| # TODO | |
| - FastAPI | |
| - Generate: | |
| - other params? (temperature, rvc f0_method, min_eos_p) | |
| - Bark TTS | |
| - Set speech sentiment in request body, then prefix each sentence e.g. `[Happy]<input_text>` | |
| - Should I aim to combine sentences which will fit in the largest clip length? (14s?) | |
| - More consistent tone etc in bark output? | |
| - Should map the rvc model to a chosen bark voice (incl. default)? | |
| - Or set via request body? | |
| - Currently hardcoded for bark voice `v2/en_speaker_9` (works well with all tested RVC models, regardless of gender etc) | |
| - RVC | |
| - Should I re-use the Config details? (GPU info etc) | |
| - Set `CUDA_VISIBLE_DEVICES` for bark | |
| - Should I load hubert model for each request? Precious VRAM | |
| - Smaller container image | |
| - Currently used to confirm app dependencies only, don't care that it's 6 GiB | |
| # Issues | |
| - Splits on sentences. So a single sentence which takes longer than ~14 seconds will be a mess | |
| - Significantly slower bark generation when ran via API vs directly in python script | |
| - Generation time roughly equals audio length (tested on 3090) |