Organisation internationale de normalisation
Yüklə 3,08 Mb.
|
- Bu səhifədəki naviqasiya:
- 2.5Review of AHG reports
- 2.7Received National Body Comments and Liaison matters
- 3Record of AhG meetings
- 3.1.2SAOC 1400-1800
2.3Creation of Task GroupsTask groups were convened for the duration of the MPEG meeting, as shown in . Results of task group activities are reported below. 2.4Approval of previous meeting reportThe 84th Audio Subgroup meeting report was registered as a contribution, and was approved. 2.5Review of AHG reportsThere were no requests to review any of the AHG reports. 2.6Joint meetingsThe joint meetings with Audio for the week are shown below:
2.7Received National Body Comments and Liaison mattersThe NB Comments and Liaison documents for the meeting that require a response are as shown below.
3Record of AhG meetings3.1AhG Meeting on SAOC, Unified Speech and Audio Sunday 1000-17003.1.1Unified Speech and Audio 0900-1400Test Site Reports The first set of presentations reported on the test site listening facilities. These were very straightforward and factual, and in each case there was no discussion to record. Kristofer Kjörling, Dolby, presented
Anisse Taleb, Ericsson, presented
Taejin Lee, ETRI, presented
Herve Taddei, Huawei, presented
Pierrick Philippe, France Telecom, presented
Markus Multrus, FhG, presented
Dong Soo Kim, LGE, presented
Werner Oomen, Philips, presented
Eunmi Oh, Samsung, presented
Oliver Wuebbolt, Thomson, presented
Roch Lefebvre, VoiceAge, presented
System Descriptions The next set of presentations reported on the proponent system architectures. Kristofer Kjörling, Dolby, presented
The Dolby/Philips system builds on HE-AAC V2 architecture, but adding prediction-based time domain processing elements. The tools in HE-AAC V2 have been slightly modified so as to offer additional compression performance for the diverse speech and audio content. The MDCT permits a number of window lengths and block sizes, in which all transform bins have equal bandwidth. All AAC-LC tools were available in the core coder. The codec delay can be as much as 1.2 seconds, which is due primarily to encoder look-ahead and blocking into super-frames. Taejin Lee, ETRI, presented
The ETRI system combined the individual tools from AMR-WB+ and HE-AAC V2 to create a new codec, with modifications to the tools as needed to address the coding of diverse content types. It used two processing models: Steady State using linear prediction coding tools, and Complex State using transform coding tools. There is a very flexible signal state decision module that results in a hard decision to use one of the two processing models for each block. Bernhard Grill, FhG, presented
The FhG system is a collaboration between Fraunhover IIS and VoiceAge. The system combines elements from MPEG Surround, the SBR tool, AAC and ACELP linear prediction coding. The tools were modified as appropriate for better performance for coding diverse speech and audio content. MPEG Surround is used for stereo processing, although it can easily be extended for processing multi-channel signals. In the submission it only uses a single OTT box. The SBR technology employed the same filterbank as MPEG Surround, but was enhanced to support very low cross-over frequencies. It can be switched off for high-bitrate operation. The core coder is comprised of a Linear Prediction tool and a MDCT Transform coder or Time-Domain coder. In the core coder, the Linear Prediction tools can be activated on a block-by-block basis or can always be active. The LP tool is followed by either a time-domain coding tool (ACELP) or a frequency-domain coding tool (AAC). At high bitrates the system is virtually identical to AAC. The AAC tool uses an improved entropy coder and optionally uses time-warping for improved performance on speech signals. The possible configurations of the core coder are:
Philippe Gournay, VoiceAge, presented
The VoiceAge system is a collaboration between Fraunhover IIS and VoiceAge. The decoder is essentially the same system as was presented in the FhG contribution. The time-domain core coding tool is the AMR-WB ACELP coding tool. The LPC coding tool is typically switched on for speech-like signals. In this case, the core coder can switch between time-domain (ACELP) and frequency-domain coding of the residual (TCX) on a block-by-block basis. If the LPC coding tool is switched off, the coder uses frequency-domain coding. At high bitrates the codec is very similar to AAC. The VoiceAge encoder employed a different psychoacoustic model, SBR tool and MDCT coder. The MDCT did not employ time-warping. The complexity of the decoder is no more than 1.8 times that of HE-AAC V2.
The LG system switched between three coding modules for coding speech-rich signals, music-rich and mixed speech and music. The speech-rich coder employs a modified AMR-WB+ tools, music-rich coder used a modified HE-AAC V2 tools and mixed coder uses a residual coding scheme. The decoder employs a delay compensation process so that the system can switch between coding modes on a block-by-block basis. JungHoe Kim, Samusng, presented
The presentation noted that coding of music signals is most effective using atransform coder with perceptual threshold for shaping the quantization noise while coding of speech signals is most effective using a linear predictive speech model. The Samsung system uses a non-uniform bandwidth MDCT tool and any of the T/F information can be coded be either the high-temporal resolution tool or HE-AAC like quantization tools. In addition, it uses a TNS tool, a SBR tool and a parametric stereo tool. The high-temporal resolution tool is only used at the lower bitrates. In discussion it was noted that, without considering quantization errors, the variable T/F MDCT tool operates such that time-domain aliasing is always cancelled. Oliver Wuebbolt, Thomson, presented
The Thomson system uses a speech/audio switch to select either a time-domain coding tool or a frequency-domain coding tool in the core coder. In addition, it incorporates SBR and Parametric Stereo coding tools. Analysis of Listening Test Data The final set of presentations reported on analysis of the listening test data. Werner Oomen, Philips, presented
This contribution presents the Philips and Dolby experts analysis of the listening test data. The data was analyzed to determine if the requirements as specified in the Evaluation document were met, for both 95% and 99% confidence intervals. The results showed that systems 2, 4, 8, 10 and 11 were promising systems with respect to the Requirements analysis. Taejin Lee, ETRI, presented
This contribution presents an analysis of the Figure of Merit statistics. It presents a count of how often each system has the highest mean performance for each test and each signal category. Its conclusion is that there is no dominant best system. Pierrick Philippe, France Telecom R&D, presented
This contribution presented graphs of system performance as compared to VC. For each signal category it performed single-sided T-tests with respect to the best performing system
The overall conclusions are that systems 2, 4, 8, 10 and 11 all seem promising for this new work. Ralf Geiger, FhG, presented
The contribution first presents the calculation of VC for each test and each category. For the Requirement that the New Technology be no worse than VC, only sys10 and sys11 meet the requirements for very test and every category for both the 95% and 99% levels of significance. For the Figure of Merit, only sys10 and sys11 have a non-negative “d” statistic for all tests, at the 95% level of significance. When averaged over all tests, only sys2, sys10 and sys11 have a non-negative “d” statistic, at the 95% level of significance, but sys10 has the best mean “d” statistic, but this mean value is not different from that of sys11 at the 95% level of significance. Roch Lefebvre, VoiceAge, presented
This contribution presents a number of interesting views of the test results based on ranking of the proponent mean scores based on test, signal category and signal. The overall conclusion is that sys10 is the best overall proposal, followed closely by sys11. Eunmi Oh, Samsung, presented a statistical analysis of listening test data, which was provided in contribution m15567. It presented a table showing which systems had the best performance by test and signal category, and also showed the systems whose performance were not different from the best. It concluded that there is no dominant system when performance is viewed by test and signal category. Systems sys2, sys4, sys8, sys10 and sys11 all seem promising. Frederik Nagel, FhG, noted that presenting a table of results based on best mean score might be misleading, in that several systems may have mean scores that are not different from the best at the 95% level of significance. Dong Soo Kim, LG, presented
This contribution presents a statistical analysis of listening test data. This was registered as m15713 and was uploaded on Sunday July 20. For the Requirements, it showed that only sys10 and sys11 pass the Requirements for all tests and all signal categories. Furthermore, it noted that sys4 passed the Requirements most often for speech category signals and sys2 passed the Requirements most often for mixed category signals. It presents a Figure of Merit analysis by signal, sys2, sys10 and sys11 all have similarly good performance. Review of AhG Report The AhG Chairs presented the AhG report and proposed
And additionally summarize the test results as
3.1.2SAOC 1400-1800CE: Separation of real-environment signals Osamu Shimada, NEC, made short presentation that reminder the Audio experts the framework for and meaning of test1 and test2 in this core experiment. Jeongil Seo, ETRI, presented the test1 results in
Yang-Won Jung, LG, presented the test1 results in
Leonid Terentiev , FhG, presented the test1 results in
The contribution showed that all listeners in Test1 had at best “no preference” and at worst “did not prefer” the proposed technology. There was considerable discussion on the slide which presented mean score for all listeners in Test 1. Osamu Shimada, NEC, presented
The set of contributions were discussed. The Chair noted that averaging over listener responses may obscure the individual’s intent rather than minimize the noise in the listener responses. It was decided that a break-out group will consider how best to interpret the data we have at this meeting and report back later in the week. Yüklə 3,08 Mb. Dostları ilə paylaş:
|