Repository navigation
LocalSoundsAPI v1.2.1 - Faster Fish Speech & Whisper on Any GPU
A maintenance release on top of v1.2 (concurrent engines, live progress bars, locked controls). Everything in v1.2 is included.
What's New
Fish Speech stays loaded between chunks
- Fish Speech now loads its models once when you click Load and keeps them in memory until Unload. Multi-chunk jobs no longer reload the models for every chunk, so long scripts finish much sooner.
- Load and Unload now do what they say for Fish: Load puts the models on the GPU, Unload frees the VRAM.
- If a chunk fails, Fish is reloaded and the chunk is retried automatically, as before. A retried chunk uses a new seed so it doesn't repeat the same result.
- Output is unchanged: the same Fish Speech code runs with the same settings.
Whisper works on any GPU
- Fixed transcription and Whisper verification failing with
Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)when Whisper was loaded on a GPU other than the first one.
Docs
- New README with a live demo GIF and a screenshot of every engine.
- The wiki is rebuilt: a step-by-step guide for each engine with screenshots and audio samples, plus Getting Started, Launcher, Concurrency & Progress, API Reference and Troubleshooting & FAQ.
Updating an Existing Checkout
git pull origin mainNo new dependencies. Restart LocalSoundsAPI after updating.
Fresh Install
Download and extract the Source code (zip) below, or clone this tag, then double-click install.bat:
git clone --branch v1.2.1 --depth 1 https://github.com/aivrar/LocalSoundsAPI.gitThe required portable-python-env-v1.7z and bin.zip assets remain in the v1.0 dependency bundle; the installer selects and verifies them automatically.