JPS6073598A - Voice recognition system - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application of the Invention] The present invention relates to a speech recognition method, and particularly to a method for recognizing continuous utterances in units of phonemes.

[Background of the invention]

In a recognition device that has to recognize a large number of idioms, word-by-word g-strings, which have been put into practical use in the past, have many practical problems in terms of the effort required to register standard patterns and the recognition ability. Therefore, technology that recognizes phonemes and syllables is attracting attention. Phonology Consists of vowels and consonants. It is known that the phonemes in continuous speech are significantly modified by the preceding and succeeding phonemes (articulatory combination).

In general, the influence of consonants on vowels is greater than the influence of vowels on consonants. In order to recognize these influences by considering them, it is possible to use a standard pattern as a unit that takes into consideration the phonemes before and after, but this is not practical because the combinations would be very large. Therefore, there is a method in which a vowel with relatively little deformation is first recognized, and a consonant sandwiched between the recognized vowels is recognized using a standard pattern of the vowel environment. However, for certain vowels (mainly lit
and 1ul) When it is sandwiched between certain consonants (mainly voiceless consonants), it may be devoiced and dropped (devoicing phenomenon).

When a vowel becomes unvoiced, its characteristics such as spectral structure change from that of a voiced vowel, making it difficult to recognize using normal vowel recognition methods and often being overlooked. If a vowel in a word is overlooked, it is difficult to recognize the consonants before and after it, and the student may misidentify three consecutive consonants, a vowel, and a consonant, or two syllables related to these.

[Purpose of the invention]

SUMMARY OF THE INVENTION An object of the present invention is to provide a method for estimating the positions of devoiced and omitted vowels and the positions of consonants and pellicles in continuous speech.

[Summary of the invention]

In order to achieve the above object, the present invention focuses on the following facts. That is, (a) in Japanese, syllables are basically formed by pairs of consonants and vowels, and the rhythm in which syllables occur is almost constant; (C) The speed of the above rhythm changes, but in cooperative vocalizations, the changes and fluctuations are gradual; (C) The speech recognition device is basically used online, and the recognition results - The length of a syllable is related to the number of syllables that make up the voice uttered at one time, but there is a certain tendency to There are various points such as regularity. Based on these facts, the position of the devoiced vowel or dropped vowel is estimated using the following procedure.

(1) Set the average syllable length.

(2) Assuming a permissible variation range of syllable length of ±30 well as in (1) (or (8) described later), estimate the number of syllables to be uttered for the first time. For example, the method for estimating the number of syllables is based on the patent application filed in 1983-1971 by the same inventor.
No. 230 can be used.

(3) Recognize the input voice. As for the recognition method, it is possible to use, for example, the patent application No. 1983-91283 filed by the same inventor.

(4) Display and confirm the recognition results. If there is a mistake, please write it down.
Correct the R part. Various correction methods can be used, such as input from a keyboard.

(5) Based on the confirmation result, determine the number of input syllables.

(6) From the length of the input voice and the result of (5), calculate the syllable duration of the voice input at the end (just divide the total length by the number of syllables).

(7) Based on the fact in (2), correct the estimated syllable length. FIG. 1 is an example of actual measured values of the distribution of the number of syllables and syllable length of words when inventor A carefully uttered about 100 words. Japanese words have around 4 syllables and many crests, so if we use this syllable length as an average value, in the example in this figure, for example, if the input is 3 syllables,
Approximately 0.85 times the actual measurement value, approximately 1.
Just multiply it by 1. As can be seen from the figure, the lengths of voices with the same number of syllables from the same speaker vary, so corrections with the above-mentioned accuracy are sufficient.

(8) Set the new average syllable length as the weighted average of the average syllable length used so far and the syllable length obtained through actual measurements.

(9) Returning to (2), calculate the number of syllables while correcting the estimated syllable length based on the recognition results.

In the above procedure, after estimating the number of syllables in step (2), the estimated average syllable length of the speech can be estimated by dividing the speech interval by the estimated number of syllables). The length is sequentially checked, and if a section longer than 1.5 times the estimated average syllable length exists, it is assumed that a devoiced or dropped vowel exists in the center of the section. Next, for more than 70% of the section from the front of this section, the same acoustic characteristics continue (for example, the spectrum outline is the same, or there is a continuation of a section without sound)
When this happens, it is assumed that there is a consonant rather than a devoiced or dropped vowel (voiceless/dropped at the end of the word).

By following the steps described above, even if the speaking rate changes during use, it is possible to accurately estimate the number of vowels, especially the position of voiceless vowels, while following the changes. becomes.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

FIG. 2 is a block diagram illustrating an embodiment of the present invention.

In FIG. 2, a small amount of audio input to the input terminal 1 is sent to a short-time power analysis section 2 and a spectrum analysis section 3.

When the short-term power is determined, the value in the section where the value exceeds a certain value is clipped and sent to the short-time power pattern buffer register 5, and the value is also sent to the control unit 6. The control unit considers that the first vowel section of the voice has started from the moment the Bawani value exceeds a first constant value θ1 that has been added in advance, and after that, the power decreases below the second constant value θ2 for a certain period of time. When it continues for more than 500 m5 (for example, 500 m5), it is considered that the last voiced vowel section of the input speech has ended.

Let To be the length from this starting point to the ending point H. The average vowel length 1G is recorded in advance in the memory 7, and the control unit determines how many syllables in this range of to ± 30% are included in To, and calculates the estimated range of the number of syllables. At the same time, each syllable length τl within the estimation range is also determined. The control unit 6 generates a rectangular pattern in this estimated syllable number range from the rectangular pattern generation unit 9, extracts the cross-correlation coefficient with the pattern in the backup memory 5 into the correlation unit 8, and sends the result to the control unit. Intake. An example is shown in FIG. (a) is audio/1euchi
The short-time power pattern p (t) of soba/ is
(b) is the rectangular pattern 1+(1) that had the highest correlation with this pattern. Let the short measurement point of the section be To, and let the correlation value r, be lT.

FIG. 4 shows an example where r+=5Σp(t)·t+(t), and the correlation is maximum at the correct number of syllables in the above example, 5. In this example, the estimated range of syllable counts is ±3 of the average.
0% and wider range (3 to 8), the situation is shown to make it easier to understand.
This is the case of kus37. It can be seen that the vowel +01 has become devoiced and the value of the power pattern has become smaller, but as shown in FIG. 6, the correlation with the correct number of syllables, 4, is the highest. Section A in FIG. 5(a) is long, and it is assumed that the vowel has become devoiced or dropped during this period. ↑ of the rectangular pulse in (b)
The location of the parts also suggests this point. Actually in this example jul
is muted at this position.

In addition to the thresholds θl and θ2, the control unit 6 has lower thresholds θ3 and θ4, and when there is an unvoiced sound at the beginning of a word,
It is presumed that there is also a vowel that has been deformed due to devoicing/dropping etc. at that position.

On the other hand, the input signal input to the spectrum analysis section 3 is stored in the buffer memory 10 after being converted into spectrum information. The spectrum information stored in the buffer memory 10 is
A matching unit 11 performs matching with vowels, silences, unvoiced parts of unvoiced fricatives, etc. in the standard pattern memory 12 at each analysis time, and the matching results are sequentially sent to the control unit 6.

The control unit 6 determines a vowel, a consonant, or a devoiced/dropped vowel based on this result and the syllable position estimation result (described above). In principle, devoiced vowels are assumed to be 111 and lul, but in combinations that may result in l kokoro l, 1 haha1, etc., they are assumed to be 1.1 or 1.1.

Once the vowel candidates are determined in this way, the control unit 6 sequentially sends the spectral information string in the buffer memory 10 to the second matching unit 14, and creates a standard pattern of consonants narrowed to the estimated vowel. It is taken out from the baton memory 13 and mated, the result is taken into the control unit 6, merged with the vowel estimation result, and confirmed by the confirmation unit 15 as a syllable recognition result.
to be displayed. The user looks at the displayed result and if it is correct, it is OK.
If the key is incorrect, enter the key correction result. The control unit 6 obtains the correct number of syllables based on the confirmation result, determines the average syllable length of the input speech based on the information on the speech interval length,
Furthermore, as explained in FIG. 1, the length is corrected as a function of the number of syllables, and is determined as the estimated syllable length to'. A new average syllable length is calculated by combining this value to', the previous average syllable length "ox't," and a weighted average.

to−αto+βto′ α×β=1 As a result, even if the speaking speed gradually changes while using the
By following this, you can calculate the average syllable length.

According to actual measurement results, the syllable duration of sequentially uttered voices does not fluctuate more than ±30 degrees of the weighted average duration up to that point, except for extreme out-of-sequence cases (mostly in the 20th section). Therefore, the rectangular wave that correlates with the short-time power pattern of the input waveform may cover a range within ±30% of the weighted average time length. As a result, not only the processing time is shortened, but also the estimated syllable duration ib is significantly reduced, and as a result, the recognition performance is significantly improved, and a speech recognition device that is easy to use can be obtained.

The corrected average syllable length to is stored in the memory 7 and used for the next input. Also, the confirmed recognition results are output to the output section 16.
The signal is then output from terminal 17.

Note that the estimated number of syllables may have a difference of about ±1 due to diphthongization and the effects of phlegmatic sounds, but most of the errors in estimating syllable length due to this are within 25%9. This does not pose an obstacle to estimating omitted vowels or consonants.

〔Effect of the invention〕

As explained above, according to the present invention, it is possible to estimate the position of devoiced or dropped vowels while following the speech rate that gradually changes during use, and it has high recognition ability and is easy to use. A voice recognition method can be provided.

[Brief explanation of the drawing]

Fig. 1 is a diagram explaining the relationship between the number of syllables constituting speech and the syllable duration, Fig. 2 is a block diagram explaining the present invention in detail, and Figs. 3 to 6 show estimation of the number of syllables and devoicing position. It is a figure showing an example for explaining estimation. 6...Control unit.

Claims (1)

[Claims] 1. Along with the average syllable information and the all syllable length information,
It has means for estimating the syllable length constituting the input speech, and means for confirming the recognition result of the input speech, and estimates the input sound length of the input speech from the number of syllables of the input speech obtained from the confirmation introduction button. , the average sound ft6 length*ib is corrected according to the input sound length of the input sound, and is used as average syllable table information of the next input sound. 2. In the speech recognition method described in claim 1, using the estimated syllable-cum-information, devoiced vowels and dropped vowels,
A speech recognition method that is characterized by estimating the positions of consonants and consonants.