JPS6315298A - Pattern generation 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 Technical Field The present invention relates to a method for creating feature patterns in speech recognition.

A number of methods have been developed to recognize the sounds of dan-mei-ku words. Most of these systems register the voices to be used in advance, and then check which voice is most similar to the already registered voices to recognize the unknown input voice. This is based on so-called pattern matching.

FIG. 3 is a block diagram showing an example of the pattern matching method described above, in which 1 is a microphone, 2 is a filter bank,
3 is a dictionary, 4 is a similarity calculation unit, and 5 is a result display unit.As is well known, the voice to be used is registered in the dictionary 3 in advance with the switch S set to the a side, and during recognition, the switch S is set to the b side. A similarity calculation section 4 calculates the input speech and the speech registered in the dictionary 3 on the side, and displays the result on a result display section 5. This pattern matching method is widely used because it requires fewer calculations and has better recognition accuracy than other methods, such as those using discriminant functions. One of these pattern matching methods, which requires a small amount of data and can be executed with simple calculations, is binary TSP.
(Time-5 pectrum pattern) has been announced.

Figure 4 is a diagram for explaining an example of a word speech recognition method in which pattern matching is performed using binary TSP (if necessary, see Proceedings of the Acoustical Society of Japan, 1981).
In the figure, 1 is a microphone, 2 is a filter bank, 3 is a dictionary, 4 is a similarity calculation section, 5 is a result display section, 6 is a binarization section, and 7 is a local A peak detecting section 8 is a superimposing section, and in the pattern matching method, the sound input from the microphone is frequency-analyzed using a group of band-pass filters or the like.

The frequency and its time change are expressed as a pattern (TSP). Furthermore, this is ``1'' centered around the peak on the frequency.
”, and the others are binarized as rOJ to create a binary TSP (Bin
ary TSP = BTSP), and register the BTSP obtained by uttering multiple times as a standard pattern.The pattern with only the frequency peak as 1 is called the peak pattern, and the peak and its surroundings are called the peak pattern. The pattern including the total number of 1 is called a broad pattern. When an unknown voice is input, a BTSP is created for this voice in the same process as when creating a standard pattern, and compared with previously registered standard patterns to determine the degree of similarity with each standard pattern. The degree of similarity is determined from the degree of overlap of elements of "1" when the BTSP of the unknown voice and the standard pattern are superimposed (see FIG. 7). The details are described in the aforementioned journal of the Acoustical Society of Japan, so they will be omitted here.

This method has the advantage that if a standard pattern is created well, it is easy to realize a speaker-independent speech recognition device that can recognize anyone's voice. Furthermore, this method has the advantage that it is less affected by the loudness of the voice because the spectrum is expressed in binary form, but it also has the disadvantage that it is difficult to distinguish between consonants with low power and vowels with high power. As a countermeasure to this problem, it is conceivable to binarize the envelope shape of the voice power and use it together with the spectrum pattern as a characteristic pattern of the voice.

FIG. 5 is a diagram showing the configuration of the audio input section when the envelope shape of audio power is binarized. , 35 is a filter bank, and 36 is a multiplexer and A/D conversion circuit.The audio signal input from the microphone 31 is first amplified by the preamplifier circuit 32, and its high level is increased to make phonological information such as consonants more prominent. The high frequency component is emphasized by the high frequency emphasis circuit 33. Since the voice level fluctuates each time depending on the speaker and psychological and physiological speaking conditions, the amplitude is corrected or corrected by the AGC circuit 34 in order to normalize the voice level fluctuation. The output is sent to filter bank 3 without
Enter 5. The filter bank 35 has 1 channel, for example, 15 channels.
The bandpass filters (B, P, F) are used, and the center frequency of the bandpass filters is 250 to 6300Hz, they are arranged at 1/3 oct intervals, and the sharpness is Q=6. In order to extract only the amplitude envelope component of each output of the bandpass filter, a full-wave rectifier circuit and a B
The signal is transmitted to a smoothing circuit which is an e5sel type fourth-order linear phase filter. The characteristics of the voice obtained in this way indicate the approximate shape of the voice spectrum and are called the spectral envelope.

The spectral envelope obtained by the filter bank is sequentially converted into a digital code by an analog multiplexer (MPX) and a 12-bit AD converter (ADC) 36, and by sampling this at a sampling period of 10m5, the voice feature amount is It is expressed as a time series of spectral envelopes. The unknown pattern input for recognition is a peak pattern, and this pattern represents peaks with "1" and others with "0". If the patterns used to create the dictionary could also be treated as binary data, this would be an advantage in realizing an actual recognition device. From this point of view, it has been proposed to create all feature patterns from binary patterns.

Figure 6 shows the power envelope shape of the voice as described above.
This is a diagram for explaining an example of the BTSP method that performs speech recognition by converting into a value.
nary Time-5pectrum Patter
This is a block diagram for explaining an example of (n), in which 41 is a microphone, 42 is a filter bank, 43 is a least squares correction section, 44 is a binarization section, 45 is a BTSP creation section, 46 is the addition part of the one-time utterance pattern by linear expansion and contraction, 4
7 is a dictionary section, 48 is a peak pattern creation section, 49 is a pattern length matching section by linear expansion and contraction 550 is a similarity calculation section,
Reference numeral 51 denotes a result display section, which performs recognition by linearly matching dictionary patterns with input cover patterns obtained by binarizing speech uttered word by word. In the case of speech recognition for unspecified speakers, a new dictionary pattern is created as a superposition of TSPs uttered by multiple people (for details on BTSP, if necessary, Ricoh Technical Report
Na 11.

MAY, 1984. P, P4-12; Proceedings of the Acoustical Society of Japan, October 1982, P, P195-196 (3
-1-8) etc.).

This method is strong against pattern fluctuations in the frequency direction, that is, differences between individual speakers, and is suitable for a speaker-independent system. However, since the absorption of time fluctuations is based on linear expansion and contraction,
It is inferior to DP matching.

Figure 7 shows the overlap of the five normal BTSP patterns, and
Figure 8 shows an example in which time fluctuations are difficult to absorb. In both figures, (a) is a broad pattern, (b) is a peak pattern, and (c) is a pattern of (a) and a pattern of (b). In the example shown in Figure 8, the broad pattern (a) and peak pattern (b) are linearly expanded and contracted as shown by dotted lines, but in Figure 8 (c) As shown by the circle d, there is a drawback that a protruding portion is generated due to time fluctuations, making it difficult to absorb time fluctuations.

Therefore, the spectral part (BTSP) in the pattern created by this method is the 2nd to 5th part of the formant in the voice.
In order to select the band so that the third order is represented, one frame of B
In the TSP, the number of "1"s is equal to or more than the formant number, whereas in the power representing the power, only one "1" indicating an envelope does not exist. In other words, during pattern matching, the spectrum part is given greater weight than the power part, and the difference in power cannot be sufficiently reflected in pattern matching.

1-On the other hand, the present invention has been made in view of the above-mentioned circumstances,
In particular, the purpose of this invention is to provide a method for creating a pattern in which spectrum and power are treated with the same weight during pattern matching.

Structure In order to achieve the above object, the present invention has a part for recording audio and a means for frequency analysis, binarizes the result of the frequency analysis based on the level, and further converts the input audio into binarization based on the magnitude of the level. In an audio pattern creation method that binarizes the temporal change of , and combines the two into one audio pattern,
This method is characterized in that the loudness of the sound is expressed by the same number of pattern elements as the number of pattern elements characterizing the pattern on the frequency pattern. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, in which 11 is a microphone, 12 is a voice section detection section, 13 is a band pass filter, 14 is a voice power detection section, 15 is an A/D conversion section, 16 is a binarization section,
17. 18 is a register, 19 is a counter, 20 is a peak normalization section, 21 is a power pattern creation section, 22 is a combination section, 23 is a pattern section, and the audio is transmitted to the microphone 1.
1, the section detecting section 12 separates it from noise etc. and extracts only the voice section, and the band pass filter group 1
Frequency analysis is performed using 3. On the other hand, the magnitude of the power of the same signal is measured by a power detection unit]4, and is normalized so that the maximum power is constant in order to perform binary fhi processing.

The magnitude of the power at this time can be determined, for example, from the output of an all-pass filter. The frequency-analyzed pattern is binarized by a method such as drawing a least square error approximation straight line, and the number of "1"s in the binarized pattern is counted by a counter 19.
Count by Next, “1” was counted by counter 19.
The number is reflected in the power pattern, and the number "1" generated in the spectral pattern is created in the power pattern. The power pattern is as shown in Figure 2, with the envelope of the audio power being "1".
, others are represented by "0", and the number of "1"s that cause the power envelope increases in the amplitude direction according to the number of "1"s in the spectrum pattern. However, before a plosive consonant or in a silent section that appears as a consonant, the voice power becomes O and the spectrum disappears, but in this case, it is not necessary to match the number of "1"s as described above.

A pattern is created by combining the power pattern created in this manner and BTSP.

As is clear from the above explanation, according to the present invention, the spectrum part and the power part are treated with the same weight during pattern matching, so even if the spectral distributions are similar, the power This makes it easier to distinguish between speech patterns that have different magnitudes.

[Brief explanation of the drawing]

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention, FIG. 2 is a diagram showing an example of a power pattern according to the present invention, and FIG. 3 is a diagram showing an example of conventional speech recognition. The block diagram shown in FIG. An electrical block diagram for explaining an example of the BTSP method, FIG. 5 is a diagram showing the configuration of the audio input section when the envelope shape of audio power is binarized, and FIG. 6 is a diagram showing the configuration of the audio power envelope. A diagram for explaining an example of the BTSP method that performs voice recognition by binarizing shapes. Figure 7 is a diagram showing the overlapping of normal BTSP patterns. Figure 8 is an example in which hourly fluctuations are difficult to absorb. FIG. DESCRIPTION OF SYMBOLS 11... Microphone, 12... Voice section detection part, 13... Band pass filter, 14... Audio power detection part, 15... A/D conversion part, 16... Binarization part . 17.18...Register, 19...Counter, 20
... Peak normalization section, 21... Power pattern creation section, 22... Combination section, 23... Pattern section. Patent applicant: Ricoh W Co., Ltd. / Figure 2, item 01 00001000100 00011111
3 Ward No. 411! J Figure 5

Claims (1)

[Claims] It has a part for recording audio and a means for frequency analysis, and it binarizes the result of the frequency analysis based on the magnitude of the level, further binarizes the temporal change in the volume of the input audio, and combines the two. A voice pattern creation method for forming one voice pattern, characterized in that the loudness of the voice is represented by the same number of pattern elements as the number of pattern elements characterizing the pattern on a frequency pattern.