Silence insertion descriptor (sid) frame detection with human auditory perception compensation

TECHNICAL FIELD OF THE INVENTION

0001This invention relates to bandwidth improvements in digitized voice applications when no voice is present. In particular, the invention suggests that improved estimation of background noise during interruptions in speech leads to less bandwidth consumption.

DESCRIPTION OF THE RELATED ART

0002Voice over packet networks (VOPN), require that the voice or audio signal be packetized and then be transmitted. The analog voice signal is first converted to a digital signal and is compressed in the form of a pulse code modulated (PCM) digital stream. As illustrated in Figure 1, the PCM stream is processed by modules of the gateway, such as echo cancellation (EC) 10, voice activity detection (VAD) 12, voice compression (CODEC) 14, protocol configuration 16, etc.

0003Various techniques have been developed to reduce the amount of bandwidth used in the transmission of voice packets. One of these techniques reduces the number of transmitted packets by suspending transmission during periods of silence or when only noise is present. Two algorithms, i.e., the VAD algorithm followed by the Discontinuous Transmission (DTX) algorithm, achieve this process. In a system where these two algorithms exist and are enabled, VAD 12 makes the "voice/no voice" selection as illustrated in Figure 1. Either one of these two choices is the VAD algorithm's output. If voice (active) is detected, a regular voice path is followed in the CODEC 14 and the voice information is compressed into a set of parameters. If no voice (inactive) is detected, the DTX algorithm is invoked and a Silence Insertion Descriptor (SID) packet 18 is transmitted at the beginning of this interval of silence. Aside from the first transmitted SID 18, during this inactive period, DTX analyzes the background noise changes. In case of a spectral change, the encoder sends a SID packet 18. If no change is detected, the encoder sends nothing. Generally, SID packets contain a signature of the background noise information 20 with a minimal number of bits in order to utilize limited network resources. On the receiving side, for each frame, the decoder reconstructs a voice or a noise signal depending on the received information. If the received information contains voice parameters, the decoder reconstructs a voice signal. If the decoder receives no information, it generates noise with noise parameters embedded in the previously received SID packet. This process is called Comfort Noise Generation (CNG). If the decoder is muted during the silent period, there will be sudden drops of the signal energy level, which causes unpleasant conversation. Therefore, CNG is essential to mimic the background noise on the transmitting side. If the decoder receives a new SID packet, it updates its noise parameters for the current and future CNG until the next SID is received.

0004In ITU standard G.729 Annex B, the DTX and CNG algorithms are designed to operate under a variety of levels and characteristics of speech and noise, ensuring bit rate savings and no degradation in the perceived quality of sound. Though the G.729 Annex B SID frame detection algorithm yields smooth background noise during non-active periods, it detects a significant percentage of SID frames even when the background noise is almost stationary. In a real VOPN system, G.729 Annex B generates numerous SID packets continuously, even when the background noise level is very low in dB. One reason for this is that the SID detection algorithm is too sensitive to very low level background noise. Another reason is the effects of imperfect EC. The output signal of EC may have bursts or non-stationary characteristics in low level noise, even when its input noise is stationary.

0005Since SID frames have considerably fewer payload bits than voice packets, generating many SID packets should theoretically not create bandwidth problems. However, both voice and SID packets 22 must have packet headers 24 in VOPN applications (Figure 2.). The header length is the same for voice and SID packets. Sometimes the header 24 occupies most of the bandwidth in a SID packet 22. For instance, in RTP protocol, the header length is 12 bytes. One SID frame contains 2 bytes and a voice frame requires 10 bytes in a G.729 codec. Although SID frame bit rate is 20% of the full bit rate in G.729 codec, when the headers 24 are appended to the packet, the SID packet length with RTP header is about 70% of voice packet length with header. Therefore, it is very important for bandwidth savings to reduce the number of SID packets while preserving sound quality

SUMMARY OF THE INVENTION

0006The SID detection algorithm of G.729 Annex B is based on spectral and energy changes of background noise characteristics after the last transmitted SID frame. The Itakura distance on the linear prediction filters is used to represent the spectral changes. When this measure exceeds a fixed threshold, it indicates a significant change of the spectrum. The energy change is defined as the difference between the quantized energy levels of the residual signal in the current inactive frame and in the last SID frame. The energy difference is significant if it is exceeds 2dB. Since the thresholds of SID detection are fixed and on a crude basis, the generation of an excess number of SID frames is anticipated. Therefore, a SID update delay scheme is used to save bandwidth during non-stationary noise; a minimum spacing of two frames is imposed between the transmission of two consecutive SID frames. This method artificially limits the generation of SID frames.

0007The present invention creates a method to determine if a background noise update is warranted, and is based upon human auditory perception (HAP) factors, instead of an artificial limiter on the excessive SID packets. The acoustic factors, which characterize the unique aspects of HAP, have been known and studied. The applicability of perception, or psycho acoustic modeling, to complex compression algorithms is discussed in IEEE transactions on signal processing, volume 46, No. 4, April 1998; and in the AES papers of Frank Baumgarte, which relate to the applicability of HAP to digitizing audio signals for compressed encoded transmission. Other papers recognize the applicability of HAP to masking techniques for applicability to encoding of audio signals.

0008While some of these works acknowledge the applicability of HAP when compressing high fidelity acoustic files for efficient encoding, they do not recognize the use of HAP in SID detection, (i.e. background noise perceptual change identification, in voice communications). The present invention observes that modeling transitions, based upon HAP, can reduce the encoding of changes in background noise estimation, by eliminating the need to encode changes imperceptible to the HAP system. The present invention does not analyze speech for improved audio compression, but instead searches for characteristics in the perceptual changes of background noise.

0009HAP is often modeled as a nonlinear preprocessing system. It simulates the mechanical and electrical events in the inner ear, and explains not only the level of dependent frequency selectivity, but also the effects of suppression and simultaneous masking. Many factors can affect the perception of sound, including: frequency masking, temporal masking, loudness perception based on tone, and auditory perception differential based upon tone. The factors of HAP can cause masking, which occurs when a factor apart from the background noise renders any change in the background noise imperceptible to the human ear. In a situation where masking occurs, it is not necessary to update background noise, because the changes are not perceptible. The present invention accounts for these factors, by identifying and weighing each factor to determine the appropriate level of SID packet generation, thus increasing SID detection efficiency.

0010The most responsive frequency for human perception, as illustrated in Figure 3, is around 4.5kHz. For sound to be perceptible to the human ear, as the frequency of a signal increases or decreases from 4.5kHz, the sound level must increase in dB. This is illustrated by the threshold in quiet line 26. For example, a sound at 2kHz would have to be 3dB louder to be heard; a sound at 10kHz would have to be 10dB louder, while a sound at a frequency of 0.05 would have to be 47 dB greater. The threshold in quiet line, 26, illustrates the dB level necessary for audible perception.

0011Simultaneous masking, also called frequency masking, is a frequency domain phenomenon where a high level signal (masker) suppresses a low level signal (maskee) when they are in close range of frequency. Figure 3, illustrates a 1KHz pure tone masker and its masking threshold. The masking threshold, below which no signals are audible, depends on the sound pressure level and the frequencies of the masker and of the maskee. In Figure 3, if a tone is generated at 1kHz, it will not only block out any sound at the same frequency, but also blocks signals near 1kHz. The masking threshold depicts the greatest masking near the generated tone, which diminishes rapidly as the sound departs from the detectable tone sound.

0012Temporal masking, including premasking and postmasking, is a time domain phenomenon, which occurs before and after a masking signal. Independent of any of the conditions of the masker, the premasking lasts about 20 ms. However, the postmasking depends on the duration of the masker. In Figure 4, a masking signal is initiated at time 0, and is maintained for 200ms. The background noise is inaudible, by human perception, for the duration of the masking signal. Additionally, masking occurs prior to the signal for approximately 20ms and last 50 to 200ms after the signal.

0013The human ear exhibits different levels of response to various levels of loudness. As sound level increases, sensitivity becomes more uniform with frequency. This behavior is explained in Figure 5. The present invention utilizes this principle as another masking feature.

BRIEF DESCRIPTION OF THE DRAWINGS

0014For a better understanding of the nature of the present invention, reference is made, by way of example only, to the following figures and detailed description, wherein like elements are accorded like reference numerals, and wherein: <ul id="ul0001" list-style="none" compact="compact"><li>Figure 1 is a functional block diagram illustrating the separate processing paths for voice, tone and silence.</li><li>Figure 2 is a diagram illustrating a typical packet.</li><li>Figure 3 is a graph illustrating frequency masking.</li><li>Figure 4 is a graph illustrating temporal masking.</li><li>Figure 5 is a graph illustrating human perception of loudness.</li><li>Figure 6 is a functional flow diagram illustrating the process for identification of background noise estimations for generating SID</li><li>Figure 7 is a graph illustrating HAP related weighting factor determination given various energy levels.</li><li>Figure 8 is a graph illustrating loudness perception thresholds.</li><li>Figure 9 is a Bayes estimator for the selection of SID generation given different thresholds.</li><li>Figure 10 contains graphs simulation results of the HAP-based SID detection and G.729 Annex B SID detection for clean speech.</li></ul>

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

0015The underlying principle of HAP-based SID frame detection is to detect the perceptible background noise change by measuring the HAP-based spectral distance changes as well as the energy level changes between the current frame and the previous SID frame. The present invention defines HAP-based spectral distance (D) as the weighted Line Spectral Frequency (LSF) distance between the current inactive frame and the previous SID frame. The selection of LSF to represent the frequency content of the signal is due to the fact that LSF parameters are available during SID detection for most CELP based codecs. Therefore, a reduction in spectral analysis computation is achieved.

0016The flow diagram of this SID detection algorithm is illustrated in Figure 6. The first step 30 in the beginning of the process is to calculate HAP-based spectral distance thresholds and signal energy levels for each frame by using equations (1), (2) and (3):<maths id="math0001" num=""><img file="EP1229520A2_D0001.tif" /></maths><maths id="math0002" num=""><img file="EP1229520A2_D0002.tif" /></maths><maths id="math0003" num=""><img file="EP1229520A2_D0003.tif" /></maths>

0017The HAP-based spectral distance is defined in equation (1), and Figure 7 shows the selection of weighting factors (w_m(i)) given various energy levels. Weighting factors w_m(i) are the weighting factors used in ITU-T G729 Annex B standard. The weighting factors are derived from Figure 5. For low energy levels, thus low loudness levels, weighting factors increase as the frequency increases to balance the effects of different frequencies. As the loudness level increases, weighting factors become flat. The w_m(i) values in Figure 7 are experimentally selected.

0018The algorithm establishes a set of criteria for the evaluation of signal changes to determine if the signal changes will be perceptible and/or significant to the human auditory response system. One pair in this decision is the HAP spectral distance thresholds based on loudness perception. They are denoted by th_h and th_l and vary depending on the energy of the frame as shown in Figure 8. These figures are also derived by the arguments in Figure 5. It is trivial to see that as the signal energy drops, the loudness drops, too. Thresholds at low loudness levels should be higher to compensate for the low sensitivity. Maximum sensitivity is at high loudness levels, therefore lower thresholds are selected for high loudness levels. The th_l and th_h values in Figure 8 are experimentally selected.

0019These two thresholds are used in the updating process of temporal masking thresholds, th_high and th_low. Equations (3), (4), and (5), represent the HAP spectral distance threshold adaptation based on the temporary masking.<maths id="math0004" num="Equation (4)"><math display="block"><mrow><mtext mathvariant="italic">Thlow</mtext><mtext>(</mtext><mtext mathvariant="italic">n +</mtext><mtext> 1) = α</mtext><mtext mathvariant="italic">Thlow</mtext><mtext>(</mtext><mtext mathvariant="italic">n</mtext><mtext>) + (1 - α) </mtext><mtext mathvariant="italic">Th</mtext><mtext>_l</mtext></mrow></math><img file="EP1229520A2_D0004.tif" /></maths><maths id="math0005" num="Equation (5)"><math display="block"><mrow><mtext mathvariant="italic">Thhigh (n</mtext><mtext> + 1) = </mtext><mtext mathvariant="italic">αThhigh</mtext><mtext> (</mtext><mtext mathvariant="italic">n</mtext><mtext>) </mtext><mtext mathvariant="italic">+</mtext><mtext> (1 - </mtext><mtext mathvariant="italic">α</mtext><mtext>) </mtext><mtext mathvariant="italic">Th</mtext><mtext>_</mtext><mtext mathvariant="italic">h</mtext></mrow></math><img file="EP1229520A2_D0005.tif" /></maths>

0020Since the post masking is in the order of 50 to 200 ms, the time constant of above thresholds are chosen as 50 ms, i.e. a=3/4 in current implementation. Th_high 50 and Th_low 52 are used in Bayes classifier as illustrated in Figure 9.

0021Figure 6 further illustrates that if the HAP-based spectral distance 30 is greater than the higher threshold th_high 36, a SID frame is detected 38. The average LSF energy is then reset 40 and is updated based on loudness perception 32 and temporary masking 34. If the distance 30 is less than the lower threshold th_low 42, the current frame is considered as a non-SID frame. If the spectral distance falls between th_high and th_low, then the quantized energy feature q 46 is introduced to decide if the current frame is a SID. If E_q > 2dB, then a SID packet 38 is detected. If E_q < 2dB, then average LSF noise spectrum is updated 44 prior to returning to re-calculating HAP spectral distance thresholds 32 and adjusting the thresholds 34.

0022The present invention is then able to reject those transitions which represent inaudible background level changes and is able to generate SID packets 38 corresponding to the perceptible changes in background noise. Figure 10 illustrates simulation results of the HAP-based SID detection and G.729 Annex B SID detection for clean speech, with/without various added noise (babble, office or street noise) under different background noise levels. PAMS is used for objective measurements. The new algorithm either performs the same or outperforms the standard G729 Annex B SID detection algorithm in terms of YLQ in noisy conditions (Rows 7 through 15) with a significant SID percentage reduction. In other examples (Rows 1-6), although the new algorithm cannot perform the same quality as the standard SID detection algorithm, the SID reduction ratio is still significant and the YLQ difference is in a negligible range. Subjective tests also proved that there was no or insubstantial degradation in the quality. <tables id="tabl0001" num="0001"><img file="EP1229520A2_D0006.tif" /></tables>

0023Because many varying and different embodiments may be made within the scope of the inventive concept herein taught, and because many modifications may be made in the embodiments herein detailed in accordance with the descriptive requirements of the law, it is to be understood that the details herein are to be interpreted as illustrative and not in a limiting sense.