JP2834471B2 - Pronunciation evaluation method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pronunciation evaluation method for evaluating good pronunciation when a speaker to be evaluated utters a foreign language or the like.

[Related Art] Normally, a quantitative evaluation of the prosodic goodness of a foreign language pronunciation is made by a speech with a good pronunciation of the foreign language to be evaluated, for example, a person whose native language is the foreign language (hereinafter referred to as a standard speaker) ) Is used as a reference, and it is generally performed from the viewpoint of how close the voices of the same generation content to the speaker to be evaluated (hereinafter referred to as the evaluated speaker) are. in this case,
In order to ignore the difference in non-linear fluctuation of utterance in the time axis direction and to compare prosody between parts uttering the same phoneme, a fixed time interval (a frame After the temporal correspondence between the speech of the standard speaker and the speech of the speaker to be evaluated based on the spectral similarity in frame units, the spectral characteristics correspond, that is, the phoneme is one. A method is used in which the prosodic features of the evaluated speaker and the standard speaker differ between frames considered to be different from each other, and evaluation is performed.

[Problems to be Solved by the Invention] In this evaluation method, in the conventional technology, the frame correspondence is determined by directly matching the voices of the standard speaker and the speaker under evaluation by the DP matching method, the HMM, and the like, and determined by the corresponding result. I was However, the speaker is simply DP
When a matching method is applied, the frame correspondence based on the physical spectral similarity between the speaker under evaluation and the standard speaker often does not match well with the phoneme correspondence due to individual differences in the spectral characteristics of each phoneme. In the wrong state, the prosodic features of the critical speaker and the standard speaker were compared, which negatively affected the evaluation.

In comparison of the prosodic features, the conventional method of simply comparing the parameters is based on individual characteristics such as individual differences in the average fundamental frequency and average voice power and individual differences in the dynamic range, and the basic shape of the prosody parameters. Could not be evaluated separately, which also had an adverse effect on the evaluation.

An object of the present invention is to improve the accuracy of phonemic correspondence by frame correspondence between a speaker to be evaluated and a standard speaker based on spectral similarity, which was inaccurate in the conventional technology, and to enable more accurate evaluation of prosodic features. And to provide a pronunciation evaluation method capable of separately evaluating individuality such as individual difference in average fundamental frequency and average audio power and individual difference in dynamic range and difference in basic shape of prosodic parameters. .

[Means for Solving the Problems] The present invention provides a speaker adaptation of a speech of a speaker to be evaluated to a speech of a standard speaker by a method such as codebook mapping using a histogram, and after the speaker adaptation. By calculating the frame correspondence based on the spectral similarity between the speech of the evaluated speaker and the speech of the standard speaker, the accuracy of the phoneme correspondence is increased, and the evaluated speaker and the standard speaker are compared for each corresponding frame. By comparing the prosodic features, it is possible to more accurately evaluate the prosodic goodness of the pronunciation, and after normalizing the prosodic features of the evaluated speaker and the standard speaker using their average and variance values. The most important feature is that the difference between the basic shapes of the prosodic parameters can be more accurately evaluated by comparing the differences for each frame correspondence.

[Operation] The speaker adaptation clarifies the relationship between individual differences in spectral characteristics, thereby absorbing individual differences, and improving the accuracy of phoneme correspondence by a frame correspondence method based on spectrum similarity.

Further, by normalizing the average and the variance, it is possible to absorb the individuality such as the individual difference in the average fundamental frequency and the average voice power and the individual difference in the dynamic range, and separate and evaluate only the difference in the basic shape of the prosodic parameter.

[Embodiment] FIG. 1 is a view for explaining an embodiment of the present invention.
This is to evaluate the prosodic goodness of the voice uttered by the evaluated speaker.

First, the standard speaker and the speaker under evaluation utter the same voice set (word, simple sentence, etc.). Next, using the utterance data, the voice of the speaker to be evaluated is speaker-adapted to the voice of the standard speaker.
The method used here is described below.

First, the codebook generation unit 1 creates a codebook for each speaker. The voice 2 of the uttered standard speaker is analyzed by the voice analysis unit 3 on a frame basis. Analysis methods include bandpass filter analysis, linear prediction analysis, and FF
Although T analysis and the like have been proposed, analysis may be performed using either of them. Here, the p-th order LPC cepstrum coefficient is used as a feature parameter related to the spectrum of the frame. Next, the analyzed speech is clustered by the clustering calculation unit 4 to create a standard speaker codebook 5 consisting of a predetermined number n of code vectors, which is a typical spectrum pattern of the standard speaker speech. Linde, Buzo, and Gray:
“An Algorithm for Vector Quantizer Design”, IEEE T
rans.Comm., Vol.COM-28, 1980.

The speech 6 of the speaker to be evaluated is also analyzed by the speech analysis unit 7 in the same procedure, and clustering is performed by the clustering calculation unit 8 to create the codebook 9.

Next, the code stream generation unit 10 uses the code book 5 of the standard speaker created above to vector quantize the speech 2 of the standard speaker in frame units by the vector quantization unit 11 to create the vector code string 12. . Similarly, the speech 6 of the speaker to be evaluated is vector-quantized by the vector quantization unit 13 to create a vector code sequence 14.

Here, the speaker adaptation unit 15 adapts the voice of the evaluated speaker to the voice of the standard speaker. Several methods have been proposed for speaker adaptation.
o, Lee, and Reddy: “Speaker Adaptation through Vecto
r Quantization ", Proc. ICASSP-86, 49.5, 1986.

First, the matching operation unit 16 performs a matching operation between the vector code sequence 12 of the standard speaker and the vector code sequence 14 of the evaluated speaker in the same target voice, and calculates the frame correspondence of the two vector code sequences. Utilizing this correspondence, the histogram generation unit 17 represents the correspondence between the code vectors in the codebook of the standard speaker and the individual code vectors in the codebook of the evaluated speaker in the form of a histogram. Therefore, the adaptive codebook generator 18
, The code vector after speaker adaptation corresponding to each code vector of the speaker to be evaluated before speaker adaptation is weighted and averaged over the standard speaker code vector using the histogram for the code book before adaptation as a weight. Create by. Then, the code vectors after the adaptation are collected, and a code book 19 of the speaker to be evaluated, which is adapted to the speaker, is created.

Next, the frame correspondence calculation unit 20 uses the code vector in the adapted speaker code book 19 created as described above and the code vector in the standard evaluation code book 5 as feature parameters, The spectrum similarity between frames is calculated from the columns 12 and 14 with reference to the codebooks 5 and 19, and the frame correspondence 21 is determined by a matching method. Any of the proposed methods such as the DP matching method and the HMM can be used as the matching method. A report of this matching method using the DP matching method is reported by Shikano: "Word Speech Recognition by Vector Quantization of Input Speech", Symposium of the Acoustical Society of Japan, S82.
−60,1982.

Finally, the prosody evaluation unit 22 calculates an evaluation value of good prosody.

Here, the fundamental frequency and the audio power are used as the prosodic features to be evaluated, and the normalization is performed by the following method. First, the fundamental frequency of the target voice is extracted by the pitch extraction / smoothing unit 23 by a pitch extraction method such as a deformation correlation method, and then smoothing is applied to remove an extraction error such as double / half pitch. Next, pitch normalization unit 24
, The average of the voiced part of the fundamental frequency extracted to normalize the individual difference of the average fundamental frequency is subtracted from the whole, and further divided by the standard deviation of the voiced part to normalize the individual difference of the dynamic range. The audio power is also calculated by the power calculation unit 25, and after taking the logarithm in the power normalization unit 26, the average of the audio section of the audio power is subtracted from the whole to normalize the average audio power. Is divided by the standard deviation of the voice section to normalize. Normalization is thus simply the average
In addition to matching the variances, a method of nonlinearly expanding and contracting the parameters according to the human sense is also possible.

Next, prosodic features are compared based on the frame correspondence 21 after the adaptation. Here, the absolute value of the difference is used as the comparison method. In addition, various comparisons based on the frame correspondence 21 between the evaluated speaker's speech and the standard speaker's speech after speaker adaptation, such as weighting the absolute value and correlating the corresponding prosodic feature parameters, etc. A method is possible. Here, the normalized fundamental frequency and voice power are calculated by the respective prosodic comparison units, that is, the pitch comparison unit 27 and the power comparison unit 28, for each corresponding frame calculated by the above method, and the absolute value of the difference is calculated. The average is obtained only for the audio section, and the audio power is averaged for the entire audio section to obtain evaluation values 29 and 30.
These evaluation values 29 and 30 can be combined with phonological evaluations calculated by other methods to obtain comprehensive evaluation values of pronunciation.

As described above, this method has a structure in which the frame correspondence between the speech of the evaluated speaker and the speech of the standard speaker after speaker adaptation to the standard speaker to absorb personality is calculated. The phonemes of the evaluation speaker and the standard speaker can be more accurately correlated, and the average fundamental frequency is normalized by the average and variance of voiced parts of the fundamental frequency and the average and variance of speech power in the speech section. -Individuality such as individual differences in average audio power and individual differences in dynamic range can be absorbed.

As an effect, more accurate prosodic features can be compared based on correspondence of phonemes, and only the basic shape of the prosodic parameters can be separated and evaluated, so that more accurate evaluation can be performed.

[Effects of the Invention] As described above, the present invention takes a frame correspondence between the speech of the speaker to be evaluated and the speech of the standard speaker after speaker adaptation, that is, absorption of spectral personality, Benefits of improving the accuracy of phonological correspondence, more accurate prosodic feature evaluation, and separating and evaluating only the basic shape of prosodic features by normalizing prosodic parameters using their average and variance values There is.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a pronunciation evaluation method according to the present invention.

──────────────────────────────────────────────────続 き Continued on the front page (58) Fields surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 551 JICST file (JOIS)

Claims (1)

(57) [Claims] The speech of a speaker to be evaluated is speaker-adapted to the speech of a standard speaker, and the correspondence between frames of the speaker-adapted speech of the speaker to be evaluated and the speech of the standard speaker is expressed by spectrum. A pronunciation evaluation characterized by comparing the prosodic features of the evaluated speaker and the standard speaker for each of the frames associated with each other, based on the similarity and then normalized using the average and variance values of the parameters. Law.