Shop OBEX P1 Docs P2 Docs Learn Events
Adpcm — Parallax Forums

Adpcm

Jay KickliterJay Kickliter Posts: 446
edited 2010-10-15 06:12 in General Discussion
Does anyone understand ADPCM? I'm trying to get a handle on so I can write some propeller code.

I get how it works, but when I run a test in Sysquake (great MATLAB-like program), the encoder seems to take a few beats to get going. That would be fine for a long transmission, but might cause problems for packetized transmissions that need to play even if some are lost(?).

Attached is a pdf showing the original, encoded, and decoded signals. The ADPCM code is from the MATLAB website and the author adopted it from Microchip AN643.
pdf
pdf
ADCM Test.pdf
23K

Comments

  • Dave HeinDave Hein Posts: 6,347
    edited 2010-10-11 08:48
    Jay,

    In general, an ADPCM encoder will require several samples for the predictor to adapt to the incoming signal. However, it should be able to generate a coarsely quantized version of the signal while it is adapting. I'm not sure why your code generates a zero output during the startup -- It must be due to the way the encoder state is initialized.

    Try modifying the input signal by inserting 16 zero-valued samples at the beginning, and see how it adapts to that signal. I suspect it's just an initialization problem, and the encoder should work fine after the first 10 samples.

    There will be a mismatch between the encoder and the decoder if the decoder does not receive all of the packets. The encoder contains a virtual decoder within its prediction loop that tracks the decoder at the receiver. If the receiver misses data it's decoder will be out of sync with the encoder's predictor. However, G.721's predictor is very "leaky", which allows the encoder and decoder's predictors to converge very quickly. I think the time constant is on the order of 16 to 32 samples.

    Dave
  • ercoerco Posts: 20,261
    edited 2010-10-11 09:02
    Dave: WOW. I am in speechless awe of your reply. It sounds great and I'm sure it's spot-on, but quite frankly I have some studying to do before I can comprehend it in all its technical glory. I have much to learn from you and the other Forum wizards here. Thanks for posting. Best,

    erco
  • Dave HeinDave Hein Posts: 6,347
    edited 2010-10-11 09:33
    erco,

    Sorry if my response wasn't clear. I was responding directly to Jay's question, and I didn't provide much background on how ADPCM works.

    ADPCM stands for adaptive-differential-pulse-code-modulation or adaptive-differential-predictive-coding-method. PCM (pulse-code-modulation) is just representing an analog signal as digital numbers. For voice, the signal is usually sampled at 8,000 samples per second, and represented by 13 or 14 bits. For telephone transmission a companding algorithm is used to compress each sample to 8 bits using a logarithmic mapping.

    ADPCM was developed to compress the audio signal down to 4 bits per sample so that twice as many voice calls could be sent over the same digital transmission line. An ADPCM encoder predicts the value of the next sample based on the previous audio that was transmitted to the other side. It subtracts the predicted value from the input value, and transmits the difference instead of the original signal.

    This works fine as long as the decoder receives all the information sent by the encoder. If some of the transmitted data is lost, then the decoder has to guess at the what the predicted value should be. The predictor only uses the past few samples for the prediction, so eventually the encoder and decoder get back in sync.

    If your interested in more details you could check Wikipedia or do a web search on ADPCM.

    Dave
  • ercoerco Posts: 20,261
    edited 2010-10-11 13:21
    Molto interessante e mille grazie, il mio amico. I will delve. Thanks very much for adding the tantalizing technical details.
  • Jay KickliterJay Kickliter Posts: 446
    edited 2010-10-11 14:02
    Dave,

    Thanks, I was wondering if it had to do with initialization. I have no way of know if they guy who translated the original C code to MATLAB knew what he was doing. I'll try to figure out if there's some error there. The fact that this code takes in a floating point value between -1 and 1 show's there's already something different than the original ADPCM algorithm.
  • davejamesdavejames Posts: 4,047
    edited 2010-10-12 08:47
    Hey Jay- side topic...what ever happened to the op-amp wierdness a few weeks back?

    Any resolve?
  • Jay KickliterJay Kickliter Posts: 446
    edited 2010-10-12 13:26
    Dave, never did get it to work. Tried almost everything. The only way it works is buy making the second stage a 1st order low pass instead of 2nd order. Something about that strange feedback path was messing it up. But, that's not the circuit in the data sheet.
  • Dave HeinDave Hein Posts: 6,347
    edited 2010-10-13 11:16
    Jay,

    I downloaded the ITU C code for G.726 and ran it on your test data. The results for the first 50 samples are shown in the attached graph. The original data is in blue, and the decoded data is in red. It takes a few samples for the encoder to adapt to the signal, but it looks pretty good after about 5 or 6 samples.

    Let me know if you want to try the C code. I can zip it up and send it to you.

    Dave
    graph.GIF
    1150 x 812 - 23K
  • Dave HeinDave Hein Posts: 6,347
    edited 2010-10-14 08:28
    Jay,

    I realized that I might not be scaling the test data the same way you were. My original simulation used a 13-bit range. I changed this to a 16-bit range, and I now get the same results you got. It takes longer for the encoder to adapt it's gain for the larger amplitude signal.

    Dave
    graph2.GIF
    1020 x 730 - 17K
  • davejamesdavejames Posts: 4,047
    edited 2010-10-15 06:12
    "never got it to work"...rats! I was hoping for some enlightenment.

    I've a number of projects like that,

    DJ
Sign In or Register to comment.