Source commit: c9507ed
Adds Fluid Speech Pro, FluidVoice's fine-tune of Parakeet TDT 0.6B v3 and Parakeet Ultra, as an additional model (#780). It covers the same 25 languages, detects the language itself and runs on the CPU. Parakeet Ultra stays the recommended model.
On a Ryzen 7 7800X3D, Fluid Speech Pro had the lowest word error rate of the Parakeet models on 250 FLEURS clips in 25 languages (10.6 % against 14.2 % for Ultra and 19.9 % for v3) and needed 384 ms per clip at the median, against 496 ms for Ultra. The download is about 900 MB.
Changes
- Fluid Speech Pro downloads from
model-fluid-speech-pro-int8-v1and is checked against pinned SHA-256 hashes. With the GPU chosen, it runs on the CPU and says so. - Ultra and Fluid Speech Pro write
<unk>for characters Parakeet's vocabulary has no token for. The marker no longer reaches the text: it becomes "&" between capitals (M&A), the Greek final sigma "ς" at the end of a word, a space before a capital, or an apostrophe between lowercase letters, and is dropped elsewhere. This applies to every model of the plugin, on the CPU and the GPU.
Dictionary boosting, the German formatting default and the History label treat Fluid Speech Pro like the other Parakeet models from the next app update.
Fluid Speech Pro by FluidVoice. Built on NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra. CC BY 4.0.