No README information on how to compile fargan_demo

[rafael2k]

I'm trying to compile fargan_demo, but without success. Any instructions on how to build its code?

[rillian]

./configure --enable-deep-plc && make -j 8 should result in a ./fargan_demo.

If you need it for some other target, it links statically with the lpcnet and celt libraries.

[rafael2k]

Thanks a lot @rillian !
Input is pcm uint16 16 kHz samples right?

[rafael2k]

I managed to use, it works, I can encode and decode an audio, thanks! But the features file size is massive. Is FARGAN supposed to be a LPCNet substitute?

[rillian]

Yes, it's a more efficient (in time and quality) replacement. See https://arxiv.org/pdf/2405.21069

[rafael2k]

Thanks again @rillian. I took a look in the article, but I need more information about the actual encoding and decoding, and how to set the bitrate, if tunable. The is no MOS versus bitrate table / graph in the article. I'm a bit confused also as the file created by fargan_demo is much bigger than what it was supposed to be if the audio was compressed at around 1.something kbit/s.

LPCNet README suggests users are encouraged to switch to fargan, but there are no examples on how to achieve results analog to lpcnet_demo "-encode" and "-decode".

[rillian]

Sorry, I don't know much about it beyond what I've already said. Are you expecting fargan_demo to produce compressed audio? It looks to me like it just runs the feature extraction/sythesis component. The 'features' file is just just a f32 vector for each audio frame. You'd have to run the full opus encoder to produce compressed output at a target bitrate, or construct your own format if you want to code just the fargan output for efficient transmission.

[rafael2k]

Indeed, it is just the feature vector, without the actual entropy coder.
I'm playing with opus_demo, but I see no vestige of an option to enable fargan compression.