Hello David, Thank you for proceeding with the priject. The 2100 bps samples sound great and I cannot hear a big difference tot he 2500 bps samples on the loudspeaker.
Regarding differencial coding, I think the 1200 bps versions of MELP are using following proceedure: Say 5 frames are packed together. First frame is fully coded, the next frames are differentially coded. Delay of course is also 5 times higher. With 28bps fort he first and 4 times say 16 bps we would get 18.4 bps as medium. Best regards Gerhard OE3GBB -----Ursprüngliche Nachricht----- Von: David Rowe [mailto:[email protected]] Gesendet: Freitag, 28. Oktober 2011 05:31 An: [email protected] Betreff: [Freetel-codec2] 2100 bit/s candidate Hello, I am back developing Codec 2 again. 1/ I have been working on line spectrum pair (LSP) quantisers, and have attached some samples from a 2500 bit/s and 2100 bit/s version. The 2500 bit/s version uses 4,4,4,4,3,3,3,3,3,2 bit scalar LSP quantiser (36 bit/fr). The 2100 bit/s version uses scalar quantisers of 4,4,4,4 bits for the first 4 LSPs, than a 12 bit (4096 entry) VQ for the last 6 LSPs (28 bits/frame). They sound fairly similar to me, what do you think? 2/ While playing with LSPs I found that "quantising" LSPs 1 to 10 to 25,25,50,50,100,100,100,100,100,100 Hz steps has no effect on speech quality. In other words an error of +/- 50Hz in say LSP 10 (or +/- 12.5Hz in LSP 1) can't be heard. This is a little like bandwidth expansion of LSPs, or "just noticeable differences". I am wondering if this effect can be used in quantiser design or vector codebook searching as a metric for how perceptually "close" two candidate vectors are to the target. Or we could discard entries in a VQ during training that are close enough to be perceptually indistinguishable. Another idea is to use this effect in coding frame-frame differences. During perceptually important strongly voiced speech LSPs change slowly. Changes that are not perceptually noticeable don't need to be quantised and transmitted. So if we take a frame-frame deltas of LSPs e.g. 25,-25,25,25,25,25,25,25,25,25 then set changes smaller than the limits above to zero we might have a perceptually-equivalent frame-frame delta of 25,-25,0,0,0,0,0,0,0,0. Presumably this would take less bit's to encode. Cheers, David ------------------------------------------------------------------------------ The demand for IT networking professionals continues to grow, and the demand for specialized networking skills is growing even more rapidly. Take a complimentary Learning@Cisco Self-Assessment and learn about Cisco certifications, training, and career opportunities. http://p.sf.net/sfu/cisco-dev2dev _______________________________________________ Freetel-codec2 mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/freetel-codec2
