feat(server): add OpenAI-compatible endpoint #421
Conversation
-Add openai-compatible v1/audio/speech endpoint to server.py -Add lowvram command line argument to server.py (if running in cuda, switches model to cpu when idle) -Allow openai server to use built in speaker voices, cloning wavs or directory with wavs as the voice -Allow language_id as server.py command line argument in alternative to request parameter -split long input text using built in segmenter
There was a problem hiding this comment.
Cool, thank you! Can you share a link to the relevant OpenAI API spec so I can check everything works as expected? Could you also add some information and examples for this to https://github.com/idiap/coqui-ai-TTS/blob/dev/docs/source/server.md?
Yes here you go, I'm going to try to do extensive testing, though can confirm the code already works with a project that uses the openai TTS spec extensively (WingmanAI). I think the most testing is needed for models other than xtts2 as I have no idea how to use those other models coqui-tts supports. https://platform.openai.com/docs/guides/text-to-speech Also: |
-Eliminate low VRAM mode (results in approx. 1.5gb more VRAM use and creeps up over time versus lowVRAM mode variant) -eliminate models and voices endpoints since not a current part of OpenAI spec -eliminate split sentences code, relying on coqui tts' already utilized split_sentences functionality with api
Since not using lowvram mode, gc import now unnecessary
-add usage examples to server.md -fix bug with elif statement in last changes -default to using speaker_idx if any specified at server launch and no voice parameter is passed to openai server
|
OK I think this is ready for your testing. |
eginhard
changed the title
|
Thank you for merging! I'll try to turn to that underlying synthesizer issue to make sure proper segmenter is always used per language in the coming days. |
-Add openai-compatible v1/audio/speech endpoint to server.py
-Add lowvram command line argument to server.py (if running in cuda, switches model to cpu when idle)
-Allow openai server to use built in speaker voices, cloning wavs or directory with wavs as the voice
-Allow language_id as server.py command line argument in alternative to request parameter
-split long input text using built in segmenter (eg to get around hard input limits in xtts2)
-tested on windows only (I do not have mac or linux)
-tested with xtts2 but changes in theory should not impact other models