JP2005077889A - Voice packet absence interpolation system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To interpolate a voice signal included in a voice packet which is lost during transmission. <P>SOLUTION: Linear prediction coefficients are calculated from voice signals normally received before and after they are lost. The prediction coefficient in a forward direction, i.e. a precedent direction is calculated from the voice signal right before the loss is calculated and the linear prediction coefficient in a backward direction, i.e. traced back is calculated from the voice signal right after the loss. A sample which is one sample after the forward linear prediction coefficient is predicted from the voice signal before the loss and a sample which is further one sample after is predicted from the voice signal before the loss. Those operations are repeated to predict all samples of the lost part. on the other hand a sample which is one sample before, is predicted from the voice signal after the loss and the backward prediction coefficient and then a sample which is further one sample before is predicted from the predicted sample and the voice signal after the loss. Those are repeated to predict all samples of the lost part. There is provided a means for obtaining a high quality voice sample for interpolating the lost part by averaging the predicted sample of the lost voice of these two kinds. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to interpolation of missing portions of a speech signal, and more particularly to an interpolation method using linear prediction.

As a missing interpolation method that occurs when an audio signal is accommodated in a packet or the like and transmitted, conventionally, the pitch immediately before the missing portion is estimated, and the audio signal for one pitch immediately before the missing portion is repeatedly interpolated by the missing portion as many times as necessary. ITU standard G. 711 Appendix I.
ITU standard G. 711 Appendix I

  In the above prior art, if the missing section becomes longer, the number of repetitions increases, and the sound quality of the interpolation portion becomes synthetic and unnatural.

  Therefore, an object of the present invention is to provide a method in which the sound quality of the interpolation portion does not become unnatural even if the missing section becomes long.

  According to the present invention, there is provided a missing interpolation method provided with means for obtaining an interpolation signal of a missing part by linear prediction from speech immediately before missing. Further, according to the present invention, there is provided a missing interpolation method including means for obtaining an interpolation signal of a missing portion by linear prediction from speech immediately before and after the missing.

According to the present invention, it is possible to interpolate audio signals included in packets lost during transmission with natural sound quality. For example, when 30% of voice packets are lost, the average subjective sound quality of five levels is lower than 2 in the conventional method, whereas the sound quality of 2.4 can be maintained according to the present invention.

  In the present invention, linear prediction coefficients are first calculated from normally received speech signals before and after the loss. At this time, the prediction coefficient in the forward direction, that is, the preceding direction is calculated from the speech signal immediately before the loss, and the linear prediction coefficient is calculated backward from the speech signal immediately after the loss, that is, by going back in time. Next, a sample one sample ahead is predicted from the forward linear prediction coefficient from the speech signal before missing. Next, a sample one sample ahead is predicted from the predicted sample and the audio signal before missing. This is repeated to predict all samples of missing parts. On the other hand, the previous sample, that is, the last sample of the missing portion is predicted from the missing audio signal and the backward prediction coefficient. Next, a sample one more sample before is predicted from the predicted sample and the voice signal immediately after the loss. This is repeated to predict all missing samples backwards. The two types of missing speech prediction signals are averaged to obtain speech samples that interpolate the missing portions.

  Examples of the present invention will be described below with reference to the drawings, but it goes without saying that the scope of the present invention is not limited thereto.

  FIG. 1 shows a first embodiment of the present invention. Audio signals of a predetermined number of samples are collected in a format called one packet, added with additional information such as a destination, and transmitted over the network for each packet. There are several relay points during network transmission. At each relay point, each packet is once stored in a memory together with a packet transmitted from another location, and is sent to the next relay point in the order received. At this time, if the packets are concentrated and sent to the relay point, the memory is limited and the packets are discarded and the stored voice signal is lost. Even if the packet is not rejected, a long waiting time is generated until the packet is concentrated and sent to the next relay point. Even if the final arrival point is reached at any time, the packet is discarded if it is not in time for the playback time of the accommodated audio signal due to a long transmission delay. As described above, in the transmission of the audio signal using the packet, the audio signal may be lost mainly due to the above two types of factors.

  Voice packets that have reached the final arrival point are first stored in a buffer. At this time, the serial number given to the packet is monitored to determine the presence or absence of a missing packet, and this is sent to the missing interpolation unit and the playback audio source changeover switch. If the packet arrives safely and is in time for playback, the packet is disassembled, and if the next packet is not missing, it is sent to the speaker and played as it is. However, when it is determined that the next packet is missing, this audio signal is reproduced by the speaker after being smoothed by the missing interpolation unit. When it is determined that a packet to be reproduced at present is missing, a new packet is not read out, and a missing portion audio signal is generated from the accumulated audio signal immediately before the loss.

  FIG. 2 shows the configuration of the missing interpolation part in the first embodiment. The decomposed packet is first stored in a buffer. This buffer always stores the audio signal contained in the two recently received and reproduced packets. First, a linear prediction coefficient is calculated by a linear prediction coefficient calculation unit from a voice signal included in the latest voice packet. The prediction coefficient can be calculated by, for example, the Levinson-Durbin algorithm. For this algorithm, for example It is described in detail in “Adaptive Filter Theory” by Haykin (Prentice-Hall, 1996, item 254).

  Next, the audio signal included in the latest packet is first preset in the shift register. Based on this signal and the calculated prediction coefficient, the first speech signal of the missing portion is first predicted. This is shown in FIG. Next, the next sample is predicted from the predicted one sample and the audio signal stored in the shift register. This process is repeated to predict a voice signal having the number of samples corresponding to the missing packet.

  When the linear prediction is repeated using a predicted sample, the amplitude gradually decreases. In order to correct this, a variable gain is used as shown in FIG. This gain is first set to 1, is increased uniformly, and finally increases to a preset upper limit value. As a result, the amplitude of the predicted sample is kept very close to the original sound.

The audio signal included in the received packet has different characteristics from the predicted audio signal. Therefore, in order to smooth this, the voice signal included in the packet immediately before the missing voice packet is predicted in the same manner as described above by using the linear prediction coefficient. This predicted sample is multiplied by a weight 1 (ω ₁ ) that approaches 0 in the first sample and gradually approaches 1 in the last sample of the packet. Then, the voice sample actually received just before the missing is added to a value obtained by multiplying the weight 2 (ω ₂ ) gradually approaching 0 in the final sample from 1 in the first sample. Here, ω ₁ + ω ₂ = 1. This is shown in FIG. With this method, the received voice sample is multiplied by the weight of the predicted voice as the sample is far from the missing portion, and the predicted voice is multiplied as it becomes closer to the missing portion, leading to a smoothly predicted missing portion. The same smoothing is also applied to the voice immediately after missing. That is, using the predicted missing voice signal, the voice signal immediately after the missing is further repeatedly predicted. The predicted voice sample is multiplied by a weight 1 (ω ₁ ) that gradually approaches 0 in the final sample of the voice packet immediately after the missing in the sample immediately after the missing, while the received voice signal is multiplied in the first sample. Is multiplied by a weight 2 (ω ₂ ) that gradually approaches 1 from 0 in the final sample, and these two types of audio signals are added and output as an audio signal to which smoothing is applied. Immediately after the loss by this method, a large amount of weight is given to the predicted audio signal, and the weight is gradually transferred to the audio signal that has been received.

According to the first embodiment described above, it is possible to interpolate the voice signal for the voice packet missing portion with the predicted voice signal having natural sound quality.

Next, in the second embodiment, missing interpolation is performed using both a prediction signal obtained from the speech signal immediately before the loss using forward prediction and a prediction signal obtained from the speech signal immediately after the loss using backward prediction. explain. FIG. 5 shows a missing complement portion of this embodiment. The overall configuration is as shown in FIG. The forward missing interpolation part composed of the forward packet buffer, shift register, forward linear prediction speech coefficient calculation, forward linear prediction and variable gain is the same as that of the first embodiment.

On the other hand, the audio signal is also stored in the backward packet buffer. A backward linear prediction coefficient for predicting the missing part is calculated from the audio signal stored in the buffer immediately after the missing. For this purpose, the Levinson-Durbin algorithm may be applied by reversing the time of the audio signal for one packet stored in the buffer. Next, the audio signal immediately after the loss with the time axis reversed is first preset in the shift register. Based on this signal and the calculated prediction coefficient, the last speech signal of the missing part is first predicted. This situation is the same as in FIG. 3, but prediction is performed by gradually going back from the last sample in which the time axis is reversed, that is, the missing part. Next, a predicted one sample and a sample immediately before both of the audio signal stored in the shift register are predicted. This process is repeated to predict a voice signal having the number of samples corresponding to the missing packet. In the backward prediction, as in the forward prediction, since the amplitude decreases when the prediction is repeated, a variable gain that gradually increases the gain is applied.

In this way, two types of speech signals obtained from forward prediction and backward prediction for the missing portion are obtained. The interpolated signal is obtained by multiplying the first part of the missing part by the former and the latter part by multiplying the latter by multiplying the weights. This is shown in FIG. The forward prediction signal of the missing part is multiplied by a weight 2 (ω ₂ ) that becomes 1 immediately after the missing and gradually becomes 0. On the other hand, the backward prediction signal is multiplied by a weight 3 (ω ₃ ) that becomes 0 immediately after being lost and gradually becomes 1. In the missing section, ω ₂ + ω ₃ = 1. The sum of these two types of weighted signals is used as an interpolation signal. As a result, in the first half of the missing period, where the number of prediction iterations is small and the accuracy of forward prediction is still high, there are many forward prediction voices. On the other hand, in the latter half, where the accuracy of backward prediction is high for the same reason, many backward prediction voices are included.

As in the first embodiment, the signal before missing further smoothes the voice predicted from the voice signal one packet before and the voice signal of the received packet. On the other hand, the audio signal immediately after the loss is similarly smoothed. That is, the audio signal is predicted backward from the audio signal one packet ahead. This is multiplied by a weight and added to the audio signal received immediately after the loss, and smoothed. At this time, immediately after the loss, the weight 3 (ω ₃ ) of the predicted speech signal is set to 1, and this is gradually set to 0. On the other hand, the received signal weight 1 (ω ₁ ) is set to 0 immediately after being lost, and is gradually set to 1. Again, ω ₃ + ω ₁ = 1. The sum of signals obtained by multiplying the above two types of weights is defined as an audio signal. As a result, immediately after the loss, the sound signal includes many characteristics of the backward predicted sound signal and gradually includes many characteristics of the received sound signal.

According to the second embodiment described above, it is possible to interpolate the lack with a more natural and high-quality speech signal by using the normally received speech signal immediately after the lack.

Next, a description will be given of a third embodiment in which the forward predicted speech signal shown in the first embodiment and the bidirectional predicted speech signal using both forward and backward directions shown in the second embodiment are switched.

In the second embodiment, high quality interpolated speech can be obtained even if there are many missing parts. However, since the audio signal included in the lost packet is used, it is necessary to wait for reception of this packet. That is, a long delay is required. On the other hand, in the forward prediction, there is no need to wait for packet reception after the loss, so that a long delay is not necessary. Until 10% loss, there is not much difference in sound quality between forward prediction and bidirectional prediction. Therefore, the missing rate is monitored, and the forward predictive interpolation illustrated in FIG. 2 and described in the first embodiment is adopted up to a missing rate of about 10%, and more than that, illustrated in FIG. 5 and described in the second embodiment. Bi-directional prediction. This is shown in FIG.

The configuration of this embodiment is substantially the same as that of the first embodiment, but the missing signal output from the buffer is input to the missing rate calculation unit. The missing rate calculation unit estimates the moving average of the missing rate from the missing signal. The forward predictive interpolation described in the first embodiment is performed until the missing rate is about 10%, and this is reproduced from the speaker. On the other hand, when it becomes 10% or more, the bidirectional predictive interpolation described in the second embodiment is performed, and this is reproduced from the speaker.

According to the third embodiment, when the missing rate is low, high-quality missing interpolation is performed with a small delay, and when the missing rate is high, a slightly longer delay is allowed and high-quality speech interpolation can be performed. .

It is a figure which shows the 1st Example of this invention. It is a figure which shows the structure of the missing interpolation part in 1st Example of this invention. It is a figure explaining the operation | movement of the linear prediction interpolation in 1st Example of this invention. It is a figure explaining the operation | movement of the smoothing in 1st Example of this invention. It is a figure which shows the structure of the missing interpolation part in 2nd Example of this invention. It is a figure which shows the operation | movement of the smoothing in 2nd Example of this invention. It is a figure which shows the structure of the missing interpolation part in the 3rd Example of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... Buffer 12 ... Missing interpolation part 21 ... Linear prediction coefficient calculation part 22 ... Linear prediction part

Claims (8)

Means for packetizing the audio signal; means for transmitting the packet to the network; means for receiving the packet transmitted from the network; means for resolving the received packet to reproduce the audio signal; A voice communication system comprising means for detecting that a packet has not been received, and means for estimating and interpolating a voice signal contained in a missing packet from a voice signal contained in a normally received packet. The voice communication method according to claim 1, wherein linear prediction is recursively repeated using a voice signal included in a voice packet normally received immediately before the missing voice packet to interpolate the missing voice signal. 3. The voice communication system according to claim 2, further comprising a variable gain that corrects a decrease in prediction gain when interpolating a missing voice signal by repeating linear prediction. The voice communication according to claim 2, further comprising means for performing smoothing by replacing a voice signal contained in a voice packet immediately before the loss with a linear sum of the voice signal predicted by repeating linear prediction of the signal of the packet and the received voice signal. method. For the interpolation of the missing voice signal, the voice signal included in the voice packet normally received just before the missing and the voice signal contained in the voice packet normally received just after the missing are recursively repeated to perform the linear prediction, thereby missing the voice signal. The voice communication system according to claim 1 for interpolating. The voice signal contained in the voice packet immediately after the loss is smoothed by replacing the voice signal contained in the voice packet immediately before the loss with the linear sum of the signal of the packet repeated by linear prediction and the received voice signal. 6. A voice communication system according to claim 5, further comprising means for performing smoothing by replacing the signal of the packet with a linear sum of a voice signal predicted by repeating linear prediction and a received voice signal. The type and weight of a linear prediction signal used for missing interpolation and its weight, and the type and weight of a signal used for smoothing a speech signal contained in packets before and after the missing are changed according to the number of consecutive missing packets. The voice communication method described. 6. Means for monitoring the missing rate of voice packets; means for interpolating missing speech using the voice signal immediately before missing according to claim 2; and voice signals immediately before and after missing according to claim 5. A voice communication system that interpolates missing voice signals by adaptively switching the means for interpolating missing voices.

Images