JPS5857197A - Voice recognition equipment - 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] This invention is based on the method of recognizing phonemes in input speech by calculating distance scale values between feature parameters obtained from input speech through LPC analysis and standard feature parameters of phoneme. The present invention relates to a device that recognizes input speech.

In the conventional speech recognition device based on 0LPC analysis, the difference in the bandwidth of these LPC spectral envelopes is calculated using $PI/h to calculate the distance measure value of the LPC spectral envelope relationships between the input speech and the standard pattern. I hadn't considered it. Even if the same phoneme is the same, the bandwidth of the LPC spectrum envelope changes due to the articulatory combination within the speaker, and the band width of the LPC spectrum changes, resulting in a different voice depending on the difference between the speaker! Differences in bandwidth occur even for the same phoneme depending on whether there is a lisp or not. For example, as shown in Figure 1, the same phoneme and the formant frequencies fa, fmaro, fs, and fa are all in one sentence, but the articulatory combination within the speaker and the tone of the voice between the speakers are different. The spectral envelope is like a solid line and a dotted line, the sharpness of the peak of each formant frequency is different by 1 degree, and the bandwidth of the same formant is IB*.
, ΔB, l may be nine different values. In this way, when the bandwidths of the KLPC spectrum envelopes are different, it may become impossible to recognize the same phoneme as a different phoneme.

In order to prevent such misrecognition, the phonological 0LPC spectrum envelope between the input speech and the standard speech is proposed.
7197 (2) The image extraction and distance scale calculation unit 2 extracts the feature bar 2 meter by LPC analysis, and combines the feature parameter with the standard feature parameter of the phoneme stored in advance in the storage unit 3. The distance measure value 4 between the spectral envelopes of is determined. In this invention, the bandwidth is normalized when calculating the distance measure value 4. The distance scale values 4 for each standard pattern of phonemes are compared in the phoneme determination unit 5, and a phoneme base 6 that gives a distance scale value smaller than 4 or a phoneme base 6 that has a distance scale value below a certain threshold value is output. be treated. Even if there are multiple output phonological bases, it will be ignored.

The concept of bandwidth normalization for distance measure calculation is shown in FIG. Bandwidth normalization is performed using a known bandwidth expansion method. The input speech 1 is used to obtain a distance scale parameter, for example, a linear prediction coefficient αl, in the LPC analysis unit 11.
This distance scale parameter is corrected by the variable correction unit 12 to
t+g >'α1, and on the other hand, the homogeneous distance measure parameter 13 of the phonological standard pattern, that is, the linear prediction coefficient α
The variable correction unit 14 corrects il in the opposite manner to the correction performed by the correction unit 12 to obtain (1+a) −1a s + . When the correction scale 1 is changed at this time, the difference in the envelope of the LPC spectrum, for example, the LPC cepstral distance CEP
The distance calculation unit 15 calculates the distance so that the distance is minimized. The correction amount C at this time is given by equation (1).

C1 and CI'a are the LPC cepstatum coefficients of the input voice and standard turn, respectively.

To calculate the separation scale, in addition to calculating the LPC peak weighting scale WLR (We1ght*
Alternatively, ε may be determined so that this WLR value is the minimum Ke. Using the value of the nine correction amount a determined in this way, a distance scale value is calculated by LPC cepstochim distance or WLR.

Next, this distance measure calculation, that is, the operation of the feature extraction and distance measure calculating section 2 shown in FIG. 2 will be explained using FIG. The input speech 1a is subjected to autocorrelation analysis by the autocorrelation analysis section 7, and autocorrelation coefficients (pn, HI==l, 2...op) are determined. VL is the analytical order of PaLPC analysis. Next, the LPG analysis unit 8 performs p-order Lpc analysis to obtain linear prediction coefficients (αn, ne=1.2°...-1p), and LPC cepstral coefficients (Cn, nWl, 2. @
@@IQ) and in some cases aLpc correlation coefficient (yH,
Hml, 2. ass, q) is calculated using the linear prediction coefficient (from αn1, equation (2), <3).

1 Difficulty-1 Cnwm-αn--Σ(n-m)αm@cn-m
(21nm - again, y1, 2..., q, the order n of Cn is the order p of LPC analysis) When the order is large, α
Consider p+-αp4gNM@*@・=αqWMO.

r--! αmrl-m (3) m-also so, n■p+1, p+2,-..., q,
LPC correlation coefficients of order p or lower match the correlation coefficients of the output of autocorrelation analysis.

Further, the standard pattern of phonemes is also stored in the storage unit 3 in advance as a feature parameter, including the LPC cepstatum coefficient (Cn') and, depending on the case, the LPC correlation coefficient (rn@), using the above-described procedure. The distance measure value calculating section 9 calculates distance measure values between the feature parameters of the input speech, the feature parameters of the standard pattern of phonemes, and the 0LPC spectral envelope relation.

First, as described above, the bandwidth expansion method is used to calculate the compensation amount ε of the shpand width using equation (1). It has been experimentally confirmed that it is effective to set a limit on the range of this correction amount as shown in equation (4).

r-≦8≦r+ ('1 limit is set, if C is smaller than r-, the value of e is r-1c, and if g is larger than r+z g, -
Let the lid of be r+. LPC cepstral coefficients Cn of the input voice and standard pattern when such correction is performed,
Cn' Id is calculated as follows.

en-(1+g)”Cn(5
1-g n'-(1+1)-"Cn'The difference is B-(1-g Cn'(6)) Therefore, using this corrected coefficient, calculate the distance scale as L.
When calculating the PC cepst 2m distance, it is obtained by the following formula.

According to this calculation, as mentioned above, the correction amount ε is set by one coefficient Cn so that the distance between the envelopes of the LPC spectrum reaches the minimum.
and Cn@ are mutually inversely corrected so that both envelopes become closer to each other, and the band widths are brought closer to each other, resulting in a normalized distance measure value.

It's fine if you have one. To calculate the distance scale value, it is best to state KL first.
There are nine cases based on the WLR of the PC beak weighting scale, and when the distance scale value in this case is calculated by performing the above correction, )−s. Here, the values of //>i and meeting l are respectively predicted coefficients ((1+1') αt) (t-xs...pp) and (
It is obtained from (1+1)-α11)(""1e...op) via the Percoll coefficient. α1 and α1 here
1 is the linear child-side coefficient of the input voice and the linear prediction coefficient of the standard pattern.

As stated above, according to the present invention, the distance measure value between the LPC spectral envelope of the input speech and the standard phonetic pattern can be calculated under the condition that the bandwidth of the spectral envelope is normalized, and the phonological standard pattern can be calculated with high accuracy. can be identified.

By normalizing the band width variations due to articulatory coupling and normalizing the band width variations between speakers, the sound m1lli! It can improve the performance of the heart. Note that when the restriction of equation (4) is provided, by making a large correction, it is possible to prevent the recognition accuracy from decreasing.

The present invention can be applied not only to utterance qa phoneme recognition, but also to, for example, a word speech recognition device to improve recognition performance. FIG. 5 shows an example of a word speech recognition device, in which a time series of characteristic parameters for each word is stored in the storage unit 3.

The input audio 1 is the same as the standard pattern time series stored in the storage unit 3.
A PC spectral distance scale value 4 is calculated by the calculation unit 2, and time axis normalized matching is performed on the distance scale value 4 by the matching unit 17. As a result of the calculation, the distance scale value 18 in units of words is input to the word determination section 19, and the word name 20 having the smallest distance scale value 18 or the distance scale value below a certain threshold value is output.

[Brief explanation of the drawing]

Fig. 1 is a diagram showing an example in which the bandwidth of the stuttering spectrum envelope is different for the same phoneme, Fig. 2 is a conceptual diagram showing an embodiment of a nine-phoneme identification device applying the present invention, and Fig. 3 is a diagram showing the bandwidth of the stuttering spectrum envelope. FIG. 4 is a conceptual diagram showing an example of the feature extraction and distance measure calculation unit 2 shown in FIG. 1. FIG. 5 is a conceptual diagram showing the concept of the normalization scale, and FIG. It is a conceptual diagram showing an example of a device. 1: Input audio, 2: Feature extraction and distance measure calculation unit, 3:
Memory of feature parameters of standard patterns of phonemes, 4: Distance scale value, 5: Phoneme determination unit, 6: Phoneme name, 7: Autocorrelation analysis unit, 8: LPC decomposition unit, 9: Distance scale calculation unit, 17:
Time axis normalization matching section, 18: Distance scale value in word units, 19: Word determining section, 20: Word name. Patent applicant Nippon Telegraph and Telephone Public Corporation agent Military expenses Taku': name of invention
Relationship with the case of correcting audio 1ivi equipment Patent applicant: Osamu Nagayo, Nippon Telegraph and Telephone Public Corporation, Meiho, Sagami Building, 4-2-21, Shinjuku, Shinjuku-ku, Tokyo, page 111, line 17: By changing ``by doing'' I am corrected.

Claims (1)

[Claims] (1) In a speech recognition device that recognizes input speech by finding the difference (distance scale value) between a feature parameter obtained from an input speech by LPC analysis and a feature parameter of a standard pattern of phonemes stored in advance, the difference ( 1. A speech recognition device, characterized in that the calculation means (distance scale value) has the ability to normalize the difference in bandwidth between LPC spectral envelopes by paying attention to the difference between LPG spectral envelopes.