Repository navigation
v0.8.3
FastConformer recognition is faster on Metal and Vulkan, with the same models and decoding options.
- Reduce repeated FFT allocations, cache fixed attention positions, and reuse RNNT/TDT decoder graphs.
- Improve ReazonSpeech greedy warm-request medians from 72 to 54 ms on M5 and 76 to 47 ms on RTX 2080; Parakeet v3 improves from 102 to 93 ms and 128 to 82 ms respectively. See measurement conditions and accuracy checks.
- Keep existing GGUF files compatible. There is no need to download or convert models again.
- Preserve ReazonSpeech's default beam search, C API 3.1, worker protocol 3, and the Metal tensor API workaround.
Installers update the engine when run again. Download the archive for your platform below; SHA256SUMS contains its checksum.
Changes: #91, #92. Full diff from v0.8.2.