JPH07120160B2 - Pseudo sound generator - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION "Industrial field of application" The present invention is used in place of an actual voice signal when objectively evaluating the basic performance of a high-efficiency voice encoding system. The present invention relates to a pseudo voice generating device that generates a pseudo voice in which physical characteristics are compressed and expressed.

“Prior Art” Conventionally, in this type of device, the physical characteristics of the audio signal were approximated only by the long-term average characteristics. For example, the literature (CCITT, Re
commendation P51, Red book Vol.V, 1985-1988),
As the characteristics to be satisfied by the pseudo voice, it is preferable that the frequency characteristic, the instantaneous amplitude characteristic and the waveform envelope characteristic are close to those of the actual voice signal for a long time.

However, the frequency characteristics of the current audio signal greatly vary depending on the originator, the content of the occurrence, and the like, and also temporally change. In addition, the fundamental frequency (voice pitch) characteristic of a voice signal differs greatly between males and females, and its time-varying characteristic is not constant because it repeats rising and falling. Further, in the voiced sound section, the characteristic of the voice cannot be expressed only by the instantaneous amplitude characteristic, and the rising characteristic (peak coefficient characteristic) of the signal waveform is also a characteristic amount representing an important voice signal. Therefore, it is necessary to fully consider these points when generating a pseudo voice signal, and the characteristics of various transmission devices were measured using the conventional pseudo voice signal, which specified the physical characteristics of voice only for a long time. In that case, there is a risk that a value far from the result obtained by using the actual audio signal may be obtained. In particular, when a high-efficiency speech coding apparatus that adaptively uses the characteristics of a speech signal for each short time is evaluated, it is considered that the above-mentioned influence is large when a conventional pseudo speech signal is used.

SUMMARY OF THE INVENTION An object of the present invention is to provide a pseudo-speech generator that easily generates an electrical signal having characteristics closer to the physical characteristics of an actual speech signal in order to solve the above drawbacks.

"Means for Solving Problems" The present invention is directed to a variety of spectral envelope characteristics (individuality and phonology) of a voice signal, which is not included in the above-described conventional pseudo-voice, and a fundamental frequency characteristic (gender Characteristics and time variation characteristics), peak coefficient characteristics, signal section length and pause section length, and ratio of voiced section to unvoiced section. To be realized.

According to the present invention, the signal / pause interval length generation circuit is driven by the random number sequence from the random number generation circuit to generate the signal interval length and the silent interval length according to the statistical properties of the audio signal. A fundamental frequency change characteristic having a frequency rising portion and a frequency falling portion before and after each signal section length is generated in the fundamental frequency generation circuit, and the fundamental frequency signal whose frequency changes according to the generated change characteristic is a voiced sound. It is generated by the signal generation circuit. The fundamental frequency signal and the random number sequence are output from the voiced / unvoiced determination circuit according to the voiced sound section or unvoiced sound section. An amplitude change characteristic having an amplitude rising portion and an amplitude falling portion before and after the signal section length is generated in the amplitude characteristic generating circuit, and the amplitude having the amplitude changing characteristic is multiplied by the output of the voiced / unvoiced determination circuit. A driving sound source signal is generated and supplied to the voice synthesis circuit. On the other hand, a characteristic parameter memory in which a plurality of speech characteristic parameters are stored is randomly read by a random number sequence at a constant cycle, the characteristics of the speech synthesis circuit are controlled by the read characteristic parameters, and a pseudo speech signal is output. It

[Embodiment] FIG. 1 shows an embodiment of the present invention in which a random number sequence such as an M sequence is generated from a random number generator 11, and the signal / pause interval length generation circuit 12 is drive-controlled by this random number sequence. Further, the random number sequence causes the characteristic parameter memory 13 to be read out randomly at regular intervals. This feature parameter memory
As the characteristic parameter stored in 13, for example, a characteristic parameter indicating each vector at the time of so-called vector quantization of an audio signal can be used. That is, speakers, men and women,
Obtaining the autocorrelation coefficient of the voice population that includes various types of utterance content, and setting each frame as a vector,
For example, 128 kinds of centroid vectors (reference vector) are obtained from the set of the vectors, and normally each vector of each frame of the input speech is quantized into one of the centroid vectors. The reference vector for this vector quantization, That is, the codebook for vector quantization is stored in the characteristic parameter memory 13. If necessary, the reference vector is converted into K parameters and stored in the memory 13. For example, 32 reference vectors in the characteristic parameter memory 13 are randomly read by the random number sequence from the random number generation circuit 11.

The feature parameters randomly read out from the feature parameter memory 13 at regular intervals are interpolated by an interpolation circuit, if necessary.
The voice synthesis circuit 15 is controlled by the interpolated feature parameter at 14, and the filter coefficient of the synthesis filter of the voice synthesis circuit 15 is set.

From the signal / pause interval length generation circuit 12, the signal interval length T and the pause interval length P are obtained by, for example, the expressions (1) and (2) with reference to the statistical analysis result of the voice signal.

T =-{log (1-x)} / Bts (1) P =-{log (1-x) -0.2Ats} / Ats (2) where x is a uniform random number of 0.1, and Bts and Ats are Bts = 1 / Tpb, Ats = 1 / (Ptb-0.
2), Tpb and Ptb indicate the average utterance duration and pause duration of the actual speech signal, which are usually 0.98 and 1.
30 is used.

The basic frequency characteristic generating circuit 16 is driven by the signal section T generated by the signal / pause section length generating circuit 12, and the temporal change characteristic of the basic frequency is generated in the basic frequency characteristic generating circuit 16 during that period. In accordance with the fundamental frequency change characteristic, the voiced sound signal generation circuit 17 generates a driving sound source signal having the fundamental frequency. The basic frequency change characteristic of the basic frequency characteristic generation circuit 16 is “trapezoidal”, and a frequency rising part (20% of voiced sound section) and a falling part (50% of voiced sound section) are provided before and after one signal section. The starting point Fs of the ascending part is given by equation (3).

Fs = Fc−1.3Ak (3) where Ak is the fundamental frequency change constant Ak = (Bf / 2) (VkT /
Tpb) and Fc are central fundamental frequencies, and in one example, male and female average fundamental frequencies (Fc = 125 Hz and 250 Hz) are randomly used for each signal section. Bf indicates the fundamental frequency fluctuation range, and 120 Hz is used in one example. Vk is the ratio between the voiced sound section and the unvoiced sound section, which is about 0.3 in one example. This voiced sound section and unvoiced sound section are judged by the voiced / unvoiced judgment circuit 18 by using the first-order (K ₁ ) of the K parameter representing the outline of the speech spectrum obtained from the output of the interpolation circuit 14 and using one signal. In the section (T), the unvoiced sound section is concentrated in the front part of the voiced sound section so that the characteristics of the actual voice signal can be approximated.

As the drive sound source signal from the voiced sound signal generating circuit 17 in the voiced sound section, a modified triangular wave having a basic period 20 as shown in FIG. 2 or a Rosenberg wave can be used in consideration of the rising characteristics. In one example, the descending point position 21 of the peak is randomly oscillated back and forth to bring it closer to the sound source signal waveform of the actual audio signal. These constants depend on the statistical analysis result of the voice signal. As described above, the voiced / unvoiced decision circuit 18 is determined by the K parameter first order (K ₁ ) if set up.
When it is determined that the voiced sound is generated, the voiced sound signal generation circuit 17
The drive sound source signal from is output to the multiplication circuit 22, and when it is determined to be unvoiced, the random number sequence from the random number generation circuit 11 is output to the multiplication circuit 22 through the low-frequency emphasis circuit 23. Since the fundamental frequency characteristic and the amplitude characteristic of an actual voice signal have a fairly strong correlation, the time-varying characteristic of amplitude is similar to the fundamental frequency change characteristic, so that the characteristic of the actual voice signal can be efficiently determined. Can be simulated. Therefore the signal /
During the signal section T, the amplitude special provision generating circuit 24 is driven by the output of the pause section length generating circuit 12, so that a characteristic output similar to that of the fundamental frequency characteristic generating circuit 16 is obtained.
Is supplied to the amplitude correction circuit 25, and the amplitude correction circuit 25 receives the K from the interpolation circuit 14.
The amplitude is corrected using parameters and the like, and the corrected amplitude characteristic and the voiced / unvoiced signal are multiplied by the multiplication circuit 22,
The output is used as the driving sound source signal of the voice synthesis circuit 15. In this way, a pseudo voice signal is obtained from the voice synthesis circuit 15,
The pseudo audio signal is output to the output terminal 26 as a digital signal as it is, or is sequentially output from the DA converter through the low pass filter 28 to the output terminal 29 as an analog audio signal.

An example of the amplitude characteristic and the fundamental frequency characteristic of the driving sound source signal obtained by the above procedure is shown in FIGS. 3A and 3B.

The speech synthesis circuit 15 is, for example, a PARCOR type speech synthesizer or LS.
Use P-type speech synthesizer. The characteristic parameter memory 13 may store several to several tens of types (32 types in one example) of K parameters and LSP parameters according to each composition method, so that the memory capacity is about several tens to several hundreds of words. Very few and suitable for downsizing of equipment. The extraction of the characteristic parameter can be efficiently performed from a large amount of voice signals uttered by a large number of speakers in advance by using techniques such as vector quantization and principal component analysis.

"Effects of the Invention" FIGS. 4A and 4B show waveforms and spectrum characteristics of pseudo-voice signals obtained by the conventional apparatus and the invention apparatus, respectively,
FIG. 4C shows the waveform and spectrum characteristics of an actual audio signal. Comparing these three, it is understood that the pseudo voice signal obtained by the device of the present invention is closer to the actual voice signal than that obtained by the conventional device.

An example of the improvement effect of the present invention is shown in FIG. This example uses a pseudo voice signal obtained by the device of the present invention and a conventional device to realize an actual high efficiency voice encoding device (PCM, ADPCM, AP).
This is the result of performance evaluation of CAB). The objective measures used for evaluation are the LPC cepstrum distance measure that represents the envelope distortion of the signal spectrum and the segmental SNR that represents the waveform distortion. The vertical axis of FIG. 5 represents the error in the evaluation result between the case where the actual voice signal and the pseudo voice signal are used, and the smaller the value, the higher the performance of the pseudo voice signal. From this example,
It is clear that an evaluation result closer to an actual voice signal can be obtained using the device of the present invention than using the conventional device.

As described above, according to the present invention, an electrical signal having physical characteristics closer to that of an actual voice signal can be easily realized by a small-scale hardware, and therefore, the following advantages can be obtained.

(1) It can be used as a test signal when performing a characteristic test of various transmission devices. Since the physical characteristics of the test signal are close to the actual voice signal and are compressed with good performance, it is possible to measure distortion and the like of the transmission device with good measurement accuracy and in a short time, which is useful as such a test signal. .

(2) For the same reason, it is useful for evaluation, maintenance, design, or standardization of evaluation methods of audio signal transmission devices such as a high-efficiency audio encoding device, an echo control device, and an audio detection device.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the invention device, FIG. 2 is a diagram showing an example of a driving sound source signal used in a voiced sound section of the invention device, and FIG. 3 is a driving sound source signal used in the invention device. FIG. 4 is a diagram showing an example of the amplitude characteristic and fundamental frequency characteristic of FIG. 4, FIG. 4 is a diagram showing an example of each of a pseudo voice signal and an actual voice signal obtained by the device of the present invention and a conventional device, and FIG. It is a figure which shows an example of the improvement effect at the time of using.

Claims (1)

[Claims] 1. A random number generation circuit, and a signal / pause section length generation circuit which is driven by a random number sequence from the random number generation circuit and generates a signal section length and a silence section length according to the statistical properties of a voice signal. A basic frequency generating circuit for generating a basic frequency change characteristic having a frequency rising portion and a frequency falling portion before and after each signal section length generated by the signal / pause section length generating circuit; A voiced sound signal generation circuit that generates a fundamental frequency signal whose frequency changes according to the generated change characteristic, and a basic frequency signal from the voiced sound signal generation circuit and a random number sequence from the random number generation circuit. A voiced / unvoiced decision circuit that outputs depending on whether it is a voiced sound section or an unvoiced sound section, and an amplitude increasing section and an amplitude decreasing section before and after each signal section length generated by the signal / pause section length generation circuit. An amplitude characteristic generating circuit for generating an amplitude change characteristic having a section, a characteristic parameter memory which stores a plurality of voice characteristic parameters, and which is randomly read by a random number sequence from the random number generating circuit at regular intervals, and its characteristic parameter memory A multiplication circuit for amplitude-correcting according to the read output of the above, and multiplying the amplitude having the amplitude characteristic generated by the amplitude characteristic generation circuit by the output of the voiced / unvoiced determination circuit, and the output of the multiplication circuit And a voice synthesizing circuit that outputs a pseudo voice signal, which is controlled by the characteristic parameters input from the above-mentioned characteristic parameter memory, and outputs a pseudo voice signal.