JPH0449958B2 - - Google Patents

Description

[Detailed description of the invention]

FIELD OF THE INVENTION The present invention relates to a speech recognition device based on spectral similarity evaluation. Configuration of a conventional example and its problems Problems of a conventional speech recognition device will be explained using a word recognition method as an example. In FIG. 1, input speech is analyzed frame by frame by spectrum analysis means 1, energy normalized by energy normalization means 2 to correct variations in vocalization intensity, and then processed and stored in advance by the same means. A spectral similarity calculation means 3 calculates the spectral similarity between the reference pattern 6 and the standard pattern 6, and the spectral similarity obtained for each frame is calculated over the entire word length by a similarity accumulation means 4. The final judgment is made by the judging means 5. In general, the strength of human speech differs each time even when pronouncing the same word. The energy normalization means 3 is effective for correcting such variations in utterance intensity, but on the other hand, it is effective for correcting such variations in utterance intensity, but on the other hand, it is difficult to correct for intervals where there should be a difference in energy intensity due to the nature of speech (for example, vowel intervals and voiceless consonant intervals). However, the problem is that energy normalization is performed regardless of its properties. FIG. 2 specifically shows this situation. In FIG. 2A, the solid line represents the spectral intensity of vowels, and the broken line represents the spectral intensity of consonants. Further, the vertical striped portion between both spectra indicates spectral dissimilarity (the opposite concept of similarity), and can be considered to represent the original spectral distance. On the other hand, FIG. 2B shows the spectrum after the vowel and consonant spectra described above have been subjected to energy normalization processing. In the figure, the spectrum of the consonant, which originally had low energy, has been expanded and has become comparable to the spectral energy of the vowel, so the spectral dissimilarity indicated by the vertical stripes in the figure is greater than that in Figure 2A. is also getting smaller. That is, according to the energy normalization process,
The spectra of a vowel and a consonant, which originally had clearly different energy intensities and a large degree of spectral dissimilarity, are evaluated as having relatively equal energy intensities and a small degree of spectral dissimilarity. By the way, when performing word recognition, the normalized spectral similarity is accumulated for each frame between the registered speech template and the speech to be recognized. Therefore, speech intervals that are considered to be clearly linguistically different [for example, the vowel interval and consonant interval in the above example]
The dissimilarity of is given greater weight than the dissimilarity of speech intervals that are considered to be linguistically similar (e.g. vowel intervals of the same type, but differing only in vocalization intensity). Evaluation can prevent erroneous recognition from occurring between languages that are considered to be clearly linguistically different in the final recognition result. That is, although energy normalization processing can eliminate factors that cause instability in energy intensity during vocalization, it has the problem of obscuring differences that should be linguistically clear and dominant. Purpose of the Invention The present invention solves the above-mentioned conventional drawbacks, and while taking advantage of the above-mentioned advantages of energy normalization processing,
At the same time, it is an object of the present invention to provide a speech recognition device that can improve the recognition rate by preventing as much as possible the occurrence of erroneous recognitions that are considered to be clearly disadvantageous from a linguistic point of view. Structure of the Invention In order to achieve the above object, the speech recognition device of the present invention includes a spectrum analysis means for analyzing the frequency characteristics of input speech for each frame, and a method for calculating the energy of the speech from the spectrum features obtained by the spectrum analysis means. A means for normalizing energy to obtain a normalized spectrum by excluding elements caused by strength and weakness, and a spectral similarity calculation that calculates the similarity between the normalized spectrum of the registered speech and the normalized spectrum of the speech for recognition for each frame. a similarity accumulating means for accumulating over the entire speech unit length to recognize the similarity for each frame obtained from the spectral similarity calculating means; and a phonological sequence analysis means for determining a phonological sequence of input speech. , a phonological similarity evaluation means for evaluating the similarity between the phonological sequence of the registered speech and the phonological sequence of the speech for recognition over the entire length of a speech unit to be recognized; and a spectrum obtained from the similarity accumulating means. This configuration includes a determining unit that makes a final determination using both the similarity and the phonetic similarity obtained from the phonetic similarity evaluation unit. DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 3 is a block diagram of a speech recognition device according to an embodiment of the present invention, and the same components as those shown in FIG. 1 are given the same reference numerals and their explanations will be omitted. In FIG. 3, 7 is a phoneme sequence analysis means, 8 is a phoneme similarity evaluation means, and 9 is a phoneme sequence standard pattern. The input audio is analyzed frame by frame by the spectrum analysis means 1, and after correction for variations in the generated intensity is performed by the energy normalization means 2, a spectrum standard pattern 6 which has been previously processed and stored by the same means is used.
The spectral similarity between them is calculated by the spectral similarity calculating means 3, and the spectral similarity for each frame thus obtained is accumulated over the entire length of the word by the similarity accumulating means 4.
On the other hand, the input speech is also input to the phoneme sequence analysis means 7, and the phoneme sequence obtained here and the phoneme sequence standard pattern 9, which has been previously processed and stored by the same means, are input to the phoneme similarity evaluation means 8. Then, the phonological similarity is determined. Further, the spectral similarity and phonological similarity accumulated over the entire length of the word are input to the determining means 5. The determination by the determining means 5 is performed as follows. Note that in the following explanation, for convenience, dissimilarity, which is a concept opposite to similarity, will be considered, and this will be referred to as distance. Now, the table below shows the smallest distance [hereinafter referred to as the first candidate distance (d ₁ )] among the multiple spectral distances obtained from the combination of each registered speech and recognition speech, the smallest distance, and the second smallest distance. The difference between the distance and the distance [hereinafter referred to as the second candidate distance difference Δ ₂ ]
Regarding the two values, the magnitude relationship that they can have with respect to thresholds determined separately in advance [hereinafter referred to as the first candidate distance threshold (θ ₁ ) and the second candidate distance difference threshold (θ ₂ )] Each case of combination is shown.

【table】

Claims (1)

[Claims] 1. Spectrum analysis means for analyzing the frequency characteristics of input speech on a frame-by-frame basis, and normalization by excluding elements caused by the strength and weakness of the energy of the speech from the spectral features obtained by the spectrum analysis means. an energy normalization means for obtaining a spectrum; a spectral similarity calculation means for calculating the similarity between the normalized spectrum of the registered speech and the normalized spectrum of the speech for recognition for each frame; A similarity accumulating means for accumulating over the entire speech unit length to recognize the similarity for each frame, a phonological sequence analysis means for obtaining a phonological sequence of the input speech, and a phonological sequence of the registered speech and the phonological sequence of the recognition speech. a phonological similarity evaluation means that evaluates the similarity over the entire phonetic unit length for which the similarity between the sequences is to be recognized, and a spectral similarity obtained from the similarity accumulation means and a spectral similarity obtained from the phonological similarity evaluation means. A speech recognition device comprising: a determination means that makes a final determination using both the phoneme similarity and the phoneme similarity. 2. The determining means determines the largest similarity among a plurality of spectral similarities obtained between each registered speech and the recognition speech, and the similarity difference between the largest similarity and the second largest similarity. 2. The speech recognition device according to claim 1, wherein the speech recognition device is configured to make the determination using three factors: 1. and phonological similarity. 3 The determination means is configured to make a determination when the largest degree of similarity is greater than a first predetermined threshold and the difference in similarity is less than a second predetermined threshold. A speech recognition device according to claim 2. 4 The determining means compares the phoneme sequence of the registered speech that gave the largest spectral similarity and the phoneme sequence of the registered speech that gave the second largest spectral similarity with the phoneme sequence of the speech for recognition, and 4. The speech recognition device according to claim 3, wherein the registered speech whose phoneme sequence is evaluated to be more similar according to the rules set forth above is used as the recognition result. 5 The determining means compares the phoneme sequence of the registered speech that gave the largest spectral similarity and the phoneme sequence of the registered speech that gave the second largest spectral similarity with the phoneme sequence of the speech for recognition, and Based on the rules, when both phoneme sequences are evaluated to be similar to the phoneme sequence of the recognition speech, the registered speech that gave the largest spectral similarity is taken as the recognition result, and one of the phoneme sequences is used as the recognition result. If the sequence is similar to the phonological sequence of the recognition speech and the other phonological sequence is evaluated as not similar to the phonological sequence of the recognition speech, the registered speech has a phonological sequence that is evaluated to be similar. Claim 3, wherein the recognition result is determined as a recognition result, and when it is evaluated that both phoneme sequences are not similar to the phoneme sequence of the recognition speech, the recognition result is not output because satisfactory recognition cannot be performed. Speech recognition device described in section.