Method and apparatus for selecting an encoding rate in a variable rate vocoder

ES 2 240 602 T5

DESCRIPTION

Method and apparatus for selecting an encoding rate in a variable rate vocoder.

Background of the invention

I. Field of the invention

The present invention relates to vocoders. More particularly, the present invention relates to a new and improved method for adding lock frames.

II. Description of related techniques

Variable rate speech compression systems often use some form of rate determination algorithm before encoding begins. The rate determination algorithm assigns a higher bit rate coding system to the segments of the audio signal in which speech is present and a lower rate coding system to the silence segments. In this way, a lower average bit rate is obtained, while the quality of the reconstructed speech remains high. Therefore, to effectively use a variable rate speech coder requires a robust rate determination algorithm that can differentiate between speech and silence in a variety of background noise environments.

One such variable rate voice compression system or variable rate vocoders is disclosed in pending US Patent No. 5,414,796 entitled "Variable Rate Vocoder" and assigned to the assignee of the present invention. In this particular implementation of the variable rate vocoder, the input speech is encoded using Code Excited Linear Prediction (CELP) techniques at one of several rates determined by the level of speech activity. The level of oral activity is determined from the energy of the input audio samples, which may contain background noise in addition to speech. In order for the vocoder to be able to perform high quality speech coding with varying levels of background noise, it is necessary to employ a threshold adjustment technique to compensate for the effect of background noise on the rate decision algorithm.

Vocoders are often used in communication devices, such as cell phones or personal communication devices, to effect digital compression of an analog audio signal, which is converted into a digital signal for transmission. In a mobile environment where a cell phone or personal communication device can be used, high energy levels from background noise make it difficult to differentiate by the velocity determination algorithm between low-energy muffled sounds and quiet silence. background noise, using a speed determination algorithm based on signal energy. Therefore, voiceless sounds are often encoded at lower bit rates, which leads to a degradation in voice quality, since consonants such as "s", "x", "ch", "sh", "T", etc. they are lost in the reconstructed voice.

Vocoders that base their rate decisions solely on the energy of the background noise fail in their attempt to account for the signal strength relative to the background noise when setting threshold values. Vocoders that base their threshold levels solely on background noise tend to compress the threshold levels together as background noise increases. If you want the signal level to remain fixed, this is the correct way to set the threshold levels; however, if you want the signal level to rise with the background noise level, compression of the threshold levels is not the best solution. Therefore, there is a need for an alternative procedure to establish threshold levels in variable rate vocoders, which takes into account the signal strength.

Finally, the problem related to music reproduction by speed decision vocoders based on the energy of the background noise remains to be solved. When speaking, people must pause to breathe, allowing threshold levels to reset to the appropriate background noise level. However, when streaming music through a vocoder (as is the case with music on hold), there is no pause and the threshold levels continue to increase until the music begins to be encoded at a slower than speed. full speed, and in such circumstances the variable speed encoder provides confusing music with background noise.

The document K. Srinivasan and A. Gersho: “Voice activity detection for cellular networks”, Proceedings: IEEE Workshop on speech coding for telecommunications, October 13-15, 1993, pages 8586, XP002204645, University of California, deserves special attention. The document refers to algorithms for the detection of voice activity in the presence of vehicular noise and mixing noise. In particular, it discloses an algorithm for detecting voice activity in which an adaptive blocking period of between 40 ms and 180 ms is introduced. The actual blocking period is based on the relationship, r, between the output capacity of the noise suppression filter and the corresponding adaptive threshold.

ES 2 240 602 T5

The document Paksoy E et al: “Variable rate speech coding for multiple access wireless networks”, Electrotechnical Conference, 1994, Proceedings., 7 deserves special attention.^th Mediterranean Antalya, Turkey April 12-14, 1994, New York, NY, USA, IEEE, April 12, 1994, pages 47-50, XP10130866 ISBN: 0-7803-1772-6 referring to voice encoding variable speed for multiple access wireless networks, which particularly mentions a voice activity detection with an adaptation of the blocking period to the detected signal levels. The document WO-A1-93 / 13516, which discloses the calculation of the VAD blocking period using an SNR, and the GSM Recommendation 06.32, "Voice activity detection", February 1992, which discloses the addition VAD blocking to speech bursts that exceed a specified duration.

According to the present invention, there is provided a method and apparatus for adding blocking frames to a plurality of frames encoded by a vocoder, as set forth in claims 1 and 3.

Summary of the invention

The present description describes a new and improved method and apparatus for determining a coding rate in a variable rate vocoder. The first objective is to provide a method by which the probability of encoding low-energy muffled sounds as background noise can be reduced. The input signal is filtered to obtain a high-frequency component and a low-frequency component. The filtered components of the input signal are then separately analyzed for the presence of speech. Because voiced sounds have a high-frequency component, it is easier to differentiate their intensity from background noise in a high-frequency band than in the entire frequency band.

The second objective is to provide a means of setting threshold levels that take into account both the energy of the signal and the energy of the background noise. The establishment of speech detection thresholds is based on an estimate of the signal-to-noise ratio (SNR) of the input signal. In the example, the signal energy is estimated as the maximum signal energy during periods of speaking activity and the background noise energy is estimated as the minimum signal energy during periods of silence.

The third objective is to provide a method for encoding music using a variable rate vocoder. In the example, the rate selection apparatus detects a group of consecutive frames during which the threshold levels have increased and checks the periodicity relative to said group of frames. The presence of music is indicated by the periodic character of the input signal. If the presence of music is detected, thresholds are set at levels that allow encoding of the signal at full speed.

Brief description of the drawings

The features, objects, and advantages of the present invention will become more apparent from the detailed description provided below and illustrated by the drawings, in which the same reference numerals are used for equivalent parts, and in which:

Figure 1 is a block diagram of the present invention.

Detailed description of the preferred embodiments

Referring to Figure 1, the input signal, S (n), is provided to the subband energy calculators 4 and 6. The input signal S (n) consists of an audio signal and background noise . The audio signal is usually voice, but it can also be music. In the exemplary embodiment, S (n) is provided in twenty millisecond frames of 160 samples each. In the exemplary embodiment, the input signal S (n) has frequency components between 0 kHz and 4 kHz, which is approximately the bandwidth of a human speech signal.

In the exemplary embodiment, the 4 kHz input signal, S (n), is filtered to obtain two separate subbands. The two separate subbands are between 0 and 2 kHz and 2 kHz and 4 kHz, respectively. In an exemplary embodiment, the input signal can be divided into subbands by subband filters, the design of which is widely known in the art and described in US Patent No. 5,644,596, entitled "Frequency Selective Adaptive Filtering." , and granted to the assignee of the present invention.

The impulsive responses of the subband filters are indicated by h_L(n) and h_H(n) for the band pass filter and the high pass filter, respectively. To calculate the energy of the resulting subband components of the signal and obtain the RL (0) and RH (0) values, simply add the squares of the subband filter output samples, in a way well known in the art. of technique.

ES 2 240 602 T5

In a preferred embodiment, when the input signal S (n) is provided to the subband energy calculator 4, the energy value of the low-frequency component of the input frame, R_L(0), is calculated as:

<img file="ES2240602T5_D0001.tif" />

where L is the number of taps of the band pass filter with impulse response h_L(n) and R_S (i) is the autocorrelation function of the input signal, S (n), provided by the equation:

for ie [0, L-1] (2) where N is the number of samples in the frame and R_hL is the autocorrelation function of the band pass filter h_L(n) provided by:

L-1 * h_L(») = XhLÍnJ-hLCn-i) 'n = O = 0 for ie [0, L-1] (3) in the other cases

The high-frequency energy, R_H(0), is calculated similarly in subband energy calculation item 6.

The values ​​of the autocorrelation function of the subband filters can be calculated in advance to reduce the number of calculations. Also, some of the calculated values ​​of R_S(i) are used in other calculations of the encoding of the input signal, S (n), which further reduces the net calculation burden of the encoding rate selection method of the present invention. For example, obtaining the tap values ​​of the LPC filter requires the calculation of a group of autocorrelation coefficients of the input signal.

The calculation of the LPC filter tap values ​​is well known in the art and is described in detail in the aforementioned US Patent No. 5,414,796. If speech is to be encoded with a procedure that requires a ten-tap LPC filter, only the R values ​​need to be calculated_S(i) for the values ​​of i between 11 and L-1, in addition to those used in the signal coding, because in the calculation of the tap values ​​of the LPC filter, R is used_S(i) for values ​​of i between 0 and 10. In the exemplary embodiment, the subband filters have 17 taps (ie, L = 17).

Subband energy calculation item 4 provides the calculated value of R_L (0) to the subband speed decision element 12 and the subband energy calculation element 6 provides the calculated value of R_H(0) to the subband speed decision element 14. The speed decision element 12 compares the value of R_l(0) with two default threshold values ​​T_{L1 / 2} and T_Lfull and assigns a recommended encoding speed, RATEL, based on the result of the comparison. The speed assignment is carried out as follows:

<td>RATE_l = eighth speed</td><td>Rl (0) s T_L1/₂</td><td> (4)</td>

<td>RATE_l = medium speed</td><td>Tli / 2 &lt;Rl (0) - T_LF_orll</td><td> (5)</td>

<td>RATEl = full speed</td><td>Rl (0)&gt; TLfull</td><td> (6)</td>

The subband rate decision element 14 works in a similar way and selects a recommended encoding rate, RATEH, based on the high frequency energy value RH (0) and based on a different set of threshold values ​​T_{H1 / 2} and T_Hfull. The subband speed decision element 12 provides the speed

ES 2 240 602 T5 encoding recommended, RATE_L, to the encoding rate selection element 16, and the subband rate decision element 14 provides the recommended encoding rate, RATE_H, to the encoding rate selection element 16. In the exemplary embodiment, the encoding rate selection element 16 selects the higher of the two recommended rates and provides the higher rate as the selected encoding rate.

The subband energy calculation element 4 also provides the low frequency energy value, RL (0), to the threshold matching element 8, where the threshold values ​​T are calculated_{LJ / 2} and T_Lfull for the next input frame. Similarly, the subband energy calculator 6 provides the high frequency energy value, R_H(0), to the threshold adaptation element 10, where the threshold values ​​T are calculated_{HJ / 2} and T_Hfull of the next input frame.

The threshold matching element 8 receives the low-frequency energy value, R_L(0), and determines whether S (n) contains background noise or an audio signal. In an exemplary embodiment, the procedure by which the threshold adaptation element 8 determines whether or not an audio signal is present consists in examining the normalized autocorrelation function NACF, which is given by the following equation:

<img file="ES2240602T5_D0002.tif" />

where e (n) is the formant residual signal obtained after filtering the input signal, S (n), by means of an LPC filter.

The design and filtering of a signal by an LPC filter are well known in the art and are described in detail in the aforementioned US Patent No. 5,414,796. The input signal, S (n), is filtered by the LPC filter to eliminate the interaction of the formants. The NACF function is compared with a threshold value to determine the presence or absence of an audio signal. If the NACF function is higher than a predetermined threshold value, it means that the input frame has a periodic characteristic that indicates the presence of an audio signal (voice or music). It should be noted that, although the voice and music parts are not periodic and have low NACF function values, the background noise usually does not present any periodicity and almost always has low NACF function values.

If S (n) is determined to contain background noise, the value of the NACF function is less than the threshold value TH1 and then the R value is used._L(0) to update the current BGN background noise estimate value_L. In the exemplary embodiment, TH1 is 0.35. R_L(0) is compared to the current value of the BGN background noise estimate_l. If R_L(0) is less than BGN_L, then the BGN background noise estimate_L is set to R_L(0), regardless of the value of the NACF function.

The BGNL background noise estimate only increases when the NACF function is less than the threshold value TH1. If R_l(0) is greater than BGN_L and the NACF function is less than TH1, then the background noise energy BGN_L is set to a_j. BGN_L, being a_j a number greater than 1. In the exemplary embodiment, a_j equals 1.03. BGNL continues to increase, while the NACF function is lower than the threshold value TH1 and RL (0) is higher than the current value of BGN_L, until BGN_L reaches a predetermined maximum value BGN_max, at which point the BGN background noise estimate_L is set to BGN_max.

If an audio signal is detected (which happens when the value of the NACF function exceeds a second threshold value TH2), the signal energy estimate, SL, is updated. In the exemplary embodiment, TH2 is set to 0.5. The value of RL (0) is compared to the current estimate of the bandpass signal energy, SL. If RL (0) is greater than the current value of SL, then SL is set to RL (0). If RL (0) is less than the current value of S_L, then S_L is set to a₂.S_L, also this time only if the NACF function is higher than TH2. In the exemplary embodiment, a₂ is set to 0.96.

Next, the threshold adaptation element 8 performs the estimation of the signal-to-noise ratio according to the following equation 8:

<img file="ES2240602T5_D0003.tif" />

(8)

ES 2 240 602 T5

Then, the threshold matching element 8 determines the index of the quantized signal-to-noise ratio I_Snrl, according to equations 9 to 12 below:

<img file="ES2240602T5_D0004.tif" />

for 20 &lt;SNRL &lt;55, (9) = 0, for SNRL &lt;20, (10) = 7 for SNRL&gt; 55 where nint is a function that rounds the fractional value to the nearest integer.

The threshold matching element 8 then selects or calculates two scale factors, k_{L1 / 2} and k_Lfull, according to the signal-to-noise ratio index, I_Snrl. Table 1 below is an example of a scale values ​​lookup table:

TABLE 1

<img file="ES2240602T5_D0005.tif" />

These two values ​​are used to calculate the threshold values ​​for speed selection based on the following equations:

Tli / 2- Kli / 2 BGNL, and (11)

TLfull⁼ KLfull 'BGNL, (12) where TL1 / 2 is the threshold value for medium speed and low frequency and TLfull is the threshold value for full speed and low frequency.

The threshold adaptation element 8 provides the adapted threshold values ​​TL1 / 2 and TLfull to the speed decision element 12. The threshold adaptation element 10 works in a similar way and provides the threshold values ​​T_{H1 / 2} and T_Hfull to the subband speed decision element 14.

The initial value of the estimate of the energy of the audio signal S (which can be S_L or S_H) is set as follows. The initial signal energy estimate, S_INIT, is set to -18.0 dBm0, where 3.17 dBm0 is the signal intensity of a full sine wave which, in the example embodiment, is a digital sine wave with an amplitude range between -8031 and 8031 .S_INIT it is used while the presence of any acoustic signal is not determined.

The method by which an acoustic signal is initially detected is to compare the NACF function value with a threshold. When the NACF function exceeds the threshold for a predetermined number of consecutive frames, then the presence of an acoustic signal is determined. In the example embodiment,

ES 2 240 602 T5 the NACF function must exceed the threshold for ten consecutive frames. When this condition is satisfied, the signal energy estimate, S, is set to the maximum signal energy in the preceding ten frames.

The initial value of the BGN background noise estimate_L is initially set to BGN_max. As soon as a subband frame energy is received that is less than BGN_max, the background noise estimate is reset to the value of the received subband energy level, and the background noise estimate BGNL is generated in the manner described above.

According to the invention, a blocking state is entered when a low rate frame is detected after a series of full rate speech frames. In the exemplary embodiment, when four consecutive speech frames are encoded at full rate that are followed by a frame in which the encoding rate is set lower than full rate and the calculated signal-to-noise ratios are lower than a predetermined minimum SNR, the encoding rate for that frame is set to full rate. In the exemplary embodiment, the default minimum SNR is 27.5 dB as defined in equation 8.

In the preferred embodiment, the number of frames in the blocking period is a function of the signal-to-noise ratio. In the exemplary embodiment, the number of frames in the lock period is determined as follows:

no. of blocking period frames = 1 no. of blocking period frames = 2 no. of blocking period frames = 0

22.5 &lt;SNR &lt;27.5,

SNR &lt;22.5,

SNR&gt; 27.5, (13) (14) (15)

The present description also provides a method by which the presence of music is detected and which, as described above, lacks the pauses that allow the background noise measurements to be reset. The procedure for detecting the presence of music assumes that there is no music at the beginning of the call. This enables the encoding rate selection apparatus to correctly estimate the initial background noise energy, BGNinit. Since music, unlike background noise, has a periodic characteristic, the value of the NACF function is examined to differentiate music from background noise. The music detection procedure calculates an average NACF function based on the following equation:

<img file="ES2240602T5_D0006.tif" />

(16) in which the NACF function is the one defined in equation 7, and

T is the number of consecutive frames in which the estimated value of the background noise has been increasing from the initial background noise estimate BGN_init.

If the BGN background noise has been increasing for a number of consecutive T-frames and the NACF function_birdexceeds a predetermined threshold, then music is detected and the BGN background noise is reset to BGN_init. It should be noted that, to be effective, the T value must be low enough that the encoding rate does not drop below full rate. Therefore, the value of T must be established as a function of the acoustic signal and of BGN_init.

The preceding description of the preferred embodiments is provided to enable all those skilled in the art to create or use the present invention. The various possible modifications of these embodiments will be apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without the need for inventive step. Therefore, the present invention should not be limited to the described embodiments, but its scope will be determined by the appended claims.