Hi , I would like to convert Melp 2400 bps(Ver. 1.2) to 1200 bps for my academic thesis,Please could any one send me any idea ,comments or source code as soon as possible. Regards machd musafi |
|
Low Bit Rate Coding
Started by ●November 8, 1999
Reply by ●November 15, 19991999-11-15
Dear Musafy, I had a student dealing with the MELP coder, so I have a list of references to papers, and web pointers to code and the draft, in case you would like to have it (probably you already have all that, as you speak about version 1.2). Otherwise we did not think about methods to reduce the bit rate even further. There is an article in ICASSP'99 from the TI people that developed MELP. They increase the bit rate to 4 kbps, to improve the quality. Maybe this can give you ideas, on where to find the trade-offs. A possibility could be to increase the frame size, from 22.5, to let's say 30 ms. But you should do some adaptation work, and check if the coder support that, and that you still get an acceptable quality. The algorithmic delay will also be bigger. I also got the impression that they have have different information to model the residual. One is 10 (?) FFT magnitude peaks they transmit, and they say this improve quality of reconstructed speech. Maybe you could try to remove, or reduce this information, and see if the quality is still aceptable for your application. Regards, Sara wrote: original article:http://www.egroups.com/group/speechcoding/?startG > Hi , > > I would like to convert Melp 2400 bps(Ver. 1.2) to 1200 bps for my academic > thesis,Please could any one send me any idea ,comments or source code as > soon as possible. > Regards > machd musafi > |
Reply by ●November 16, 19991999-11-16
Dear Musafi, I was checking the references I have on the MELP, and I found two articles, which I have forgotten about, that maybe useful for your work: [1] Alan McCree and Juan Carlos De Martin, "A 1.7 KB/S MELP CODER WITH IMPROVED ANALYSIS AND QUANTIZATION", ICASSP'98. [2] Alan McCree and Juan Carlos De Martin, "A 1.6 KB/S MELP CODER FOR WIRELESS COMMUNICATIONS", IEEE SPEECH CODING WORKSHOP, PENNSYLVANIA, SEPTEMBER 1997, pp.23-24. I check them quickly, and contrary to what I advised you in my last posting, they reduce the frame length from 22.5 ms to 20 ms. They also improved pitch and voicing estimation and added a noise supression front-end. The bit rate decrease comes from: - better LSP quantization, instead of 25 bits, 21 bits in [1] and 20 bits in [2]. - removed the Fourier magnitudes (8 bit saving). - transmit the gain only once per frame, as the frame is now shorter (from 8 to 5 bits). - reduce from 7 to 6 the bit used for pitch and overall voicing - changed (removed, saving of 1 bit) the aperiodic flag by a "pitch contour perturbation technique" (please do not ask me what it is :-) - reduced the number of bits for bandpass voicing from 4 to 2, by selecting from a catalog of of 4 partial voicing patterns. - the sync bit was removed. I have the impression that they really squezed every possible bit, and still they are at 1.6 kbps. Maybe for further reduction you would lose some quality. Do not forget to check the 4 kbps in the ICASSP'99. Regards, Sara |