Sara Lyme- > i like to know what is the characteristic of the input signal that > i will used in melp program (format ,frequency ,bit,mono )and i > know that melp work well for male speakers than female speakers please could you > explaint why. > please i need this information as soon as possible. Suggest to use 8 kHz sampling rate, 16-bit samples. Each MELP encode/decode is one channel, therefore "mono" in audio terminology -- if you want stereo, then you need two (2), fully independent MELP channels in your algorithm. Why does MELP tend to work better for male speakers? In general because female speakers tend to have somewhat higher pitch and also slightly more bandwidth -- that makes it more difficult for the algorithm to compress and encode information as accurately as with lower pitch and less bandwidth. You will find the same phenomenon with other languages, for example Chinese which is more tonal and for which speakers tend to use a bit more bandwidth. This happens with other LBR codecs, not only MELP. -Jeff |
|
Re: [Fwd: ask help]
Started by ●February 11, 2004
Reply by ●February 15, 20042004-02-15
Sara- > thank you for your help ,please I have Run Melp on dskc6711 ,and in out (speaker) > sometimes it give me (signal) but not speech and sometime it give me nothing > ,please could you explain me why and what I have to do(i must to have a speech) and > please how i can remove noise . I don't think MELP or MELPe C code can run in real-time on DSK C6711. Even single-channel full-duplex will take too many MFLOPS unless it is optimized. MELP is a computationally intensive algorithm -- the C54xx of MELPe implementation had to be hand-optimized to fit into 100 MIPS on C5409, and encode/decode combination takes about 65 MIPS on C55xx. That's a lot. -Jeff |
Reply by ●February 18, 20042004-02-18
Sara- > i have another problem when i run MELP on my dsk c6711(TI),like i told you > sometimes give me signal (not speech) and sometimes give me nothing ,and > when i tried to run it for many time the dsk (be saturation it not work)i > swich off the dsk and after i swich on(used just power (cable > d'alimentation) i hear some frame of my speech. > please can you tell what happend Are u using audio input? If so, then my guess is is that u can sometimes hear the first several frames before the algorithm overruns (runs out of real-time) or has some other problem. Why don't u change the number of frames to run to be just a few -- say 20, which would give you about 1/2 second (22.5 msec per 180 pts), and see if u can get more repeatable behavior? Also, try running just low level (silence) and see if u still hear the noise u mention above. If u can get something stable, then u can start making small changes to see what is going on. -Jeff |