FluidVoice-Labs/fluid-speech-mini-gguf

Fluid Speech Mini (GGUF)

An English speech-to-text model built for dictation on phones and laptops. It matches the accuracy of Moondream's Parakeet Ultra (a 1.26 GB download) on the Open ASR Leaderboard's public English test sets, and was trained to keep transcribing the right person when someone else is talking in the room.

Fluid Speech Mini is the speech model inside FluidVoice. The iPhone app is in TestFlight: for access, send us a message on X at @FluidVoiceApp.

This is the GGUF build, for C++ apps on Windows and Linux: one 347 MB file that loads in the stock parakeet.cpp runtime, with about 485 MB of memory on a 15-second clip.

This is a port. Fluid Speech Mini was designed and trained for Apple devices, where Core ML stores its weights at about 3 bits each in 253 MB. ggml's packed formats carry a little more per weight (2.6 and 4.5 bits), so this port is somewhat larger; accuracy is unchanged.

All builds

Build	For	Download	Memory, 15 s clip	Repository
Core ML	iPhone, Mac	253 MB	about 277 MB	FluidVoice-Labs/fluid-speech-mini-coreml
ONNX	Windows, Linux, CPU or NVIDIA GPU	302 MB	about 500 MB	FluidVoice-Labs/fluid-speech-mini-onnx
GGUF	Windows, Linux, C++ apps	347 MB	about 485 MB	FluidVoice-Labs/fluid-speech-mini-gguf

Numbers

Word error rate (WER) in percent; lower is better. All rows are our own runs with the Open ASR Leaderboard's evaluation code on identical clips, not entries on the public board.

Model	Download	Average, 8 public English sets	Another voice in the room
Fluid Speech Mini (Core ML build; this build: 4.53 and 11.1)	253 MB	4.54	10.9
Parakeet Ultra (Moondream)	1.26 GB	4.56	14.3
Parakeet TDT 0.6B v3 (NVIDIA)	2.51 GB	4.86	20.4
Phonon-2 (Fermion)	164 MB	5.13	24.6
Parakeet Redux (Moondream)	178 MB	5.62	25.7
Whisper large-v3	about 3 GB	5.77	11.8

Set by set: AMI 8.34, Earnings-22 6.29, GigaSpeech 7.67, LibriSpeech clean 1.45, LibriSpeech other 3.25, SPGISpeech 3.56, Monsoon 3.96, VoxPopuli 1.76.

"Another voice in the room" is Google SVQ (English, 28,004 phone recordings) in its background-speech condition. On the same clips with traffic noise, the larger Parakeet models and Whisper large-v3 score a couple of points better than this model; its advantage is other voices.

This build

This build holds the same weights as the ONNX build: fine-tuned from the Core ML model so that every packed row stores exactly in this format. Those weights score 4.53 % on the same eight test sets (Core ML: 4.54 %) and 11.1 % with another voice in the room (Core ML: 10.9 %), measured in PyTorch. Through parakeet.cpp itself we measured 1.43 % on a 266-clip sample of LibriSpeech test-clean, against 1.41 % for the PyTorch reference on the same clips.

The encoder's weights are stored as ggml Q2_K and Q4_K tensors and multiplied in that packed form on CPU and on CUDA.

Use it

hf download FluidVoice-Labs/fluid-speech-mini-gguf --local-dir fluid-speech-mini-gguf
parakeet-cli transcribe --model fluid-speech-mini-gguf/fluid-speech-mini-q2k-q4k.gguf --input recording.wav --threads 4

Build parakeet-cli from parakeet.cpp; the file loads in the stock runtime, on CPU and on CUDA. Tested with commit 5c9b0fb. Pass --threads explicitly (4 is a good default); the runtime's default thread count can be slow on a busy machine.

Limitations

• English only. The vocabulary is Parakeet v3's 25-language one, so other languages may come out, but they were not trained or measured and are expected to be worse.
• Its advantage in noise is specifically other voices; with traffic noise, larger models do better.
• Scores above are self-reported runs of the leaderboard's code, not entries on the public board.

License and attribution

CC BY 4.0. If you use the model, please credit it as Fluid Speech Mini by FluidVoice.

Built on NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra, both CC BY 4.0. The decoder and joint network are Parakeet Ultra's.