JPS5947903B2 - Digital call audio introduction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a digital call voice insertion method, and is capable of suitably absorbing freeze-outs that occur during voice insertion.

With the advancement and economicization of digital signal processing technology, when transmitting multiple digitized audio signals, it is necessary to detect the audio portion of each audio signal where audio portions exist, and transmit only the detected audio portions. Digital call voice insertion method (
DigitalSpeechInterpola-ti
on (abbreviated as DSI) has been adopted.

In this case, the audio signal is divided into a predetermined number of unit blocks that are a series of sample values, and a sound detector determines whether or not audio exists within the unit block for each unit block. Transmit existing unit blocks. In addition, in order to compress the bandwidth, a predictor predicts the current sample value from a group of digitized past sample values, and a subtracter calculates the difference (prediction error) between this predicted value and the current sample value. Later, there is a predictive coding method in which only this prediction error is quantized and coded again using a quantizer and transmitted at a low bit rate. This predictive coding method uses the delta modulation method, which encodes with l bits, and the DPCM (Differential modulation method, which encodes with 2 or more bits).
D
The PCM method also uses adaptive DPCM (ADP
CM) method. Therefore, the DSI method and the DPC
A high-efficiency DSI method can be considered that uses the M method (or ADPCM method) in combination. This high-efficiency DSI method makes it possible to make dramatic use of the transmission path by transmitting only the audio part, thereby effectively utilizing the transmission path, and by transmitting this audio part at an even lower bit rate to compress the bandwidth. It's summery. In other words, the DSI gain (B:) is determined from the ratio of the detected audio portion to the entire audio signal (average operation rate), and the original number of encoded bits of the audio signal is If the predictive coding gain is the reciprocal of the reduction rate of the number of predictive coding bits for
This is because the overall gain of the SI method is given by the product of the DSI gain and the predictive coding gain. By the way, since the average operation rate of voice calls is about 40%, theoretically a DSI gain of about 2.5 times should be obtained.

When powering up or inserting audio, the number of input trunks with audio in the DSI input trunks, that is, the number of active input trunks, exceeds the number of DSI output channels, and some active input trunks cannot be connected. figure,
There are cases where the audio is cut off midway through. This state is called freeze-out. Therefore, in order to avoid this freeze-out, conventionally the DSI gain is suppressed to about twice. On the other hand, regarding the predictive coding gain, the voice quality (signal to quantization noise ratio, ▼) is 8 to 6 bits P.
Assuming that it corresponds to a CM, 4-bit predictive coding is used. As a result of this Kb, the total gain is about 4 times that of 64τPCM of 8 bits and 8 KHz sampling.

As mentioned above, even in the high-efficiency DSI system, the following two points can be cited as reasons why the conventional total gain is not very high at about 4. In other words, in order to avoid frequent if-freeze-outs,
NTO-
The number of bits (for example, 4 bits) required for the low part is applied as is to the high part of 1Nq, so there is still redundancy in predictive coding.

SUMMARY OF THE INVENTION In view of the above-mentioned drawbacks of the prior art, it is an object of the present invention to provide a digital call voice insertion system with a significantly high overall gain. The configuration of the present invention to achieve such an object divides a plurality of digitized audio signals into unit blocks, each of which is a predetermined number of series of sample values, and predictively encodes only the unit blocks in which audio exists. In the digital voice insertion method for transmission using variable bits, the predictive coding is performed using a variable bit type . At the same time, the characteristics of the voice within the unit block in which the voice exists is determined, and if freeze-out occurs when inserting voice, prediction is performed in order from the unit block where the signal to quantization noise ratio is considered to be good. It is characterized by reducing the number of encoding bits. do. Hereinafter, the present invention will be explained in detail based on the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention. In this embodiment, the input audio signal is 8-bit, 8-kHz sampling PCM, and the unit program is a series of 32 sample values. . In FIG. 1, reference numeral 1 denotes an expander which in this case converts an 8-bit non-linear PCM signal NL into a 13-bit linear PCM signal L. Therefore, if the audio signal is originally a linear PCM, this expander 1 is unnecessary.
2 is a buffer memory in which the audio signal L of each trunk is stored for each unit block.

3 is a voice detector, 4 is a data signaling detector, 5
is a variable bit type ADPCM encoder for audio signals, 6 is an ADPCM encoder for data, 7 is an allocation controller, 8 is a multiplexer, 9 is an allocation status signal generator, and 11 is a coded bit length display signal generator. It is a vessel.

That is, in this embodiment, data and signaling in the audio band signal are separated from the audio signal, removed from the DSI channel CH, encoded by a dedicated ADPCM encoder 6, and transmitted, and only the audio signal is placed on the DSI channel CH. It's becoming like that. However, the same ADPCM may be operated without separating the ADPCMs for voice and data. Here, the portion related to the audio signal is extracted and shown in FIG. In FIG. 2, the variable bit type ADPCM encoder 5 is an ADPC
It consists of an M encoder 19 and a bit length controller 10. 12 to 16 constitute an ADPCM encoder 19, 12 is a subtracter, 13 is an adaptive quantizer, 1
4 is an adaptive predictor, 15 is an inverse quantizer, and 16 is an adder. Note that 17 is an input terminal for the audio signal L converted into a linear PCM signal, and 18 is an output terminal for an output signal predictively encoded after audio insertion according to the method of the present invention. Incidentally, although FIG. 2 is diagrammed as one independent system, as can be understood by comparing it with FIG. While the variable bit type ADPCM encoders 5 can be shared by the input trunk TK, the same number of variable bit type ADPCM encoders 5 as the number of DSI output channels CH on the output side of the buffer memory 2 are prepared. The voice detector 3 detects for each unit block whether or not voice exists in the unit block, and when voice exists, calculates the signal to quantization noise ratio (
? It predicts the quality of the product (Wai).

Then, the unit block determined by the voice detector 3 as having voice is read out from the buffer memory 2 by a control signal from the allocation controller 7 and encoded into variable bit type ADPCM of the allocated channel CH. It is output to the device 5. At the same time, the voice detector S unit 3 outputs the quality information of one of the unit blocks to the Nq bit length controller 10.
When a freeze-out occurs, the bit length controller 10 instructs the quantizer 13 about the unit block whose predictive coding bit length is to be reduced and its bit length.

At this time, the unit blocks whose bits are to be reduced are designated in descending order of the good bits. The bit length information of the unit block is outputted from the encoded bit length display signal generator 11. Also,
According to the control signal from the allocation controller 7, an allocation status signal for associating the input trunk TK with the output channel CH is outputted from the allocation status signal generator 9. The voice detector 3 includes a voice detection section that detects whether or not voice exists in a unit block by using the internal power and zero-crossing number in the unit block, and a voice detector that detects the nature of the voice in the unit block, that is, the signal. and a voice quality determination unit that determines whether the quantization-to-noise ratio is good or bad.

By the way, regarding the signal-to-quantization noise ratio (L) of speech, in general, parts containing many low-frequency components, that is, voiced sounds, are easy to predict, so the prediction error is small.
is said to be good, but it is difficult to predict parts that contain many high-frequency components or parts where the frequency spectrum is flat, that is, unvoiced parts, or transient parts between voiced and unvoiced sounds, and the prediction error is large and - is bad. I can say that it will be. Therefore, it can be considered that - is good for voiced sounds, which are mostly vowels, etc., and that - is bad for unvoiced sounds such as fricatives and plosives, and the transient parts before and after them. From the above, in the present invention, instead of detecting the signal to quantization noise ratio (L), it is determined whether the sound is voiced or unvoiced, and if it is a voiced sound, 1 is good, and if it is unvoiced, - is bad. It is determined. For this reason, the speech quality determining section may focus on the number of zero crossings and determine that a unit block with a particularly large number of zero crossings has poor speech, or may use a predictor set to a low frequency range to detect frequent mispredictions. Although a unit block for determining the quality of audio may be used, in this embodiment, the unit block shown in FIG. 3, which focuses on internal power and frequency components, is used. In FIG. 3, 20 is a power calculator, 21 is a zero-crossing counter, 22 is a threshold selector, 23 is a polarity bit matrix processor, 24 is a judger, 25 is a voice detection display terminal, and 26 is a voiced/unvoiced sound. This is a display terminal.

The power calculator 20 calculates the internal power St of each unit block. This internal power St is input to the determiner 24 and compared with a reference value, for example STh, so that St<
When STh, the audio of unit block t is silent (Sil
ence). When STh≦St, it is immediately determined that there is a voice, and a voice detection signal is output from the terminal 25. Furthermore, we pay attention to the main polarity series, which is the series of polarities of each sample value of unit block t itself, and the subpolarity series, which is the series of polarities of each sample value after the unit block t is filtered, and calculate their periodicity. It is determined whether the center is a voiced sound or an unvoiced sound, and a voiced/unvoiced sound display signal is output from the terminal 26. For this reason, the polarity bit matrix processor 23 calculates the similarity of the main and sub-polarity sequences to a reference polarity sequence group, which is a polarity sequence of each sample value of each of the digitized sine waves and cosine waves of a plurality of frequencies. calculate. The reference polarity sequence group is a polarity bit sequence matrix H consisting of ±1 elements given by the following equations (1) and (2). - On the other hand, if the characteristics of the digital filter are specified by l (=l~5), the main and sub polarity series can be determined by Z'
It is written as (t).

Therefore, the main polarity series is Z2(t) and Z1(t
) is the low-passed sub-polarity series, Z3(t) to Z5(t
) are high-passed (pre-emphasized) secondary polarity sequences.

Therefore, if the pattern matching degree Y between the main and sub polarity series Z8(t) and the reference polarity series H is determined by the following equations (3) and (4), the polarity pattern matching power P5k corresponding to the Fourier transform is (t) is given by the following equation (5). This polar pattern matching power Plk(t) is nothing but the similarity. However, m=32 and k=r1 to R2. To summarize the above, polarity bit matrix processor 2
3 performs a matrix operation on the main/secondary polarity series Z1(t) and the reference polarity series group H, and outputs the similarity Plk(t).

This degree of similarity P5k(t) indicates that the original and filtered unit blocks t, denoted by l, are 125Hz and its harmonics, denoted by k. This shows how similar each combination of l and k is. The degree of similarity P5k(t) thus calculated by the polarity bit matrix processor 23 is input to the determiner 24, where it is compared in magnitude with the reference value Plh.

This base. The quasi-value PTh is set using, for example, the filter characteristic 1, the frequency k, and the zero crossing numbers Zct, zcpt as parameters. The zero crossing number Zct is the original unit block t(7) zero crossing number, and the zero crossing number Zcpt is, for example, in the digital filter shown in FIG. 4, when a1=1, A2
This is the number of zero crossings of the unit block t after pre-emphasis when A3=-0.5 and A3=o. In addition, in Figure 4,
D is a delay device, and a1 to A3 are coefficients. These zero crossing numbers Zct and zcpt are counted by the zero crossing number counter 21 for each unit block t, and are inputted one after another to the threshold selector 22, and from the threshold selector 22, the zero crossing number Zct is counted.
, zcpt, is output to the determiner 24. The settings of the reference value PTh, as shown in Table 1 with k and i, are as follows: Z1(t) is set to the low range, Z2(t) is set to the mid-low range, Z3(t) is set to the mid range, and Z4 is set to the middle range. (t) for mid-high range,
Further, Z5(t) is determined based on the fact that each filter 1 is specified so as to be easily similar to a high frequency.

In this case, if STh≦St, the determiner 24 determines the nature of the voice of the unit proc t based on criteria such as A or Mouth, which will be described later. iPlk(t)≧PT
If hl or P2k≧PTh2 holds true for at least one k, it is considered a voiced sound, and the other sounds are considered unvoiced. ClP3k(t)≧PTh3, P(t)≧PTh
4 and P5k(t)≧PTh5 holds for at least one k or Plk(t)
≧PThl, P2k(t)≧PTh2, P3k(t)≧
PTh3, P(t)≧Plh4, P5k(t)≧PTh
If none of the above conditions hold true, the unit block t is considered unvoiced, and the rest are voiced. Therefore, the allocation controller 7 detects the occurrence of a freeze-out, and when a freeze-out occurs, the unit blocks of unvoiced sounds are predictively encoded using the basic bit length, while the unit blocks of voiced sounds are sequentially encoded in the basic bit length. Predictive encoding is performed using a bit length shorter than the bit length.

As a result of the above, even if a freeze-out occurs, a new channel can be secured by gathering the reduced number of bits, and the freeze-out can be avoided.

In this case, since the unit block with a good signal to quantization noise ratio (?) is subjected to bit reduction, the overall voice quality is hardly affected. Here, the relationship between DSI gain and freeze-out will be explained. Average operating rate is 35-38
%, the relationship between DSI gain and freeze-out occurrence rate is as follows: If the DSI gain is 3 times the basic bit length (4 bits), 240 trunks will have a channel shortage of about 10%, and the average operating rate will be 35 to 38%. For a DSI gain of the reciprocal of 240 trunks, there will be a channel deficit of about 5%. On the other hand, if all the bits are reduced to 3 bits, the freeze-out will be approximately 5.5 x 10-2% for an average operation rate of 38%, and less than 3 x 10-3% for an average operation rate of 35% ( channel shortage), and even if the DSI gain is set to 3 times, most of the freeze-outs can be absorbed. As specifically explained above in conjunction with the embodiments, according to the present invention, the predictive encoder is of variable bit type, and the bit length is reduced from the unit block that is considered to have a good signal to quantization noise ratio. ,
Freeze-outs can be avoided with little impact on audio quality,
The overall gain of high-efficiency DSI can be greatly improved.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is a block diagram showing its essential parts, Fig. 3 is a block diagram further showing its essential parts, and Fig. 4 shows an example of a digital filter. It is a block diagram. In the drawing, 2 is a buffer memory, 3 is a voice detector, 5 is a variable bit type ADPCM encoder, 7 is an allocation controller, 9 is an allocation status signal generator, 10 is a bit length controller, and 11 is a bit length display signal. generator, 19 is an ADPCM encoder, 2
0 is a power calculator, 21 is a zero-crossing counter, 22 is a threshold selector, 23 is a polarity bit matrix processor, 2
4 is a determiner.

Claims (1)

[Claims] 1. In a digital call voice insertion method that divides a plurality of digitized voice signals into unit blocks each of which is a predetermined number of series of sample value groups, and predictively encodes and transmits only the unit blocks in which voice exists, the prediction Encoding is performed using a variable bit type predictive encoder, and the characteristics of the voice within the unit block in which the voice exists is determined, and if freeze-out occurs during voice insertion, the signal to quantization noise ratio is good. A digital call voice insertion method characterized by reducing the number of predictive coding bits in order from unit blocks that can be regarded as .