dear Dr Omar Thank you for the followup. I am forwarding your reply to the egroup; I guess it was, by mistake, sent to me alone. Sameer. -----Original Message----- From: omar nasr [mailto:] Sent: Wednesday, February 04, 2004 3:23 AM To: Subject: RE: [speechcoding] Can Huffmann coding be directly applied to speech samples ? Hi all, well, about how to find the probability distribution, this can be easily got by drawiing the spectrogram of the samples of the speech signal, and know the probability distribution for the samples. of course.... it is a bad way for compression since the lossy compression like MP3, LPC-10,...etc gives better performance, but u can use it in simple non-real time compression, so if u have , say 1 hour of speech, start by drawing the histrogram to know the PDF of the speech signal,most probably u will give lower number of bits for low amplitudes cause they r in general comes more often than higher amplitudes, and do the huffman coding, and send the huffman table with the coded speech this may give u 50% compression in some cases Omar AHmed Nasr comm-DSP TA -faculty of Engineering- Cairo University -----Original Message----- From: Sameer Kibey [mailto:] Sent: Tuesday, February 03, 2004 2:19 PM To: 'Jaydeep Inamdar'; Subject: RE: [speechcoding] Can Huffmann coding be directly applied to speech samples ? hi Jaydeep to derive the huffman codes for a given source, you need to know the probablity distribution for various symbols. Speech signal is highly random.. the loudness of speech can vary over abt 30 dB for the same speaker and generally there will be more than one speaker. hence I doubt if it is possible to estimate an optimal model of the probability distribution at the source, as far as speech is concerned. moreover, huffman coding is a lossless coding technique. most standard speech codecs use codebooks, which are lossy, but still give great compression. hence, directly representing speech samples using hufman coding may not be a very effective way of compressing speech. best regards, Sameer. |
|
FW: Can Huffmann coding be directly applied to speech samples ?
Started by ●February 4, 2004
Reply by ●February 4, 20042004-02-04
Hi All
To add on more to this topic, speech exhibits a Laplacian PDF
function.
Laplacian tends to be a squeezed in version of the general guassian
function.
This is the basis for mu-law and A-law compression which is inherently
supported by DSPs. Here more bits are allocated to low amplitudes and lesser to
the higher amplitudes.
This also goes for most waveform based compression techniques such as
ADPCM.
In fact if you take a look at mu-law and A-law its a sort of a reverse
huffman based method which assigns more bits to low amplitudes as compared to
high amplitudes.
cheers
Sameer Kibey <s...@tataelxsi.co.in> wrote: dear Dr Omar Shree Jaisimha |