Organisation internationale de normalisation
Yüklə 7,54 Mb.
|
| Wednesday 11AM
Max Neuendorf, FhG, presented new information on this CE. The presenter noted that, in his opinion, the CE tries to do something that the RM coder can already do in the FD coding mode. He presented the results of a listening test. The systems under test are
The middle two systems in the list above derive the following mode decisions from ETRI‘s RM8_CE bitstreams:
These mode decisions are imposed on the RQE encoder, but instead of using long-overlap TCX frames, the encoder uses FD mode with a higher SBR cross-over:
The last condition uses current regular short overlap in LPD
With this configuration of “reg”, the only difference between the technology in the two systems is the ETRI technology. The listening test results at 12 kb/s mono show that there is no difference in the absolute scores. In analysis of the differential scores, no difference is observed either between the CE and LovlpAsFD_T80 nor between the CE and LovlpAsFD_T80T40" Taejin Lee, ETRI, continues his presentation on the USAC-TCXW CE. In summing he made the following points:
Bernhard Grill, FhG, noted that the point of this discussion is whether a new tool (TCXW) is needed, or whether the current USAC encoder can deliver the same quality with additional tuning and/or classification modules. Taejin Lee, ETRI, noted that a new encoder tuning that shows merit for this CE may not represent a reference quality encoder tuning. Bernhard Grill, FhG, gave a presentation on his view of the structure USAC. The main point was that, neglecting ACEL mode, USAC is an AAC codec with two means to shape the quantization noise: scale factors or LPC frequency-domain weighting. He noted that one can always improve one coding mode, but what is important is to improve the overall coder performance. For example, he stated that in the FD transform, one could investigate the following options:
Hence one could explore 16 configurations, whereas AAC standardized only 4 configurations (128, 1024; sine, Kaiser-Bessel). He contends that the other configurations would bring small, if any, improvements at the cost of significant increase in encoder control. The basic problem is to select long-overlap window for tonal items and short-overlap for speech items. All tools for this are already in USAC, and the ETRI CE technology is not needed. The Chair asked if the CE imposes any significant constraints on the encoder architecture. Taejin Lee, ETRI, noted that the CE encoder/decoder has a small additional delay (192 samples). The Chair called for a show of hands of experts that do not support adopting the CE technology, and 9 experts raised their hands.
David Virette, Huawei, presented
The contribution presents the results of a listening test a 8 kb/s mono operating point. The two systems under test were Sys1 RM8
Sys2 RM8+CE There is no difference in the absolute scores. For differential scores, the CE was better for 1 item. Philippe Gournay, VoiceAge, presented
The contribution presents the results of a listening test a 8 kb/s mono operating point. There is no difference in the absolute scores. For differential scores, the CE was better for 5 items and for the mean. The presenter noted that the items were better by 1 or 2 MUSHRA points and had small confidence intervals indicate large agreement amongst listeners. The contribution also reports on CE decoder verification: the CE bitstreams decoded to exactly the CE waveforms. Takehiro Moriya, NTT, presented
The contribution reports that the CE was slightly revised, in that it had new Huffman codebooks. This resulted in more than 6 bits savings per ACELP frame (an average of 3% savings for ACELP mode frames at 8 kb/s). The presenter noted that speech items coded at low rates may use ACELP mode for 50% of the frames. Saved bits were were returned to the bit reservoir, but an additional encoder tool was used to steer additional bits to ACELP and to increase the ACELP excitation mode when appropriate. The presenter noted that the CE proposal is purely lossless coding with respect to RM8 operation. Differential coding was exploited, but an ACELP superframe remains a random access entry point. Listening test results for data pooled over all listening sites was presented: Only cross-check sites: 2 better Proponent and cross-check sites: 6 better and mean better. Some improvements were as large as 3 points. When investigating consistency amongst test sites Of 7 items that are better for pooled data, 2 items were judged better at 2 of 3 test sites. Complexity at 8 kb/s: 0.04 WMOPS typical, 0.15 worst case 1025 bytes of Huffman tables The presenter noted that one step higher in ACELP excitation pulse coding requires 32 bits per superframe, hence 6 bits per superframe savings permits the sytem to jump to a higher rate in one of every 6 superframes. However, for voiced regions the savings can be much more and so one could jump to a higher-rate excitation mode more often. MM, FhG, asked what material was used to train Huffman tables. The presenter stated that they are trained based on 100,000 ACELP frames outside of the RM0 test set and also on the RM8 reference quality bitstreams Audio experts requested additional information on:
This discussion will be continued Thursday at 1400 hrs. Takeshi Norimatsu, Panasonic, presented
The contribution presents the results of a listening test at 20 kb/s mono. The systems under test were: Sys1 RM8
Sys2 RM8+CE When analysing absolute scores there were no significant difference. When analysing differential scores: 1 item better (Id4) Eunwoo Song, Yonsei University, presented
The contribution presents the results of a listening test at 16 kb/s mono. When analysing absolute scores: 1 item better (Id4) When analysing differential scores: 1 item better (Id4), 1 item worse (SpeechOverMusic_1) Tomasz Zernicki, Telcordia, presented the following contributions
The presentation reviewed the problem that the CE proposes to solve, namely that the SBR tool in USAC is unable to reconstruct signals with only high tonal components (since there are no low-band harmonics to patch up to the high band). The proposed solution is to use a sinusoidal coding tool to represent the high-band tonal components. The complexity of the decoding tool is: 0.5 WMOPS. 0.64 Kbyte ROM 48 kBytes RAM now, but anticipated to be 10 kBytes after further optimization Listening test results for all data pooled together were presented: Poznan, 16 kb/s: 2 better, some with mean value better by more than 15 MUSHRA points Telcordia, 20 kb/s: 5 better, mean better Poznan, Yonsei, 16 kb/s: 2 items better, 1 item worse Telcordia, Poznan, Panasonic, 20 kb/s: 5 better, mean better The presenter noted that the improved items are for the most part the new test items (e.g. Id4). He further noted that for the one item worse (at Yonsei), the Tonal Components tool was never used; hence any actual negative impact would be due to the 1 bit/frame allocated to the enable/disable flag. Very likely it might be due to a statistical anomaly. It was clarified that the USAC CfP items were coded as one concatenated file, while the new items were encoded separately. It was also clarified that the average tonal components bit budget was subtracted from the nominal USAC bit rate to get an average bit rate for USAC to code the residual signal. Kristofer Kjörling, Dolby, note that this is different from how an actual system would work, where the per-frame tonal components bit requirements would be subtracted from the USAC per-frame residual coding bit budget. The Chair agreed that not modelling the instantaneous bit requirements leaves quite a bit of uncertainty in evaluating the subjective performance data. Max Neuendorf, FhG, noted that there is an 8 frame latency in the encoder since the tool encodes 8 frames as a unit. There was considerable discussion on exactly how this data was transmitted over the channel. Werner Oomen, Philips, reported on a very informal in-house listening test in which the violin item was graded as having worse quality. Kristofer Kjörling, Dolby, summed up his concerns as:
This discussion will be continued Thursday at 1400 hrs. David Virette, Huawei, presented
The contribution summarized the Enhanced Pulse Indexing CE technology. The technology is able to save 1 bit per ACELP track in certain coding modes. This results in an average bit savings of
The Chair noted that the situation of this CE is that the CE technology is guaranteed to deliver increased compression, in that it is a lossless compression technology. However, the difficulty with the CE is that it has been unable to demonstrate that this bitsavings can be translated into an increase in subjective performance. The Chair called for a show of hands of experts that do not support adopting the CE technology, and 10 experts raised their hands. Based on this, the Chair concluded that there is no consensus to adopt the CE technology.
The contribution proposes changes to the USAC bitstream that address the requirements listed in the workplan from the 94th meeting. Possible bitstream changes fall into the following three categories, and the contribution addresses the first two categories:
Config
Payload
The Chair noted that every change to the Extension Element ID will require an amendment to the USAC specification, and suggested that the alternate mechanism of a Registry be considered. He further suggested that the Independency flag be mapped back to the Config, e.g. that every 100 frames are Independent frames and hence random access entry points. Heiko Purnhagen, Dolby, noted that in the philosophy in the proposed design, such information would be at a higher level (e.g. Systems), which is not covered in this contribution. The presenter hopes to bring information on the impact of the syntax design for various application scenarios, e.g. high-bitrate versus low-bitrate and file-based versus broadcast. The presenter requested that the following syntax elements be adopted into the USAC DIS.
The Chair stated that there was no more time to discuss this item. The discussion will continue in Friday’s Audio plenary. Yüklə 7,54 Mb. Dostları ilə paylaş:
|