[feature request]? runtime meda data

[landist]

Hi all, I am exciting to find audio.cpp! I am very new so forgive me if my question is already answered or silly.

The first time I generate music using yue2 and the server UI really blow my mind. The output *.wav file is really nice. I play quite a few songs by modifying things here and there. But I was in trouble to tell which output correspond to which input. My bad at keeping records and things in place.

I think the runtime meta data like the prompt, the style, the lyrics, seed, any options and model info etc for each output should be saved together (or at least a user option - a checkbox maybe) with the output. This would allow the reproduction of the output, comparison of outputs, and output from different models etc. I believe the image generation stable-diffusion webui or comfyui have this features, where the meta data is just appended at the end/beginning of the image. I understand audio maybe more difficult, but there must be a way to do it.

If it is already done, please let me know where to look as I am very new. Thank you very much.

[0xShug0]

@landist Thanks for your kind words! Enriching the UI is on the TODO list and is expected to be community-driven. I'm not really a UI person and have messed up the UI several times before 😅 . Also, if you start the server with --log, it will print extra information. I'm wondering if that might be useful.

[landist]

@0xShug0 Thanks a lot for the --log tip. It is useful.
But my question is rather about the option to save runtime metadata (about model used, prompt, flags, seed, lyrics etc) information into the output file e.g. wav. So that the output file is stand alone and contain enough information for reproducing with the same setting or comparison. e.g for tts output, that will contains the text for the outputted speech, voice reference file etc. It can be a user option flag.

It probably a minor thing comparing to including more new models, optimizing current models performance etc. But I think it will be very useful for users to keep track their outputs with input information. Image/video generation tool like stable diffusion or comfyui already has this feature - the images contain the metadata. llama.cpp with the build in UI server also have a list of history for past conversation/chats etc so that the same prompt or setup can be easily reuse to test different models. I think audio.cpp should have something similar as well.

[0xShug0]

@landist Generation export/import and preliminary (non-persistent) history support will be available in the next release!