JP2001343985A - Method of voice switching and voice switch - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method of voice switching by which elimination of the beginning of a word is prevented and the quality of communication is improved. SOLUTION: In the method of voice switching, received signals are attenuated using a first attenuation amount and reproduced signals are outputted, residual signals are generated by subtracting pseudo reflected signals that are generated based on the reproduced signals from sound collected signals, the residual signals are attenuated using a second attenuation amount and made into transmitting signals, and the amounts of the first and the second attenuation are controlled based on the information of the received voice and the residual signals. In the method, probability of received voice being included in the received signals in every short time segment is computed, moreover, probability of transmitted signals being included in the residual signals is computed from received and residual signals for every short time segment. Based on these probabilities, distribution of the first and the second attenuation amounts is computed in a single or plural frequency bands and the amount of the first and the second attenuation is controlled.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used for voice communication in a loudspeaker communication system, and it is an object of the present invention to suppress the reverberation generated when a reproduced received voice is transmitted to a caller and to ensure the voice quality. .

[0002]

2. Description of the Related Art In a loudspeaker system, there is a problem of reverberation in which a loudspeaker received by a speaker is picked up by a microphone. This reverberation phenomenon is also called an acoustic echo, and when the loop gain of a closed loop formed with the loudspeaker communication system at the ground is larger than 1, howling occurs and communication becomes impossible. Also, when the loop gain is smaller than 1, adverse effects such as trouble in communication and discomfort occur.

As a method for solving such a problem of the loudspeaker communication system, a reception signal and a transmission signal are detected by calculating a signal power of a voice and treating a high power portion as a voice section.
There are a voice switch method for suppressing the reverberation by switching the amount of attenuation of the received or transmitted signal in accordance with the above, and a reverberation canceling (echo canceller) method by reverberation simulation. The echo canceling method by echo simulation cannot be completely eliminated, and is usually used in combination with the voice switch method.

[0004]

However, a loudspeaker communication system or a TV conference system having a voice switch method for switching the amount of attenuation between the transmitting side and the receiving side by a simple power comparison has a problem in that the beginning of the voice is cut off. Was. This is because, in a call environment in which a noise source is present in the surroundings or in a call environment in which the amount of echo cancellation is insufficient, when the transmission / reception determination is incorrect, the attenuation on the transmission / reception side is rapidly switched. In particular, since the voice power of a consonant is less than 1/10 of a vowel, voices that start with a consonant are likely to be truncated. The relative comparison of the power of each sound in speech is described in, for example, "Hearing and Speech" (edited by T. Miura, edited by the Institute of Electronics, Information and Communication Engineers, 1979, pp. 295-297).

On the other hand, in the voice section detection method for voice recognition, voice pitch information instead of voice power or LPC (linear
It is known that by performing detection based on a feature vector including a speech parameter by predictive coding (spectral analysis), the amount of calculation is greatly increased, but the consonant detection ability and the detection accuracy of a speech section are improved. For example, in the voice section detection method disclosed in Japanese Patent Laid-Open No. 4-251299, a characteristic peak is a spectrum peak in a low frequency band and an average spectrum in a high frequency band after LPC spectrum analysis.

[0006]

SUMMARY OF THE INVENTION Accordingly, the present invention provides a method for preventing the beginning of a speech from occurring even in a voice starting with a consonant.
A feature vector is extracted from the received signal and the transmitted signal, and a probability that the received voice and the transmitted voice are included for each short time interval is determined. Controls the distribution of the inserted attenuation.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS (First Embodiment) FIG. 1 shows a configuration example of a voice switch according to the present invention. The reception signal input to the voice switch unit 1 is band-divided by an N band division filter 141 in a reception attenuation unit 14, and the N band reception attenuation unit 1
At 42, each band is attenuated, and the N band synthesis filter 14
3 reproduces a reproduced signal after being subjected to band synthesis, and is converted into an acoustic signal by the reproducer 3. The echo canceller 2 subtracts the pseudo echo signal generated by the pseudo echo generator 21 from the reproduced signal from the sound pickup signal of the sound pickup device 4 by the subtractor 22 to reduce the received echo component and obtain a residual signal. The residual signal is band-divided by the N-band splitting filter 151 in the transmission attenuation unit 15, attenuated for each band by the N-band transmission attenuation unit 152, band-synthesized by the N-band synthesis filter 153, and transmitted. It becomes a talk signal.

[0008] The reception voice probability calculation unit 11 calculates the probability that the reception voice signal includes the reception voice in the current short time interval,
The transmission voice probability calculation unit 12 calculates the probability that the transmission signal is included in the residual signal in the current short time period from the reception signal and the residual signal. In addition, the transmission / reception attenuation distribution calculating unit 13 determines the attenuation distribution ratio p to the receiving side based on the probability r that the reception voice exists in the short period of the reception signal and the existence of the transmission voice in the short period of the transmission signal. Based on the probability s, the function p = s (1−r / 2) shown in FIG. 2, for example. Normally, in a voice switch, the product of the attenuation of the reception attenuation unit 14 and the attenuation of the transmission attenuation unit 15 is controlled to be constant. For example, a value obtained by multiplying the reference value by the attenuation distribution ratio is determined as the attenuation of the reception attenuation unit 14. The attenuation of the transmission attenuation unit 15 is determined as a value obtained by dividing the reference value by the attenuation distribution ratio. By controlling the distribution of attenuation in this way,
When there is no transmission (speech) (when s = 0), the attenuation on the transmitting side is the maximum, and as the existence probability of the transmitted voice increases, the attenuation on the transmitting side decreases. In these processes, an input signal is divided into N frequency bands and independently performed for each band, and attenuated components for each band are combined into the entire band. With such a configuration, the attenuation can be controlled so that the loop gain does not exceed 1 for only some of the bands, and the control of the attenuation with a finer grain can be realized.

FIG. 3 shows a configuration of the received voice probability calculator 11 and the transmitted voice probability calculator 12 for calculating the probability that the received signal includes the received voice and the residual signal includes the transmitted voice. On the receiving side, the k-th short-term section of the received signal is the full-band power PT determined from the short-term section by the M-dimensional feature vector extraction unit 11B, the power ratio γ between the high band and the low band, the partial self-phase. Relationship number
K1, a four-dimensional feature vector consisting of the maximum value ρ of the autocorrelation coefficient of the linear prediction analysis residual

[0010]

(Equation 1) Is converted to According to Japanese Patent Application Laid-Open No. 4-100099, when the noise level NS is known, the entire band power PT obtained from the short time period of the input signal, the power ratio γ between the high band and the low band, the partial autocorrelation coefficient From K1, the maximum value ρ of the autocorrelation coefficient of the linear prediction analysis residual, a voiced sound
Unvoiced sound and a silent section can be determined as follows. (1) Since a voiced sound generally has a higher power than an unvoiced sound, it is compared with a noise level NS measured in advance, and γ> 0
When PT> NS + 30 dB, voiced sound is assumed. In addition,
Since this condition does not function effectively unless the S / N ratio is sufficient, attention is paid to the periodicity of the signal in that case. That is,
A voiced sound has periodicity due to vocal cord vibration, and the time delay (period) corresponding to ρmax corresponds to the fundamental frequency of the vocal cord, and the magnitude depends on the periodicity of the vocal cord vibration. When the periodicity of the signal is ρmax> 0.25, the voiced sound is determined. (2) The partial autocorrelation coefficient K1 indicates the slope of the spectral outline of the voice, and the smaller the value is, the flatter the spectrum is. Therefore, it is assumed that the voice is unvoiced when K1 <0.3. (3) The condition of a silent section is PT <NS, and when the condition is satisfied, the determination of the corresponding frame is determined to be a silent section.

A feature vector time-series pattern determining unit 11
In D, the feature amount vector of [Equation 1] is classified into a voiced section (V), an unvoiced section (U), and a non-speech section (N) by the above-described determination method. The probability that the received voice is included in the short-term section of the k-th received signal is, for example, the time-series pattern including the voiced section (V), the unvoiced section (U), and the non-voice section (N) and the hidden pattern shown in FIG. Calculated from Markov model. On the transmitting side, the residual signal is converted into a feature vector by the M-dimensional feature vector extraction unit 12B.

[0012]

(Equation 2) In the feature vector time-series pattern determination unit 12D, for example, a 16-dimensional vector obtained by combining the feature vectors of the four blocks up to now is obtained.

[0013]

(Equation 3) Is calculated. If this is less than or equal to a predetermined threshold, it is determined that the residual signal is almost a received echo and does not include a transmitted signal. If the threshold value is exceeded, it is determined that the transmission signal is included in the residual signal, and the voiced sound section (V), unvoiced sound section (U), and silent section (N) extracted from the residual signal by the above-described determination method. )
Is calculated from the time-series pattern consisting of and the hidden Markov model.

In this embodiment, a four-dimensional feature quantity vector for classifying a speech section into three is used.
It is also possible to use a feature vector having a larger dimension. FIG. 4 shows each state transition of the hidden Markov model used in the embodiment of the present invention. The probability that the voice is included in the short time interval is calculated as follows using the hidden Markov model of FIG. This Markov model has four states S0,
A transition is made between S1, S2, and S3, and a silent section (N), an unvoiced section (U), or a voiced section (V) is output at the transition. In each of the states S0 to S3 of the Hidden Markov Model, the probability of including a voice when staying in that state is set in advance. For example, in the Markov model of FIG. 4, the voice probability is set to 0% for S0, 30% for S1, 70% for S2, and 100% for S3.

The numbers above the arrows in FIG. 4 indicate the state transition probabilities and the output probabilities of the silent section (N), unvoiced section (U) or voiced section (V). For example, the transition probability from the state S0 to the state S1 is 0.7, the probability of outputting the voiced sound section (V) during this transition is PV = 0.0, and the probability of outputting the unvoiced sound section (U) is PU =
1.0, the probability of outputting a silent section (N) is PN = 0.0. The transition probability from the state S0 to the state S2 is 0.3, and during this transition, a voiced sound section (V) is output with a probability of 100%. State S
The transition probability from 1 to the state S2 is 0.6, and the probability of outputting the unvoiced interval (U), voiced interval (V), and silent interval (N) during this transition is PV = 0.8, PU = 0.2, PN = 0.0.

Now, the probability that a time series pattern of U → U → V is extracted and exists in each of the states S0 to S3 is represented by
Pn = (S0n S1n S2n S3n). For example, P1 is a probability that U is extracted first and exists in each state of S0 to S3, and P3 is a time series pattern UUV extracted and S0 to S3.
The probability of being in each state of 3. The initial state is P0 = (10
When set to (0 0), P1 = (0 1.0 × 0.7 × S0 ₀ 0.0 × 0.3 × S0 ₀ 0) = (0 0.7 0 0) P2 = (0 0.8 × 0.4 × S1 ₁ 0.2 × 0.6 × S1 ₁ 0) = (0 0.224 0.12 0) P3 = (0 0.8 x 0.6 x S1 ₂ + 0.8 x 0.4 x S2 ₂ 1.0 x 0.6 x S2 ₂ ) = (0 0 0.1459 0.072).

When the time series UU is extracted, the probability of being in state S1 is 0.224 / (0.224 + 0.12) and the probability of being in state S2 is 0.12 / (0.224 + 0.12). The probability that a voice is included in a short time period can be calculated by the following equation, for example. 30% × (0.224 / (0.224 + 0.12)) + 70% × (0.12 // 0.224 +
0.12)) ≒ 44% The probability estimation from the observed symbol sequence using the hidden Markov model is described in detail in, for example, Seiichi Nakagawa, “Speech Recognition by Probability Model”, edited by IEICE, Chapter 3.

A Markov model used for speech processing as shown in FIG. 4 generally has an initial state and a final state, and does not have a backward transition to the state S0. Therefore, after a while from the start of the voice, the probability that the voice exists in a state other than the final state becomes almost zero, and the voice remains in the final state even in the next voice section. To prevent this, the hidden Markov model is initialized when it is detected that a silence period (N) has continued for a certain number in the reception signal. Further, when it is detected from the residual signal that the silence section (N) has continued for a certain number, the hidden Markov model on the transmitting side is initialized. As a result, the hidden Markov model that has already reached the final state can be returned to the initial state, and the voice switch can be controlled from the beginning of the next voice section.

In this embodiment, a hidden Markov model having four states and outputting three symbols is used. However, it is also possible to use a hidden Markov model having a larger number of states and a larger number of output symbols. Furthermore, when speech coded at high efficiency using, for example, linear prediction is transmitted through a transmission path, each element of a speech code including a speech spectrum outline and a pitch period parameter is used as a feature amount of a speech switch. . This makes it possible to implement the voice switch of the present invention with a smaller amount of processing as compared with a method of extracting a plurality of feature amounts from a decoded signal.

[0020]

As described above, according to the present invention, a feature amount vector that more reflects the features of the voice is extracted from the reception signal and the residual signal at regular time intervals, and the current short-term value is extracted based on the time series. The probability that a voice is included in the time voice section is calculated, and the distribution of the amount of attenuation of the reception / transmission is controlled based on the probability. For this reason, the voice switch of the present invention does not cut off the beginning of the voice even in voices starting with a consonant having a low voice power, as compared with the conventional voice switch method of switching the attenuation amount between the transmitting side and the receiving side by a simple power comparison. Can be.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of a voice switch of the present invention.

FIG. 2 is a diagram showing a function for determining a receiving-side attenuation distribution ratio p from a received voice existence probability r and a transmitted voice existence probability s.

FIG. 3 is a diagram showing a configuration of a reception voice probability calculation unit and a transmission voice probability calculation unit.

FIG. 4 is a diagram showing an example of a hidden Markov model used in the embodiment.

Continuation of the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat II (reference) H04M 9/08 G10L 9/00 C H04R 3/02 D (72) Inventor Shigeaki Aoki 2-3 Otemachi, Chiyoda-ku, Tokyo No. 1 Nippon Telegraph and Telephone Corporation F-term (reference) 5D015 DD03 5D020 CC05 5K027 BB03 DD07 DD10 5K038 CC00 DD06 FF13

Claims (8)

[Claims] An attenuated received signal is attenuated by a first amount to output a reproduced signal, a pseudo echo signal is generated based on the reproduced signal, and the pseudo echo signal is subtracted from a collected signal to obtain a residual signal. A voice switching method for attenuating the residual signal by a second attenuation amount to generate a transmission signal, and controlling the first and second attenuation amounts based on information on the reception signal and the residual signal. In the control of the first and second attenuation amounts, a step of calculating a probability that a received voice includes a received voice signal for each short time section, and a step of calculating a residual error from the received signal and the residual signal for each short time section Calculating a probability that the transmitted voice is included in the signal; and calculating a distribution of the first and second attenuation amounts in a single or a plurality of frequency bands based on the probability. Switch method. 2. The voice switching method according to claim 1, wherein the step of calculating the probability that the received voice signal includes the received voice for each short time period comprises extracting a feature vector from the short time period of the received voice signal. And a step of calculating the probability that the received voice is included from the time series of the feature amount vector. The step of calculating the probability that the transmitted voice is included in the residual signal for each short time interval includes: A voice switch, comprising: extracting a feature vector from a short time period of a difference signal; and calculating a probability that a transmitted voice is included from a time series of the feature vector. 3. The voice switching method according to claim 2, wherein the step of calculating a probability that a voice is included for each short time period from the time series of the feature amount vector includes a probability calculation using a “Hidden Markov Model”. And a step of initializing a hidden Markov model used for calculating a probability when a non-voice section is detected. 4. The voice switching method according to claim 2, wherein the reception signal and the transmission signal are coded signals,
A voice switch method, wherein the feature vector comprises a code element. 5. A receiving attenuator for attenuating a received signal by a first attenuation to output a reproduced signal; generating a pseudo echo signal by referring to the reproduced signal; and generating the pseudo echo signal from the collected sound signal. An echo canceller that outputs a subtraction residual signal, a transmission attenuator that attenuates the residual signal by a second attenuation and outputs a transmission signal, and inputs the received signal and the residual signal, A voice switch including a control unit for controlling the first and second attenuations, wherein the control unit calculates a probability that the received signal includes the received voice for each short time interval; A transmission voice probability calculation unit for calculating a probability that the transmission signal is included in the residual signal from the reception signal and the residual signal for each time section; and a first or a plurality of frequency bands based on the calculated probability. And transmission / reception talk attenuation distribution calculation for calculating the distribution of the second attenuation An audio switch comprising an output unit. 6. The voice switch according to claim 5, wherein said received voice probability calculation unit includes: a means for extracting a feature vector from a short time period of the received signal; and a received voice from a time series of the feature vector. Means for calculating a probability to be transmitted, wherein the transmission voice probability calculation unit comprises: means for extracting a feature vector from a short period of a received signal and a residual signal; and transmission voice from a time series of the feature vector. A voice switch, comprising: means for calculating a probability that a character is included. 7. The voice switch according to claim 6, wherein the means for calculating a probability that a voice is included for each short time period from the time series of the feature amount vector includes calculating a probability using a “hidden Markov model”. And a means for initializing a hidden Markov model used for calculating a probability when a non-voice section is detected. 8. The voice switch according to claim 6, wherein the reception signal and the transmission signal are coded signals,
An audio switch, wherein the feature vector comprises a code element.