowel of the syllable /na/ (B)
? Here, the syllable interval estimation parameter PWSP has a peak. When there are two or more peaks of the syllable interval estimation parameter pwsp of the input voice obtained in this way, the interval between the peaks of the syllable interval estimation parameter pwsp corresponding to the syllable (i.e. A-B Let the interval Q) be the average syllable length of the syllable /ta/.

However, if the interval between the peaks of the syllable interval estimation parameter pwsp is extremely long, there is a possibility that two or more syllables are included between the peaks, so the average syllable length is not estimated. In this case, for example, the average syllable length estimated for the immediately preceding syllable is used as the average syllable length of the syllable for which the average syllable length is not estimated. This is based on the fact that the conventionally known consonant-to-vowel transition interval has less variation than the vowel-to-consonant transition interval. That is, as shown in FIG. 2, since the average syllable length Q of the syllable /ta/ estimated by the syllable interval estimation parameter pvsp is approximately equal to the average syllable length of the syllable /na/, the average syllable length Q of the syllable /na/ If an estimated value of the syllable length cannot be obtained, it is possible to use the average syllable length Q of the immediately preceding syllable /ta/ as the average syllable length of the syllable /na/.

In addition, if many peaks of the syllable interval estimation parameter pwsp appear in an interval that is estimated to be one syllable interval, an interval that is appropriate as a syllable length among the intervals between each peak (for example, a preset input syllable (judged from 7 seconds) and average them to find the average syllable length.

In this way, by using a composite parameter of the power change and spectral change extracted by the feature extraction unit 3 as a parameter for syllable interval estimation, it is possible to estimate the syllable interval more accurately than by using only feature quantities such as power change and spectral change. The average syllable length can be estimated stably.

Next, syllable segment extraction by the syllable segment extracting section 5 will be explained using a specific voice input example.

Figure 3 shows the power change when the user utters /ringo/ into the microphone l at the optimal speaking speed. In the figure, the upper row (a) at the bottom of the power change diagram shows the average syllable length (a preset value) according to the conventional method and the syllable section cutout result obtained from the average syllable length, and the lower row at the bottom of the power change diagram (b) shows the average syllable length (the value calculated for each human voice by the average syllable length estimating unit 4 as described above) and the syllable section extraction results obtained from the average syllable length according to this embodiment.

First, syllable boundary candidates C, D, E, F, and G are detected based on the feature amounts extracted by the feature extractor 3. Then, the detected syllable boundary candidates C and D.

Among E, F, and G, syllable boundary positions C, F, and G are determined based on changes in the feature amounts (such as sudden changes in power and spectrum changes shown in FIG. 3). The syllable boundary positions determined in this way are indicated by vertical solid lines. Furthermore, a syllable boundary candidate, E, which could not be determined using only the above-mentioned feature amount, is indicated by a broken line.

Next, based on the average syllable length calculated by the average syllable length estimation unit 4 as described above, syllable boundary candidates that have not been determined as the syllable boundary positions are determined.

The syllable boundary position is determined for E. the result,
When the conventional method is based on a preset average syllable length, the syllable boundary position E is determined as shown in FIG. 3(a). Furthermore, when based on the average syllable length determined for each human voice according to this embodiment, the syllable boundary position E is determined as shown in FIG. 3(b).

In this case, since the speaking speed of the speaker is the optimum speed, the average syllable length preset by the conventional method and the average syllable length estimated for each input voice by this embodiment are almost the same. Moreover, both average syllable lengths are approximately the same as the actual syllable lengths. Therefore, the syllable boundary position determined based on any average syllable length becomes the same syllable boundary position E.

FIG. 4 shows the power change when /mandarin/ is uttered at about twice the normal utterance speed. In the figure, the upper part (a) of the lower part of the power change diagram shows the average syllable length (a preset value) by the conventional method and the syllable segment extraction results obtained from the average syllable length, as in Figure 3. , the lower part (b) of the power change diagram shows the average syllable length (
(value calculated for each human voice by the average syllable length estimation unit 4)
The results of syllable segment extraction obtained from the average syllable length and the average syllable length are shown below.

Similar to FIG. 3, syllable boundary candidates Nl1H, [, J, KL are detected based on the feature amounts extracted by the feature extractor 3. Then, the detected syllable boundary candidates H, I,
In J,, the syllable boundary position H,1, JL determined from among L based on the sudden change in the feature amount is shown by a vertical solid line. Furthermore, syllable boundary candidates K that could not be determined using only the feature m are indicated by broken lines.

Next, based on the average syllable length calculated by the average syllable length estimation unit 4, the syllable boundary position is determined for the syllable boundary candidate K that has not been determined as the syllable boundary position. As a result, when based on the preset average syllable length according to the conventional method, the syllable boundary candidate is not determined as a syllable boundary position as shown in FIG. 4(a). Furthermore, when based on the average syllable length determined for each human voice according to this embodiment, the syllable boundary candidate K is determined as the syllable boundary position as shown in FIG. 4(b).

In this case, since the speaking speed of the speaker is about twice the normal speed, the average syllable length pre-cented by the conventional method will be about twice the syllable length of the actual input voice. Therefore, the syllable boundary candidate is not determined as a syllable boundary. However, the average syllable length estimated for each input voice according to this embodiment is almost the same as the actual syllable length. Therefore, the syllable boundary candidate K is determined as the syllable boundary position.

Therefore, in the conventional example, syllables that are incorrectly cut out as two syllables are correctly cut out as three syllables, /mi/, /ka/, and /n/, according to this embodiment.

The same thing as described above occurs when the average syllable length obtained from a plurality of words inputted in the past is used as the average syllable length in the conventional example.

The point is, unless you find the average syllable length for each human-generated voice,
When the speech rate of a speaker changes, there is always a difference between the average syllable length and the syllable length of the actually input voice, and syllable sections may not be correctly extracted.

In the above embodiment, the syllable boundary position is determined only based on the average syllable length estimated for each input voice by the average syllable length estimator 4.

However, the invention is not limited thereto. That is, in actual input speech recognition, there are cases where it is actually difficult to estimate the average syllable length for each human speech. In such cases, if it becomes possible to determine the syllable boundary position using the average syllable length obtained from multiple words inputted in the past and estimate the average syllable length again for each human voice. In this case, the average syllable length may be estimated by the average syllable length estimating section 4.

<Effects of the Invention> As is clear from the above, the speech recognition device of the present invention has the following effects:
It has a syllable length estimation unit and a syllable interval extraction unit, calculates a syllable interval estimation parameter for each human voice based on the feature quantity extracted from the input speech, and calculates a syllable interval estimation parameter for each human voice based on the syllable interval estimation parameter. The average syllable length is estimated for each input voice, and syllable sections are extracted based on the estimated average syllable length and the above feature amounts, so syllables can be correctly extracted without being affected by changes in the speaker's speaking rate. can be recognized.

Furthermore, the speech recognition device of the present invention calculates, for each human voice, a parameter for estimating a syllable interval, which is a combination of the power change and the spectrum change of the feature m, by the syllable length estimator, and based on this parameter for estimating a syllable interval. Since the average syllable length is estimated for each human voice, syllable intervals can be accurately extracted.

Furthermore, the speech recognition device of the present invention calculates for each input speech a parameter for estimating a syllable interval whose value is larger in the transition from a consonant to a vowel than in other parts.
Syllables can be recognized by more accurately and stably estimating the average syllable length.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the speech recognition device of the present invention, FIG. 2 is a diagram showing an example of changes in parameters for syllable interval estimation in actual input speech, and FIGS. 3 and 4 are syllables produced by the speech recognition device described above. It is an explanatory diagram of a recognition result. l...Microphone, 2...Amplifier, 3...Feature extraction section,
4. Average syllable length estimation section, 5.. Syllable interval extraction section, 6
. . . Similarity calculation unit, 7. Standard pattern memory, CPU, 9. Interface unit.

Claims (3)

[Claims] (1) In a speech recognition device that recognizes syllables by extracting features from the input speech, cutting out syllables based on the extracted features, and calculating similarity on a syllable basis, the input speech is Based on the features extracted from
a syllable length estimator that calculates a syllable interval estimation parameter for each input voice and estimates an average syllable length for each input voice based on the syllable interval estimation parameter; What is claimed is: 1. A speech recognition device comprising: a syllable section extraction unit that extracts a syllable section based on the syllable section. (2) The syllable length estimation unit uses the power and spectrum of the feature extracted for each input voice to calculate a parameter for estimating a syllable interval, which is a combination of power change and spectrum change, for each input voice. A speech recognition device according to claim 1 that calculates. (3) The syllable interval estimation parameter calculated by the syllable length estimation unit, which is a combination of the power change and the spectrum change, has a value larger in the transition from a consonant to a vowel than in other parts. Speech recognition device described in section.