TalkMate/fluid-speech-mini-gguf

Fluid Speech Mini (transcribe.cpp GGUF)

Fluid Speech Mini by FluidVoice. Built on NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra.

This is FluidVoice's GGUF build of Fluid Speech Mini, converted so it loads in transcribe.cpp. TalkMate downloads it on Windows and Linux. It is English only.

Changes from FluidVoice's file

• Tensors and metadata renamed to transcribe.cpp's Parakeet layout. The tokenizer is Parakeet TDT 0.6B v3's, which Fluid Speech Mini shares.
• The 2-bit (Q2_K) encoder weights are stored as Q4_K, because transcribe.cpp does not load Q2_K. Each weight is within 0.38 of a 2-bit step of FluidVoice's value; the root-mean-square difference is 0.16% of the weights' size.
• The pointwise convolutions are stored as F16, because transcribe.cpp's convolutions take only F16 or F32.
• Each LSTM layer's two bias vectors are summed into one.

On 100 LibriSpeech test-clean clips, this file and FluidVoice's original gave the same transcripts.

The conversion script is scripts/fluid-speech-mini-gguf in TalkMate's repository.

File	Size	SHA-256
fluid-speech-mini-q4k.gguf	506,663,584 bytes	0bff22e9947f79cd1fe0d1a5f3a95fcca07ecf90a7e9cb56a09bc923b1d15297

License

CC BY 4.0, as for the original model; see LICENSE and NOTICE. Please credit it as Fluid Speech Mini by FluidVoice.