JPS5918718B2 - Speech analysis/synthesis system - Google Patents

Description

[Detailed description of the invention] The present invention relates to a speech analysis/synthesis system.

In recent years, advances in speech analysis equipment such as linear predictive coding methods for speech signals have made it possible to encode and transmit speech signals at a low bit rate, and then synthesize the original speech signal on the receiving side based on information from the transmitting side. A speech analysis/synthesis system has been realized. However, in conventional systems of this kind, the voice restored by synthesis on the receiving side has a high degree of clarity but is extremely poor in naturalness, resulting in a so-called mechanical voice.
It had the disadvantage that the individuality of the speaker was lost.

In other words, conventionally, a speech analysis device extracts the spectral envelope information and pitch frequency information shown by the broken line from the audio signal shown by the solid line in Figure 1a, and compares these with the average power information and the voiced/unvoiced discrimination information. The voice synthesizer on the receiving side synthesizes a voice signal as shown in FIG. 1 based on this information.

More specifically, the spectral envelope information is extracted and transmitted as LPC parameters by a so-called analysis filter, and the receiving side uses a synthesis filter whose frequency characteristics are determined by the LPC parameters as shown by the broken line in Figure 1b. was used for synthesis. In this case, L.P.
As the number of C parameters increases, the spectral envelope of the reproduced audio signal shown by the broken line in FIG. 1b becomes asymptotic to the spectral envelope of the original audio signal shown by the broken line in FIG. 1a. However, in this conventional method, only pitch frequency information was extracted and transmitted as information regarding the pitch frequency component of the audio signal, so the fine structure of the spectrum of the original audio signal was It has a maximum value at the pitch frequency interval, and there is a slight spread of the frequency spectrum before and after the maximum value, whereas the reproduced audio signal has a complete line spectrum, as shown by the solid line in Figure 1. - This was a major cause of the lack of naturalness.

The present invention provides a speech analysis/synthesis system that is equipped with a means for extracting information regarding the fine structure of the spectrum of an audio signal, thereby making it possible to analyze the audio signal more accurately and re-voice synthesized speech with excellent naturalness. The purpose is to

Since the reproduced audio signal (t) shown in FIG. 1B is a set of line spectra having components at integral multiples of the pitch frequency Fp, it can be expanded into a Fourier series as shown below.

On the other hand, the original audio signal S(t) shown in FIG. 1a is expressed as follows.

Here, 6k is a constant, but Cn(t) is a function of time t and consists of low frequency components. This low frequency component is generated by minute vibrations of the pitch frequency, and the 1cn(t river) determines the fine structure of the spectrum.The gist of this invention is to extract information regarding this 1cn(t river.Hereinafter, this The invention will be explained in detail by way of examples.

FIG. 2 is a diagram showing a speech analysis/synthesis system according to an embodiment of the present invention. In the voice analysis device 10 provided on the transmission side, the voice signal introduced into the input terminal 1 is D/
After being converted into a digital signal by the A converter 11, it is applied to an analysis filter 12, an average power calculation circuit 13, and a spectral fine information extraction circuit 14. The analysis filter 12 extracts LPC (Linear Predictive Coding) parameters as the spectral envelope information of the audio signal, and is already known (for example, in the Acoustical Society of Japan Conference Proceedings 2-2-5, published in October 1972). ). The analysis filter 12 outputs a residual signal in addition to the LPC parameter C, and supplies this residual signal to the pitch information extraction circuit 15. Pitch information extraction circuit 15
determines whether it is a voiced or unvoiced sound from the residual signal, and for example, the discrimination information B is "1" for a voiced sound and 0 for an unvoiced sound.
In addition, in the case of a voiced sound, pitch frequency information A is also output.

The average power calculation circuit 13 is composed of, for example, a RAM (random access memory) and an adder, stores input signals for a certain period of time in the RAM, calculates the sum of their absolute values, and outputs information D proportional to the average power. do. The spectral fine information extraction circuit 14 is newly added according to the present invention, and in this example, it includes an absolute value circuit 141 as a nonlinear circuit for superimposing the spectrum of the audio signal on the low frequency band, and a low-pass filter 142. It is composed of.

The absolute value circuit 141 is a circuit that takes the absolute value of the audio digital signal input from the A/D converter 11. For example, if the input digital signal is a polar amplitude display, the most significant bit is set to "0", and its complement is For display, when the most significant bit is "1", the most significant bit is inverted, and "1" is added to the signal with the most significant bit inverted. Low-pass filter 1
42 is a filter having a cutoff frequency lower than the pitch frequency of the audio signal and higher than the syllable change frequency;
In this example, it is a digital filter. At the output of this low-pass filter 142, information E corresponding to the fine structure of the spectrum of the audio signal is obtained. Pitch frequency information A, voiced and unvoiced sound discrimination information B, which are obtained in the form of digital signals as described above. The LPC parameter C, average power information D, and spectral fine information E are input to the encoding circuit 16, where they are encoded with an appropriate bit rate, and a synchronization signal is added as necessary, and sent from the output terminal 2 to the transmission line 3. Sent out.

In this case, since the newly added spectral fine information E is a signal in a low frequency band, it can be transmitted at a low bit rate like other information. On the other hand, in the speech synthesis device 20 provided on the receiving side, the signal transmitted to the input terminal 4 via the transmission line 3 is decoded by the decoding circuit 21, and each information of A, B, C, D, and E is decoded. is separated and supplied to each part.

The impulse sequence generator 22 is a circuit that generates an impulse sequence at intervals equal to the pitch based on the pitch frequency information A, and the white noise generator 23 is a circuit that generates a random pattern sequence at a bit rate higher than the pitch frequency.
These series signals are input to the switch circuit 24. The switch circuit 24 is switched by voiced/unvoiced sound discrimination information B, and when this information B is "1-", that is, an unvoiced sound, the output signal of the impulse sequence generator 22 is switched;
In other words, in the case of unvoiced sounds, the output signals of the white noise generator 23 are passed through and supplied to the first multiplier 25 . The first multiplier 25 multiplies the signal input from the switch circuit 24 by the spectral fine information E, and is newly provided in conjunction with the provision of the spectral fine information extraction circuit 14. The multiplication result of the first multiplier 25 is input to the second multiplier 26 , where it is further multiplied by the average power information D, and then input to the synthesis filter 27 . The synthesis filter 27 is a digital filter whose frequency characteristics are determined by the LPC parameter C, and its output signal is converted back to an analog signal by the D/A converter 28 and then sent to the output terminal via the low-pass filter 29. 5 as a reproduced audio signal. In the manner described above, audio signals are analyzed, transmitted, and synthesized.

Now, in the above embodiment, in the case of a voiced sound, the information E output from the vector fine information extraction circuit 14 is expressed as the above Cn(
Although it is not the t-river itself, M Σ 1cn (information that has an extremely high correlation with the t-river; n = blind).

The spectrum of the signal obtained by multiplying this information by the impulse sequence supplied from the impulse sequence generator 22 via the switch circuit 24 in the first multiplier 25 in the speech synthesizer 20 is as shown in FIG. 3a. Then, Fig. 3b
When the signal passes through the synthesis filter 27 having the frequency characteristics shown by the broken line in FIG. 3B, the reproduced audio signal has a spectrum shown by the solid line in FIG. This spectrum is very similar to the spectrum of the original audio signal shown by the solid line in FIG. 1a, and therefore highly natural sound is reproduced. In the case of unvoiced speech, the output signal of the white noise generator 23 in the speech synthesis device 20 is input to the first multiplier 25 via the switch circuit 24, so that the output signal of the multiplier 25 is converted into spectral fine information. Since the pattern sequence is random regardless of the state of E, the multiplier 25 has no effect on the operation.

As detailed above, the speech analysis device according to the present invention is capable of extracting information regarding the fine structure of the spectrum of the audio as one element of information regarding the audio signal.
It is possible to reproduce extremely natural sounds during synthesis.

Furthermore, if the speech analysis device of the present invention is applied to the input stage of a speech identification device, there is an advantage that speaker recognition can be performed more easily and accurately than before by comparing spectral fine information. FIG. 4 is a diagram showing another embodiment of the present invention, and the difference from FIG. 2 is that the audio signal introduced into the input terminal 1 is input to the spectral fine information extraction circuit 14'' in the form of an analog signal. , the output thereof is supplied as spectral fine information E to the encoding circuit 16 via the A/D converter 17.According to this embodiment, the absolute value circuit 141''
Another advantage is that the low-pass filter 147 can be easily constructed using an analog circuit. In this case, the A/D converter 17
It can also be omitted by using the D converter 11 in a time-sharing manner. Note that the present invention can be implemented with various modifications; for example, the input signal to the spectral fine information extraction circuit is not limited to the audio signal itself, but may be any signal containing pitch frequency components; It may also be a residual signal.

Furthermore, a low-pass filter having an appropriate cut-off frequency may be inserted before the absolute value circuit. Moreover, the above absolute value circuit is 2
It may be replaced with any other suitable nonlinear circuit such as a multiplier circuit or a half-wave rectifier circuit.In short, it has the effect of superimposing (concentrating) the spectrum of the input signal on the low frequency band, and its M output is passed through a low-pass filter. If passed through Σ (Cn(t)
1 is a schematic configuration diagram.

10...Speech analysis device, 14... Spectrum fine detailed information extraction circuit, 20...Speech synthesis installed.

Claims (1)

[Claims] 1. A speech analysis device that extracts and transmits pitch frequency information, voiced/unvoiced speech discrimination information, spectral envelope information, and average power information from an input speech signal, and receives each of the above information transmitted from this speech analysis device. After extracting an impulse sequence and a white noise sequence generated based on pitch frequency information based on voiced/unvoiced sound discrimination information, the original audio signal is obtained by multiplying by average power information and adding spectral envelope information. In a speech analysis/synthesis system comprising a speech synthesis device, the speech analysis device receives the input speech signal or a signal containing a pitch frequency component in the speech signal, and converts the spectrum of the signal into a low frequency band. and a low-pass filter provided on the output side of the nonlinear circuit and having a cutting frequency lower than the pitch frequency of the input audio signal and higher than the syllable change frequency. , wherein the speech synthesis device is provided with means for multiplying the impulse sequence and the white noise sequence extracted based on the voiced/unvoiced sound discrimination information by the information from the spectral fine information extraction circuit. Synthesis system.