JPS6363920B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a speech recognition method. Configuration of conventional examples and their problems The conventional method divides input speech into phoneme units and recognizes them as combinations of phonemes (called phoneme recognition), and outputs the recognition result by calculating the similarity with a word dictionary written in phoneme units. A block diagram of the word recognition system is shown in FIG. First, the audio analysis unit 1 analyzes the voices of multiple speakers for each frame (one frame is 10 msec) using a filter bank, and the feature extraction unit 2 analyzes the voices of multiple speakers using a filter bank based on the obtained spectrum information. Find the feature parameters by. Standard patterns are created for each of the five vowel and consonant phoneme groups from these feature parameters and are registered in the standard pattern registration section 3. Next, a segmentation unit 4 performs segmentation using the feature parameters determined by the feature extraction unit 2. Based on this result, the phoneme discrimination unit 5 determines the phoneme by comparing it with the standard pattern in the standard pattern registration unit 3. Finally, the time series of phonemes created as a result is sent to the word recognition unit 6, and the word corresponding to the item with the highest degree of similarity to the word dictionary 7 similarly expressed in the time series of phonemes is output as a recognition result. Here, in segmentation, when the shape of the temporal change in the overall power 8 is concave as shown in Figure 2 (this is called a dip), the frame in which the power shows the minimum value is set as n ₁ . , n ₁ , frames in which the rate of change in power over time (this is called a power difference value) 9 exhibits maximum negative and positive values are denoted by n ₂ and n ₃ . Also, if the difference value at a certain frame n is WD(n), then WD(n ₂ ),
When WD(n ₃ ) satisfies the following conditions: WD(n ₂ )≦−θ _w WD(n ₃ )≦θ _w , the interval from n ₂ to n ₃ is defined as a consonant interval. Here, θ _w is a threshold value for preventing addition of consonants. Next, characteristic parameters indicating the characteristics of phonemes are determined for each frame for this consonant section, and compared with standard patterns for each phoneme prepared in advance to perform consonant classification for each frame. This result is applied to the consonant classification tree to classify consonants into those that match the conditions. As described above, in the conventional method, first the whole range power of input speech is determined, the power dip is used to perform segmentation of mid-word consonants, and then consonant classification is performed for each frame. Finally, the results of the consonant classification for each frame are applied to a consonant classification tree, and the consonants are classified into those that meet the conditions, which is a time-consuming and time-consuming method with extremely complex algorithms. Purpose of the Invention The present invention has been made to solve the above-mentioned conventional problems, and provides a speech recognition method that can extremely easily segment input vocal sounds and classify consonants. The purpose is to Structure of the Invention In order to achieve this object, the present invention uses both low frequency power and high frequency power as characteristic parameters,
By applying the size of the dip caused by the temporal change in each power to a two-dimensional discriminant diagram, segmentation of middle consonants and general classification of consonants can be performed simultaneously and easily. . DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. In this embodiment, low frequency power and high frequency power are used together as feature parameters. This is because voiced consonants tend to exhibit characteristics in high-frequency power, and voiceless consonants tend to exhibit characteristics in low-frequency power, so by using both low-frequency and high-frequency power, almost all consonants can be handled. In this example, the phonemes |P|, |t|, |k|, |
c|, |b|, |d|, |m|, |n|, |s|, |h
| is a voiceless plosive {|p|, |t|, |k|, |c
|}, voiced plosive {|b|, |d|}, nasal {|m
|, |n|}, voiceless fricative {|s|, |h|}
An example of major classification is shown below. In this embodiment, the low-frequency power is obtained by passing the input audio signal through a low-frequency band filter, obtaining logarithmic power for each frame (in this embodiment, one frame is 10 msec), and smoothing the logarithmic power. Further, high frequency power can be obtained in the same manner using a high frequency band filter. Next, FIG. 3 shows how to determine the size of the power dip. In the figure, numerals 10 and 11 represent temporal changes in the low-frequency power and high-frequency power, respectively, and numerals 12 and 13 represent temporal changes in the difference values. The frame where the difference value of low-frequency and high-frequency power is the maximum negative and positive value is nA ₁ ,
Let nA ₂ , nB ₁ , nB ₂ be the difference values in each frame as WD (nA ₁ ), WD (nA ₂ ), WD (nB ₁ ), WD
(nB ₂ ). Therefore, the size of the low-frequency power dip PL and the size of the high-frequency power dip PH
is defined as PL = WD (nA ₁ ) - WD (nA ₂ ) PH = WD (nB ₁ ) - WD (nB ₂ ). As the consonant section, a dip section is taken, and if the low-frequency and high-frequency power dips overlap, the intersection of both dip sections is taken. The consonant section in FIG. 3 is section C from nA ₁ to nB ₂ . The size of the dip in this consonant section is PL and PH.
become. For example, the phonemes |k|, |d|, |n|, |
When examining the distribution of PL and PH in the case of s|, Figure 4~
It will look like Figure 7. In the diagram, the horizontal axis represents PL, and the vertical axis represents PH, and the numbers in the diagram represent the number of phonemes. As is clear from the figure, |k|, |d|
Phonemes that exhibit plosiveness, such as , have large PL and PH; in particular, the voiceless plosive |k| has a large PL, and the voiced plosive |d| has a large PH. In addition, the phonemes |n| and |s|, which do not exhibit plosiveness, have small PL and PH, but are classified as shown in Figures 6 and 7 depending on whether they are voiced or voiceless consonants. As described above, by using both the low-frequency power and the high-frequency power, it is possible to discriminate between voiced consonants and voiceless consonants. Furthermore, by using the size of the dip, rupture can be detected. Based on these properties, we created a two-dimensional discriminant diagram for PL and PH, and used voiceless plosives {|p|, |t|, |k
|, |c|}, voiced plosives {|b|, |d|}, nasals {|m|, |n|}, voiceless fricatives {|s|, |h|}
It can be broadly classified into four groups. FIG. 8 shows a discriminant diagram created using a large amount of data. In the figure, region 14 is considered to be a consonant added to a vowel, and is not segmented as a consonant section. Also, area 15 is nasal sounds {|m|, |n|}, and area 16 is voiced plosive sounds {|
b|, |d|}, region 17 is a voiceless fricative {|s
|, |h|}18 region is voiceless plosive sound {|p|, |
t|, |k|, |c|}. By finding the dip size of low-frequency power and high-frequency power for unknown input speech and applying it to the discriminant diagram in Figure 8, segmentation of middle consonants and general classification of consonants can be performed simultaneously and easily. can be done. Effects of the Invention Using this example, data (approximately 2100 words) of 10 men and women who uttered a set of 212 words was evaluated. The results are shown in Table 1. Regarding the number of dropped consonants in the middle of words, there were a little more dropouts for nasal sounds {|m|, |n|}, but overall the number of dropped consonants was small and better results were obtained than in the past. Furthermore, the recognition rate for major categories of consonants is high at approximately 85% to 90%. As described above, the present invention uses the dip size of low-frequency power and high-frequency power as parameters and applies them to a two-dimensional discriminant diagram, which is a very simple method for segmentation of consonant intervals. It is possible to determine the major classification of consonants with high accuracy. 【table】

[Brief explanation of drawings]

Fig. 1 is a block diagram of a speech recognition system, Fig. 2 is an explanatory diagram of a conventional method for segmenting middle consonants, and Fig. 3 is an illustration of the dip in low-frequency power and high-frequency power in the present invention. The other explanatory diagrams, FIGS. 4 to 7, are diagrams showing the distribution of the dip size of low frequency power and high frequency power for each phoneme, and FIG. 8 is a discrimination diagram in an embodiment of the present invention. 1... Acoustic analysis section, 2... Feature extraction section, 3...
Standard pattern registration unit, 4... Segmentation unit, 5... Phoneme discrimination unit, 6... Word recognition unit, 7...
...word dictionary, 8...wide range power, 9...difference value of full range power, 10...low range power, 11...high range power, 12...difference value of low range power, 13...
Difference value of high frequency power, 14... Area of addition, 15
... Nasal sound region, 16 ... Voiced plosive sound region, 1
7...A region of voiceless fricatives, 18...A region of voiceless plosives.

Claims (1)

[Claims] 1 It has a procedure for phoneme recognition, which calculates the low-frequency power and high-frequency power of the speech spectrum, focuses on the size of the dip (concave part) caused by the temporal change of each, and converts them into two-dimensional data. A speech recognition method characterized by simultaneously performing segmentation of mid-word consonants and general classification of consonants by applying a discriminant diagram.