JPS62111300A - Voice analysis/synthesization circuit - 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] Industrial Application Field The present invention is a low bit rate digital signal converter (hereinafter referred to as a codec) for use in voice feature extractors for voice recognition and telephone voice digital signal converters (hereinafter referred to as codecs). This relates to a speech analysis and synthesis circuit used for conversion.

BACKGROUND OF THE INVENTION In recent years, the trend toward digitalization of telephone equipment is thought to be promoted along with the expansion of functionality of equipment and the expansion of the use of communication lines. However, due to this, there is a limit to the services that can be provided to a large number of subscribers using limited lines, and it becomes necessary to use the lines effectively. As a countermeasure against this problem, it has become necessary to compress the audio signal, that is, to convert the audio signal at a low bit rate. Conventional methods for audio codecs at such low bit rates include those that take advantage of variations in the level fluctuations of the audio signal to compress and expand the signal by applying a nonlinear curve to the signal, and those that use the correlation of the audio signal to compress and expand the signal. There have been methods to reduce the bit rate by encoding only the changing components of the signal waveform to increase the compression rate, and methods to reduce the bit rate by approximating the audio spectrum parameters and their residuals with a plurality of pulses. In addition, in speech recognition, in order to extract the characteristics of the input voice and compare the input voice with the dictionary using pattern matching, a huge amount of input voice data must be compressed and stored as parameters in the dictionary, and the input voice is also the same. Since vocal tract parameters must be extracted as parameters, similar to the above, linear prediction coefficients and the like were used to extract vocal tract parameters as features.

An example of the conventional speech analysis and synthesis circuit described above will be described below with reference to the drawings.

FIG. 5 shows the configuration of a conventional speech analysis and synthesis circuit. In FIG. 5, the input voice is used by a voice synthesizer 62 and a pulse generation circuit 61 to generate a pseudo voice, and a subtracter 66 outputs the error component of the input voice. The pulse position of the pulse generation circuit 51 and each pulse signal level are controlled by the pulse generation control circuit 63 so as to minimize the pulse position. Further, the parameters used in the sound synthesizer 52 are configured by linear predictive analysis 64 of input speech. Here, the pulse position obtained as above,
The pulse signal level and the linear prediction harameter were used to extract the features of the input voice using the quantization multiplexer 66. Such an example is shown, for example, in B. S. Mutal U.S. Pat. No. 3,263,371, issued December 1, 1981.

Problems to be solved by the invention Generally speaking, speech or non-speech pronunciation mechanisms (e.g.
When modeling a wind instrument such as a trompon, it can be expressed as a zero-pole model. Sound is generated by driving a filter with filter coefficients of this zero-pole model with a noise source.

In speech, there are many factors involved in the production of consonants, and the airflow that passes through a narrow part of the vocal tract can become turbulent and constitute a noise source. In conventional feature extraction methods, this vocal tract model is positioned as an all-pole model approximated by linear prediction coefficients, but the residual components after being inverse filtered with the all-pole model include nasal consonants and vocal folds. Since the lung part does not fit into the polar model and is treated as a zero point in the acoustic system, it contains a lot of information, and consonants that occur due to the narrowing of the mouth in the vocal tract are not included in the model at all. For this reason, the residual component contained a lot of information of the audio signal.

Considering that it is possible to completely identify the developmental system by introducing the zero model system, it is clear that using the vocal cord developmental model is effective for extracting features for speech recognition, and allows for efficient compression. This improves the audio quality of audio codecs.

However, in such conventional circuits, there is no correspondence other than vocal tract parameter extraction from the speech generation mechanism, and a lot of speech information remains in the residual components after being inverse filtered using the all-pole model. Instead, a simple pulse generation circuit was used instead. In this case, especially voice parts such as unvoiced sounds cannot be efficiently compressed, resulting in deterioration of voice quality and inability to extract complete voice characteristics.

In view of the above-mentioned problems, the present invention focuses on extracting features by creating a pole-zero model of speech or other acoustic signal generators, and uses a simple impulse or a noise source as a sound source and input acoustic signals. It is characterized by optimizing the coefficients of the synthesis filter to minimize the error between the signal and the synthesis filter consisting of a pole-zero model, so it is possible to completely model speech or other acoustic signals. becomes possible. Furthermore, the present invention provides a speech analysis circuit which is effective for data compression of speech analysis used in speech codecs or speech recognition devices by using a quantizer for the coefficients of the modeled synthesis filter.

Means for Solving the Problems In order to solve the above problems, the speech analysis and synthesis circuit of the present invention analyzes input speech as grid-type linear prediction coefficients, extracts output points from each grid point in parallel, and calculates each The output points are weighted by a weighting circuit, each weighted output is added by an adder, and the difference between the added output and the output signal or white noise of the above-mentioned lattice filter is subtracted by a subtracter, and the output error is calculated. has a square error minimization control circuit that controls the coefficients of the weighting circuit so as to minimize the square of the amplitude obtained by passing the signal obtained from the output or error output of the lattice filter through a correlator, Audio analysis that outputs the correlation time that shows the highest correlation, quantizes parameters such as the coefficients of the lattice filter, the coefficients of the weighting circuit, the amplitude obtained from the correlator, and the correlation time, and transmits them to a line etc. using a multiplexer. The apparatus has a structure that is the inverse of the lattice filter described above, and inserts weighted white noise from a white noise source through a weighting circuit at each lattice point. Each.

Insert into the lattice filter coefficients and weighting circuit coefficients,
Furthermore, it consists of a speech synthesizer that restores the audio signal by inserting a pulse signal having a period corresponding to the correlation time and controlling it with the amplitude obtained above from the opposite side of the lattice filter, and outputting it from the opposite side. ing.

Operation The present invention uses the above-described configuration to separate input speech into vocal tract parameters using a lattice filter configuration and coefficients obtained from each lattice point via a weighting circuit, and has a simplified pulse or noise as a sound source. If the input signal cannot be identified as a vocal tract parameter, the residual component is controlled by a square error minimization algorithm using a weighting coefficient as a parameter. In other words, it is not only possible to deal with input sounds other than speech, but also when noise etc. are superimposed on the input speech, the feature extraction circuit used for speech codec and speech recognition consisting of conventional linear prediction filters can be used. It is also effective against deterioration in the identification accuracy of the speech production system because it can be regarded as a pole-zero model.

Embodiment Hereinafter, a speech analysis and synthesis circuit according to an embodiment of the present invention will be described with reference to the drawings.

FIGS. 1 and 4 show the configuration of a speech analysis and synthesis circuit in an embodiment of the present invention. In Figure 1, 1,
7.13 is a correlator, 2, 3, 8° 9.14.15 is a multiplier, 4, 5, 10, 11° 16.17.25 is an adder,
6, 12. 18 is a unit delay element, 19 is a correlator, 20,
21, 22 and 23 are multipliers that constitute a weighting circuit. 24 is an adder; 21 is a square error minimization control circuit.

The operation of the speech analysis and synthesis circuit configured as described above will be explained below with reference to FIG.

First, in Fig. 1, the input voice has vocal tract parameters of correlators 1, 7, 13, multipliers 2, 3, 8°9, 14, 1
5, unit delay element 6, 12.1B, adder 4, 5, 10
, 11, 16, and 17, the input speech is analyzed into parameters K that depend on the vocal tract. For this analysis, for example, the Burg method
New Analysis Technique 7 Oh Time
Series Data” (Mu New Analysis T
echnique for Time 5eries D
ata-Proc Nmu TOmu dvanced stud
y 1nsstitut on 5j4nal pr
oc-XruschadeNetherlands 1
968) etc. The output signal of the adder 16 is sent to the correlator 1
9 to obtain the amplitude, period, etc. of the residual signal. Furthermore, the output of each grid point is multiplier 20, 21.22.2
3 to the adder 24. The output of adder 24 is added to adder 26 for difference with white noise. Here, the error is passed through the square error minimization control circuit 26 and the coefficients (
weighting factors).

Note that the obtained coefficients of the coefficient section 1 weighting circuit of the vocal tract characteristic filter, the amplitude value obtained from the correlator 19, and the time period are sent to a quantizer and a multiplexer to constitute a speech analysis system. Furthermore, various parameters obtained by the above analysis system are restored to the original speech by the following synthesis system.

The synthesis system will be explained below with reference to FIG. In FIG. 4, a pulse oscillation amplitude control section 1o1 applies to an adder 105 a pulse signal having the period and amplitude obtained from the correlator 19 of the analysis system. Furthermore, similarly, ko parameter input section 1
Multiplier 117゜118.112, 113, 107, 108 that controls the coefficient k of the lattice type filter of the 02-synthesis system
In addition, the weighting coefficient W is inserted from the W parameter input section 103 into each multiplier 123, 122, 121, and 120. Here, the lattice type filter of the synthesis system has an inverse relationship to the lattice type filter of the analysis system. Here, since a signal weighted with white noise is added to each grid point of the grid filter of the synthesis system, it is possible to simulate the narrowing of the vocal tract when the child voice is uttered.

As described above, according to the present embodiment, regarding the speech analysis and synthesis circuit, for example, the vocal tract parameter analysis section and the transversal filter composed of W parameters can be independently controlled, and the lattice filter can be controlled independently. By controlling the coefficients of the above-mentioned transversal filter using an algorithm using stability and a learning identification method, it is possible to compensate for the stability of the constituent system. More importantly, since features can be extracted based on the speech production mechanism in this way, even if unnecessary ambient noise is present at the same time as the speech, it is possible to extract features using only the conventional vocal tract model (all-pole model). It is possible to improve the voice quality in the voice codec without causing deterioration of vocal tract parameters as would be the case in the case of dependency, and it is also possible to improve the accuracy of voice feature extraction in voice recognition.

A second embodiment of the present invention will be described below with reference to the drawings.

FIG. 2 is a block diagram of a speech analysis and synthesis circuit showing a second embodiment of the present invention. Note that the same components as in the first embodiment are given the same numbers.

In the figure, 1 to 17 are the same as the linear predictive analyzer (vocal tract) using a lattice filter in Figure 1 (used as a metric analysis), and 20 to 26 are weighted with a weighting coefficient W from each lattice point. The sum of the signals is obtained by an adder 24, which is compared with white noise by an adder 26, and the weighting coefficient W is controlled by a square error minimization control circuit 26 so as to minimize the square error. Here, the configuration is such that the pulse amplitude M and the pulse period T are outputted from the output of the adder 26 via the correlator 19.

The operation of the speech analysis circuit configured as described above will be explained below.

In FIG. 2, when the system from the weighting circuit 20 to the squared error minimization control circuit 26 is identified by an algorithm such as the steepest method or the learning identification method, a white color is applied to the addition input of the adder 25. When noise is input, the subtraction input, ie, the output of the adder 24, is approximated to white noise. That is, in the output of the adder 26, a pulse component corresponding to the pitch that occurs when the voiced sound cannot be identified by the system is detected. Furthermore, in the case of unvoiced sound, the output error is minimized, and a series of parameters W corresponding to the narrow part of the vocal tract corresponding to the time of vocalization can be obtained. Parameters obtained from such a voice analysis system are quantized and sent to the transmission system via a multiplexer, and the receiving section restores the original voice signal using a configuration as shown in FIG.

A third embodiment of the present invention will be described below with reference to the drawings.

FIG. 3 is a block diagram of a speech analysis and synthesis circuit showing a third embodiment of the present invention. In FIG. 3 as well, the same components as in the first embodiment are given the same numbers.

In the same figure, 1 to 17 are the same as the linear predictive analyzer (used for vocal tract parameter analysis) using a lattice filter in FIG. The adder 24 obtains the sum, which is compared with the output of the adder 16 by the adder 26, and the square error minimization control circuit 26 controls the weighting coefficient W so as to minimize the square error. Here, the configuration is such that the pulse amplitude, pulse period TP, presence or absence of input voice, and voiceless determination detection are output from the output of the adder 25 via the correlator 19.

The operation of the speech analysis and synthesis circuit configured as described above will be explained below.

In FIG. 3, when the system from the weighting circuit 2o to the squared error minimization control circuit 26 is identified using an algorithm such as the steepest descent method or the learning identification method, a grid is applied to the addition input of the adder 25. When a residual signal (white noise) corresponding to the input voice is input as a type filter output, the subtraction input, that is, the output of the adder 24, is approximated to a pulse signal or white noise depending on whether the input voice is voiced or unvoiced. Ru. That is, in the output of the adder 26, a pulse component corresponding to a pitch that occurs when the voiced sound cannot be identified by the lattice filter system is detected. Furthermore, in the case of unvoiced sounds, the output error is minimized, and a series 0 of parameters W corresponding to the narrowing of the vocal tract corresponding to the utterance of speech is obtained. This error signal is passed through a correlator 19 to determine the pulse amplitude, period, and output error value using a given threshold value, thereby obtaining a signal output for determining whether the input voice is voiced or unvoiced. Parameters obtained from such a voice analysis system are quantized and sent to the transmission system via a multiplexer, and the receiving section restores the original voice signal using a configuration as shown in FIG.

Note that in the first, second, and third embodiments, the quantizer can also perform optimal sorting according to the appearance frequency of each parameter. Furthermore, in FIG. 4, the pulse oscillator and white noise generator may be configured to determine whether the input voice is voiced or unvoiced and to switch the outputs of both oscillators using a switch or the like, as described in FIG.

Effects of the Invention As described above, the present invention is capable of applying a complete model to speech and other audio input signals, and efficiently extracting speech features. Furthermore, the present invention provides a circuit that is effective for improving the voice quality in low bit rate voice codecs and for extracting features of input voice in voice recognition.

[Brief explanation of drawings]

1 and 4 are block diagrams showing the configuration of a speech analysis and synthesis circuit in a first embodiment of the present invention, and FIG. 2 is a block diagram showing the structure of a speech analysis circuit in a second embodiment of the present invention. , FIG. 3 is a block diagram of a speech analysis circuit in a third embodiment of the present invention, and FIG. 6 is a block diagram of a speech analysis and synthesis circuit in a conventional example. 1.7.13 Correlator, 2, 3, 8, 9, 1
4゜16...multiplier, 4.5.10.11,
16,17゜26...Adder, 6,12.18
... Unit delay element, 19 ... Correlator,
20~23... Multiplier, 24゜26...
- Adder, 26... Square error minimization control circuit.

Claims (3)

[Claims] (1) Analyze the input audio as a grid-type linear prediction coefficient,
Furthermore, output points are extracted in parallel from each grid point, each output point is weighted by a weighting circuit, each weighted output is added by an adder, and the added output is combined with the output signal of the above-mentioned grid filter or white noise. The output or error output of the lattice filter has a square error minimization control circuit that controls the coefficients of the weighting circuit so as to minimize the square of the output error. Pass the signal obtained from the correlator through a correlator, output the amplitude obtained and the correlation time showing the highest correlation, and calculate the coefficients of the lattice filter, the coefficients of the weighting circuit, the amplitude obtained from the correlator, the correlation time, etc. It has a voice analysis device that quantizes the parameters of and transmits it to a line etc. using a multiplexer, and has the inverse configuration of the above-mentioned lattice type filter, and inserts weighted white noise from a white noise source through a weighting circuit at each lattice point. In the configuration, various parameters obtained from the speech analysis device described above are inserted into the coefficients of the lattice filter and the coefficients of the weighting circuit, respectively, and a pulse signal with a period corresponding to the correlation time is generated with the amplitude obtained above. A speech analysis and synthesis circuit comprising a speech synthesis device which inserts a controlled pulse signal from the opposite side of a lattice filter and outputs it from the opposite side to restore an audio signal. (2) Perform linear predictive analysis on the input signal using a lattice filter, weight the output signal from each lattice point via a weighting circuit, obtain the difference between the sum of the outputs and the white noise signal, and square the difference. A squared error minimizing control circuit is provided to control and minimize the coefficients of each weighting circuit, and the amplitude and correlation time of the error signal are obtained from the error signal using a correlator. The speech analysis and synthesis circuit according to item 1. (3) Perform linear predictive analysis of the input signal using a lattice filter, weight the output signal of the lattice point from each lattice point via a weighting circuit, and calculate the difference or error between the sum of the outputs and the output signal of the lattice filter. A square error minimization control circuit is provided which obtains the signal and minimizes the square of the signal by controlling the coefficients of each weighting circuit, and the error signal is passed through a correlator to obtain the amplitude and correlation time of the error signal. A speech analysis and synthesis circuit according to claim 1, characterized in that: