Academia.eduAcademia.edu

Figure 8 – uploaded by John Mourjopoulos

See full PDFdownloadDownload figure
Fig. 4. PE bitrate requirements, expressed as function of time, for the conventional PAC method and the proposed DPAC method: (a) results for big band jazz segment (JZZ). (b) results for orchestral music segment (MZT).

Figure 4 PE bitrate requirements, expressed as function of time, for the conventional PAC method and the proposed DPAC method: (a) results for big band jazz segment (JZZ). (b) results for orchestral music segment (MZT).

Related Figures (20)

Transform Coding methods are traditionally employed for high- quality audio data compression applications mainly because they operate in the frequency domain, which in turns results in performance gains obtained from the uncorrelated nature of the spectral components and the ability to minimize perceived distortions by optimal shaping of quantization noise in this domain [3]. As will be discussed in more detail later, this advantage can be further exploited by using the masking phenomenon of the human auditory system. Fig. 1 gives the block diagram of a Transform Coder and Decoder.
Transform Coding methods are traditionally employed for high- quality audio data compression applications mainly because they operate in the frequency domain, which in turns results in performance gains obtained from the uncorrelated nature of the spectral components and the ability to minimize perceived distortions by optimal shaping of quantization noise in this domain [3]. As will be discussed in more detail later, this advantage can be further exploited by using the masking phenomenon of the human auditory system. Fig. 1 gives the block diagram of a Transform Coder and Decoder.
ig. 2. Basic block diagram of a Perceptual Audio Coder (PAC).
ig. 2. Basic block diagram of a Perceptual Audio Coder (PAC).
ig. 3. Block diagram of the Differential Perceptual Audio Coder (DPAC). The Spectral Coefficient Selection (SCS) module replaces the spectral components of a Simple Frame using the respective components of the Reference Frame, when the replacement criterion is satisfied. Furthermore, the operations of this module differ depending on its application to Simple or Reference Frames. Clearly, the SCS results in a substitution which will generate errors which differ from those given by equation (4). Here, a study of this error will be provided which will be complemented by experimental results given in section aT
ig. 3. Block diagram of the Differential Perceptual Audio Coder (DPAC). The Spectral Coefficient Selection (SCS) module replaces the spectral components of a Simple Frame using the respective components of the Reference Frame, when the replacement criterion is satisfied. Furthermore, the operations of this module differ depending on its application to Simple or Reference Frames. Clearly, the SCS results in a substitution which will generate errors which differ from those given by equation (4). Here, a study of this error will be provided which will be complemented by experimental results given in section aT
and its variance, (assuming that the vectors yp and ys are uncorrelated), can be described by: The above equation shows that the variance of the total spectral substitution error for the complete N-point block signal, i.e. Noss’, equals to the difference of the variances as and of’, only with respect to the substituted coefficients. The quantized version of the vector y.,, is o1ven hv:
and its variance, (assuming that the vectors yp and ys are uncorrelated), can be described by: The above equation shows that the variance of the total spectral substitution error for the complete N-point block signal, i.e. Noss’, equals to the difference of the variances as and of’, only with respect to the substituted coefficients. The quantized version of the vector y.,, is o1ven hv:
Assuming that the vectors of the quantization and the spectral substitution are uncorrelated, then it is:
Assuming that the vectors of the quantization and the spectral substitution are uncorrelated, then it is:
The above equation shows that the variance of the total DPAC error, which is introduced by the spectral substitution and the quantization procedures in a Simple Frame, is the summation of the variance of the quantization noise for these components which are not affected by the SCS module, the variance of the quantization noise for the spectral components of the corresponding Reference Frame which are modified by the SCS module, and the variance of the spectral substitution error as is described by equation (26). This indicates that the additional error of the proposed DPAC method, is mainly due to the difference in variance between the frames’ original and substituted spectral coefficients. The above conclusion shows that the DPAC error
The above equation shows that the variance of the total DPAC error, which is introduced by the spectral substitution and the quantization procedures in a Simple Frame, is the summation of the variance of the quantization noise for these components which are not affected by the SCS module, the variance of the quantization noise for the spectral components of the corresponding Reference Frame which are modified by the SCS module, and the variance of the spectral substitution error as is described by equation (26). This indicates that the additional error of the proposed DPAC method, is mainly due to the difference in variance between the frames’ original and substituted spectral coefficients. The above conclusion shows that the DPAC error
ig. 5. Block diagram of the Differential Perceptual Audio Decoder. The decoder has a simpler structure compared to the encoder. One bit flag which is transmitted as side information, allows t discrimination of the type of the processed frame. If a Reference Frame is processed, then the decoding operation is the same as with tl established perceptual decoders. If a Simple Frame is decoded, then tl zeroed quantized values are substituted by the corresponding values the Reference Frame. It is obvious that the differential decoder has ne ne more complex structure compared to the established perceptual decod schemes and also requires some amount of memory (i.e. N/2 memo ler ry location), in order to store the decoded values of the Reference Frame. However, this additional complexity is by no means prohibitive for its real time implementation of the DPAC method. The block diagram the differential decoder is shown in the Fig. 5. of
ig. 5. Block diagram of the Differential Perceptual Audio Decoder. The decoder has a simpler structure compared to the encoder. One bit flag which is transmitted as side information, allows t discrimination of the type of the processed frame. If a Reference Frame is processed, then the decoding operation is the same as with tl established perceptual decoders. If a Simple Frame is decoded, then tl zeroed quantized values are substituted by the corresponding values the Reference Frame. It is obvious that the differential decoder has ne ne more complex structure compared to the established perceptual decod schemes and also requires some amount of memory (i.e. N/2 memo ler ry location), in order to store the decoded values of the Reference Frame. However, this additional complexity is by no means prohibitive for its real time implementation of the DPAC method. The block diagram the differential decoder is shown in the Fig. 5. of
where x(n) is the n-th time-domain sample in the Reference Frame of block size N, and xs(n) is the n-th time-domain sample in the frame under test. The decision level A was obtained from experimental results and was also found to depend on the type of the processed audio material. The above criterion is entirely objective, since no psychoacoustic weighting is employed, and represents the difference between the two waveforms. According to Parceval's theorem this criterion has an equivalent frequency domain interpretation. A similar criterion was proposed in [11], in order to select the optimal block-size in an ATC technique with adaptive block-size MDCT.
where x(n) is the n-th time-domain sample in the Reference Frame of block size N, and xs(n) is the n-th time-domain sample in the frame under test. The decision level A was obtained from experimental results and was also found to depend on the type of the processed audio material. The above criterion is entirely objective, since no psychoacoustic weighting is employed, and represents the difference between the two waveforms. According to Parceval's theorem this criterion has an equivalent frequency domain interpretation. A similar criterion was proposed in [11], in order to select the optimal block-size in an ATC technique with adaptive block-size MDCT.
Fig. 6. MNR and corresponding PE reduction, for the different values ot parameter €: (a) €=0.1 and (b) €=0.3. i.e. the SFM can be computed as the ratio of the geometric mean of the power spectrum to the arithmetic mean of the power spectrum. The inverse of the SFM, ye, is a measure of signal predictability and typically has values between 3 and 16 for long-term power spectral densities of speech signals [4]. Viewed in a different way, the SFM represents the global tonality of the short time spectrum of the signal.
Fig. 6. MNR and corresponding PE reduction, for the different values ot parameter €: (a) €=0.1 and (b) €=0.3. i.e. the SFM can be computed as the ratio of the geometric mean of the power spectrum to the arithmetic mean of the power spectrum. The inverse of the SFM, ye, is a measure of signal predictability and typically has values between 3 and 16 for long-term power spectral densities of speech signals [4]. Viewed in a different way, the SFM represents the global tonality of the short time spectrum of the signal.
If this value was found not to exceed a predefined (from experimental results) value, then this indicate that no significant audible changes existed between the compared spectra and in this case the frame under test was assumed to be a Simple Frame.
If this value was found not to exceed a predefined (from experimental results) value, then this indicate that no significant audible changes existed between the compared spectra and in this case the frame under test was assumed to be a Simple Frame.
Fig. 9. Reference Frame allocation along a music piece (JZZ), for the different FS criteria. It is also significant that for all criteria, the amount of variance due to Spectral Substitution (equation (19)) was found to be linearly related to the audio quality, as is shown in Fig.8.
Fig. 9. Reference Frame allocation along a music piece (JZZ), for the different FS criteria. It is also significant that for all criteria, the amount of variance due to Spectral Substitution (equation (19)) was found to be linearly related to the audio quality, as is shown in Fig.8.
Variance of the Spectral Substitution error in relation to MNR for the different FS criteria (JZZ music piece).
Variance of the Spectral Substitution error in relation to MNR for the different FS criteria (JZZ music piece).
FS criterion. A good illustration of this is given by Fig. 9, where the allocation between reference and simple frames is shown, along the time evolution of a music piece (JZZ), and for each of the 3 FS criteria. In this figure, the number of reference frames allocated by each criterion was kept the same and the corresponding PE reduction and audio quality is also shown. Fig. 10. Perceptual Entropy distribution for the conventional PAC and the proposed DPAC technique, (a) results for big band jazz music (JZZ) and, (b) results for orchestral music (MZT).
FS criterion. A good illustration of this is given by Fig. 9, where the allocation between reference and simple frames is shown, along the time evolution of a music piece (JZZ), and for each of the 3 FS criteria. In this figure, the number of reference frames allocated by each criterion was kept the same and the corresponding PE reduction and audio quality is also shown. Fig. 10. Perceptual Entropy distribution for the conventional PAC and the proposed DPAC technique, (a) results for big band jazz music (JZZ) and, (b) results for orchestral music (MZT).
Fig. 11 shows the bitrate gain of the DPAC method, expressed as average reduction in the value (in bits) of the audio samples. This diagram is obtained from the difference of DPAC and PAC values, shown in Fig. 10.
Fig. 11 shows the bitrate gain of the DPAC method, expressed as average reduction in the value (in bits) of the audio samples. This diagram is obtained from the difference of DPAC and PAC values, shown in Fig. 10.
DPAC method NMR Measurements for the Spectral Substitution process and the complete coding process for different types of music.
DPAC method NMR Measurements for the Spectral Substitution process and the complete coding process for different types of music.
g.13. | Comparison in audio performance between PAC and DPAC methods, at similar bitrates: (a) 128 kBits/s, (b) 88 kBits/s, (c) 64 kBits/s Fig. 12. | Overall NMR measures for PAC and DPAC method along the evolution of the same music segment: (a) results for big band jazz music (JZZ) and, (b) results for orchestral music (MZT) (bitrate for both tests: 128 kBits/s).
g.13. | Comparison in audio performance between PAC and DPAC methods, at similar bitrates: (a) 128 kBits/s, (b) 88 kBits/s, (c) 64 kBits/s Fig. 12. | Overall NMR measures for PAC and DPAC method along the evolution of the same music segment: (a) results for big band jazz music (JZZ) and, (b) results for orchestral music (MZT) (bitrate for both tests: 128 kBits/s).
segments upon which tests were performed in order to assess th performance of the DPAC method. Fig. 14. Audio quality difference between DPAC and PAC method at different bitrates and for different types of audio material: (a) author's database and (b) ITU-R database. Audio quality difference between DPAC and PAC method at different bitrates and for different types of audio material: (a) author's database and (b) ITU-R
segments upon which tests were performed in order to assess th performance of the DPAC method. Fig. 14. Audio quality difference between DPAC and PAC method at different bitrates and for different types of audio material: (a) author's database and (b) ITU-R database. Audio quality difference between DPAC and PAC method at different bitrates and for different types of audio material: (a) author's database and (b) ITU-R
Related topics:
EngineeringComputer ScienceSignal ProcessingAudio Signal ProcessingData CompressionAudio CodingLanguage Culture and CommunicationSpeech CodingApplication SoftwarePsychology and Cognitive Sciences
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved