Chapter 4 Proposed PNS Approaches
4.1 Signal Substitution before Quantization Process
4.1.2 High Frequency Signal Substitution
The audio coding usually has the tradeoff between retaining all the frequency range and cutting off high frequency signals. As shown in Figure 13, a generic audio codec usually cuts off the signals above 16k Hz to maintain a better precision in low frequency range at 128kbps. This mechanism results in a less degree of loudness and clearness of the coded signals. However, in another example illustrated in Figure 13, the sparse bands will appear in high frequency range and the precision of coded signals in low frequency range will be lower if retaining all frequency range. This thesis shows that the tradeoff between retaining all frequency and cutting off high frequency bands needs to be reconsidered within perceptual noise substitution module.
As illustrated in Figure 13, the signals in high frequency bands can be replaced by PNS information to protect better precision on the signals in low frequency range.
Since the purpose of PNS is to substitute noise-like components, the bands with high tonality must be avoided during the high frequency signal replacement process to prevent great distortions in substituted bands. Furthermore, a limitation on the bit usage for PNS to replace high frequency signals must be conducted at low bit rates. At low bit rates, the PNS should not spend too many bits on retaining the high frequency signals and cause degradation in low frequency range since the bit rate is already low.
In this section, the design issues of proposed high frequency signals substitution approach will be brought up and discussed in detail.
Figure 13: Example of high frequency signal substitution by PNS
4.1.2.1 Start Frequency Decision and Tonal Band Avoidance
Retaining high frequency signals by means of perceptual noise substitution can reserve the loudness of the original signals, avoid sparse bands in high frequency and gain a better precision on the signals in low frequency. The start frequency of high frequency signal substitution is according to the location that sparse band appears and the appearance of the sparse band is mainly based on the encoding bit rate. As illustrated in Figure 13, the suitable start frequency of high frequency signal substitution in the example is about 16k Hz since the sparse bands appeared in the range of 16k~20k Hz if retaining all frequency range at 128kbps. Table 2 shows the suitable start frequency proposed in this thesis at different bit rates. The suggested start frequency is conducted by the objective quality measurement in MPEG 12 tracks at different bit rates.
Table 2: Suggested start frequency of high frequency substitution by PNS
Bit rate (kbps) 128 112 96 80 64
Start frequency (Hz) 17875 16500 15125 13750 12375 In MPEG-4 Advanced Audio Coding (AAC), the bit rate can be arbitrarily specified. For the bit rate which is not belonged to any entry in Table 2, the start frequency can be determined by interpolation of the nearby entries in the table.
Formula (6) is a mathematical expression of the method.
1
where B is the value of specified bit rate, Fstart
(B) is the calculated start frequency at
bit rate B, Bi and Bi+1 are the nearby bit rates found in Table 2, Si and Si+1 are the corresponding start frequencies of bit rate Bi and Bi+1.Figure 14 shows the signals that contain highly tonal components and the reconstructed signals that the high frequency signals are substituted by PNS parameters. Obviously, the tonal signals in high frequency are disappeared in the reconstructed signals. The human perception of a highly tonal band is mainly stimulated by the high tone, not the other signals. If enabling PNS module in a highly tonal band, the distortion between the original signals and reconstructed random noise can be perceived obviously even the substituted energy was adjusted to the noise floor.
Therefore, for the signals which contain plenty of tonal components, the PNS module has to avoid being active in such signals. This thesis proposes a method to detect highly tonal bands and avoid activating PNS in these bands based on the spectral
flatness measured in equation (3).
Figure 14: Example of tonal band avoidance in high frequency signal substitution
4.1.2.2 Limitation on Bit Usage
Retaining high frequency signals by perceptual noise substitution dose not always save bits. It depends on the bit usage of the band that PNS substituted. If the band that PNS substituted was allocated some bits, then PNS can save bits in this band since the cost of PNS is only a few bits. On the other hand, if the band was not allocated any bits, then PNS cannot save bits but spends costs in this band. Therefore, at low bit rates, using PNS technique to retain high frequency signals may lead to obvious degradation in low frequency range since the high frequency signals were not allocated any bit if PNS did not substitute them.
Figure 15 illustrates an example to reveal the difference. Without limitation on bit, the signals in low frequency range are sparser than the one with limitation on bit.
Such distortion in low frequency can degrade the quality significantly although it retained the signals in higher frequency range. This thesis proposes a method to prevent such situation based on the bits allocated to the frame. The bits used for PNS to substitute high frequency signals will not exceed a ratio R of the available bits in a frame. After extensive experiments on objective quality measurement and subjective listening test, this thesis suggests the suitable value for R to be 15%. With such limitation on bit, the qualities of coded signals in low frequency range will not be influenced by retaining high frequency signals via perceptual noise substitution technique.
Figure 15: Limitation on bit usage in high frequency signal substitution
Input (frequency domain)
Spectral Flatness Measurement
Highly Tonal Band ?
Noise Floor Estimation
PNS Parameters Evaluation
END
N
Y
Start Frequency Decision
Bit Rate Check
Y
N
Bit Limitation ?
N
Y
Roll Back
Figure 16: The flowchart of high frequency signal substitution