|
Hi, Unlike the speech coders from the CELP-family WI is a parametric coder. The more bits you "invest" in CELP-like coder the higher accuracy of the waveform reconstruction, but it is not so for WI. WI strongly depends on pitch detection accuracy, CW-alignment and CW-decomposition; if these procedures fail (completely or partially) any amount of bits for SEW/REW quantization (including non-quantization) will not help you. Even if in addition to non-quantized SEW- and REW-amplitude spectra you transmit original SEW- and REW-phase spectra (which is not interesting for commercial WI) you still cannot get a perfect reconstruction of the waveform because the initial phase in phase-tracking procedure is arbitraty. So, don't judge your coder according to accuracy of the waveform reconstruction. Try to focus on the following most annoying problems: 1. Good pitch detector 2. Good CW-alignment algorithm 3. Good CW-decomposition algorithm For quality measurement use PESQ or PSQM. The following paper might interest you: C.H. Ritz, I.S. Burnett and J. Lukasiak, "Extending Waveform Interpolation to Wideband Speech Coding", 2002 IEEE Speech Coding Workshop. Ilya Druker --- In , "empush2000" <mpush2000@m...> wrote: > Hi everyone, > I am currently working on waveform interpolative speech coding. I > exploit the usual scheme, but it doesn't provide accurate (perfect) > waveform reconstruction even when using no quantization. Does anyone > know any papers/info about perfect reconstruction for waveform > interpolation coder? Please advise. > > Thank you. |