[Lame-dev] mp3 decoding using a neural network

[Matt F. <fl...@fl...>]

Hi all

First of all, thanks for implementing lame!

I am wanting to train a neural network to see how it performs at 
decoding the mp3 stream.
In order to do this, I would like to have the bitstream frame by frame 
and the audio PCM frame by frame.
I have modified lame to try to extract frame data and audio PCM data 
frame by frame.

Here is the fork which includes my modifications :
https://github.com/flatmax/lame/commit/7851ba71a138e79bf433660d09fbe8f366910358

To create a row I simply write "\n" to file which indicates the start of 
the next frame's row.

When I run this code using an mp3 file like so :
~/temp/lame/frontend/lame -q --cbr -b 192 -c --tt "15.Awareness" --ta 
"Flatmax, Rob J" --tl "Wideband Cold Fusion" --ty 2008 --tc 
flatmaxstudios.com --tn 15 "15.Awareness.8k.wav" "15.Awareness.mp3"

I get bits.txt with 2537 frames
I get pcm.txt with 2147 frames.

The encoder reports 2149 frames.

Is there any way to change where I print "\n" to file to ensure that 
whole frames are written per row in the bitsream (bits.txt) file ?

I think they should both have around 2149 frames ... the PCM data is 
close, but the bitstream data is far off.

thanks
Matt