🎙️ Fluid Speech Pro (MLX)

| | |

Fluid Speech Pro for Apple Silicon, running on parakeet-mlx. Speech-to-text for English and 24 other European languages, with punctuation and capitals.

All versions

Version	Best for	Clean mode
NeMo + 🤗 Transformers	NVIDIA GPUs and servers; the reference model	Yes (NeMo)
ONNX	CPUs, with onnx-asr or sherpa-onnx	Yes
MLX (this one)	Apple Silicon Macs, with parakeet-mlx	Yes
GGUF	C++ with parakeet.cpp, CPU or GPU, no Python	No
Core ML	Apple apps in Swift, on the Neural Engine, with FluidAudio	Not yet

Description

Fluid Speech Pro is a 0.63-billion-parameter speech recognition model for fast, accurate speech-to-text, from the team behind FluidVoice. It is built on NVIDIA's Parakeet TDT 0.6B v3 and keeps its 25 European languages. It detects the language on its own, with no prompt needed.

It is the #2 open model on the Open ASR Leaderboard, with a 6.06 average word error rate. It is the most accurate open model under 1B parameters, #2 among open models on noisy speech, and it runs at about 3,700× real time on one GPU.

This model is ready for commercial and non-commercial use.

Key features

Accurate: #2 open model on the Open ASR Leaderboard (6.06 average WER)
Best under 1B: 10% fewer errors than the next open model under 1B parameters
Strong in noise: #2 open model on URGENT 2024, the leaderboard's noisy-speech set (9.83 WER)
Fast: about 3,700× real time on one GPU, so an hour of audio takes about a second
Punctuation and capitalization included
Word- and segment-level timestamps (NeMo)
Clean mode (optional): edited text without repeated words and filler words (NeMo, ONNX and MLX)
25 European languages, detected automatically
Five versions: NeMo and 🤗 Transformers, ONNX, MLX, GGUF and Core ML
Permissive license: CC BY 4.0

Performance

Figure 1: Average word error rate (up is better) against model size for open models on the Open ASR Leaderboard. Fluid Speech Pro is the only top-5 open model under 1B parameters.

Figure 2: The best open models at 1B parameters and under, ranked by average word error rate (lower is better).

Figure 3: Word error rate on URGENT 2024, the leaderboard's noisy-speech set (lower is better).

Quick start

pip install parakeet-mlx          # reads audio through ffmpeg: brew install ffmpeg
hf download FluidVoice-Labs/fluid-speech-pro-mlx fluid_speech_mlx.py --local-dir .

from fluid_speech_mlx import load

model = load("FluidVoice-Labs/fluid-speech-pro-mlx", bits=8)
print(model.transcribe("audio.wav").text)

Leave out bits=8 for the full weights.

Clean mode

Clean mode drops filler words ("um", "uh") and repeated words ("I I think"). It's a small add-on (55 MB) that the loader applies on top of either set of weights:

clean = load("FluidVoice-Labs/fluid-speech-pro-mlx", bits=8, clean_mode=True)

Why the small loader

Fluid Speech Pro is Parakeet TDT 0.6B v3 plus one small block between the encoder's subsampling and its first Conformer layer. parakeet-mlx doesn't know that block, so a plain parakeet_mlx.from_pretrained refuses these weights instead of running without it. fluid_speech_mlx.py (one file) adds the block, and computes the audio features exactly the way NeMo does. With parakeet-mlx's own features, 1 in 4 transcripts came out different in our check.

Everything else is parakeet-mlx: transcribe returns its usual result with sentence and token timestamps, and chunk_duration works for long files.

Files

File	What	Download
`model.8bit.safetensors`	8-bit weights	751 MB
`model.safetensors`	fp32 weights	2.52 GB
`clean-mode.safetensors`	clean mode add-on, for either	55 MB
`config.json`, `fluid_speech_mlx.py`

load() downloads only the files it needs. Compute runs in bfloat16 by default; pass dtype=mx.float32 for full precision.

Accuracy of this version

400 clips from 8 public test sets (AMI, Earnings-22, GigaSpeech, LibriSpeech clean and other, Monsoon, SPGISpeech, VoxPopuli), compared with the NeMo model. All runs decode greedily.

Run	Same text as NeMo	WER
NeMo (reference)	400 / 400	3.53
MLX, fp32	398 / 400	3.53
MLX, default (bfloat16)	390 / 400	3.50
MLX, 8-bit	354 / 400	3.56

Clean mode against NeMo's clean mode (WER 3.98): fp32 394 / 400 the same, WER 4.00; 8-bit 341 / 400, WER 3.97.

WER is over the 350 clips from the 7 sets with per-clip references (Earnings-22's references are whole calls).

Good to know

Speed on an Apple M5 Max, one clip at a time: fp32 about 128× real time, 8-bit about 109×.
Offline transcription only: the extra block looks a few frames ahead, so transcribe_stream isn't supported.
The reference checkpoint is the NeMo version.

Use case

For developers, researchers and companies building anything that turns speech into text: dictation, meeting notes, captions and subtitles, voice assistants, call and media analytics, and transcription services.

Release date

Hugging Face, November 2026.

Model architecture

Architecture type: FastConformer-TDT

Network architecture:

A FastConformer encoder and a TDT (token-and-duration transducer) decoder, as in Parakeet TDT 0.6B v3
One small extra block between the encoder's subsampling and its first Conformer layer
0.63 billion parameters

Input

Input type: audio, 16 kHz, mono
Input format: .wav, .flac and other formats your audio loader reads

Output

Output type: text, with punctuation and capitalization
Optional: word and segment timestamps (NeMo)

Model version

Fluid Speech Pro 1.0

Evaluation datasets

The Open ASR Leaderboard's public test sets: AMI (cleaned), Earnings-22 (cleaned, chunked), GigaSpeech (cleaned), SPGISpeech, URGENT 2024 (noisy speech), Voice Arena Monsoon (Indian English) and VoxPopuli (cleaned).

Results

The leaderboard numbers below are for the main NeMo model with 4-beam search. This MLX version runs the same model, extra block included; see Accuracy of this version for how closely it matches.

Open ASR Leaderboard

Open models on the Open ASR Leaderboard: average word error rate over its seven public and two private test sets (lower is better):

Model	Parameters	Average WER
Musci-ASR	2B	5.70
Fluid Speech Pro	0.63B	6.06
NVIDIA Nemotron 3.5 Transcribe	30B	6.13
Qwen3-ASR	1.7B	6.15
IBM Granite Speech 5.0	2B	6.68
Cohere Transcribe	2B	6.72
NVIDIA Parakeet TDT 0.6B v2	0.6B	6.93
NVIDIA Parakeet TDT 0.6B v3	0.6B	7.04
Fermion Research Phonon-2	0.6B	7.86

Results by test set

Word error rate (%) on the leaderboard's seven public test sets:

Model	Public average	AMI	Earnings-22	GigaSpeech	SPGISpeech	URGENT 2024	Monsoon	VoxPopuli
Musci-ASR	5.15	6.53	5.83	6.41	1.86	10.49	3.44	1.47
Fluid Speech Pro	5.43	7.55	5.42	7.51	2.31*	9.83	3.71	1.70
NVIDIA Nemotron 3.5 Transcribe	5.51	7.78	5.25	7.18	1.91	10.94	3.63	1.89
Qwen3-ASR	5.75	8.31	5.84	7.22	2.57	9.93	3.50	2.85
IBM Granite Speech 5.0	6.05	6.96	6.22	8.14	3.36	10.22	3.94	3.52
Cohere Transcribe	6.26	7.02	7.88	7.88	2.74	9.50	3.44	5.38
NVIDIA Parakeet TDT 0.6B v2	6.43	9.09	6.76	8.15	1.94	11.42	3.89	3.78
NVIDIA Parakeet TDT 0.6B v3	6.51	9.42	5.85	7.99	3.63	11.52	4.14	3.05
Fermion Research Phonon-2	7.23	9.32	6.96	8.35	3.70	14.90	4.91	2.46

* In clean mode, which matches SPGISpeech's edited references; the leaderboard allows one transcription mode per test set.

Other languages

Fluid Speech Pro supports 25 languages. On the Open ASR Leaderboard's multilingual track it makes 12% fewer errors than Parakeet TDT 0.6B v3, and it is better in all seven of the track's languages:

Language	Parakeet TDT 0.6B v3	Fluid Speech Pro	Fewer errors
German	4.12	3.76	9%
French	5.40	5.10	6%
Italian	4.73	4.16	12%
Spanish	3.71	3.38	9%
Portuguese	6.08	4.88	20%
Dutch	7.53	6.82	9%
English	7.04	6.06	14%
Average	5.52	4.88	12%

Word error rate (%) on the Open ASR Leaderboard's multilingual track, lower is better.

It also got better than Parakeet TDT 0.6B v3 in every one of the 24 languages beyond English. On FLEURS, scored the way the leaderboard scores its multilingual track, it makes nearly 20% fewer errors on average.

Figure 4: Word error rate on FLEURS for Parakeet TDT 0.6B v3 (grey) and Fluid Speech Pro (blue), lower is better.

Language	Parakeet TDT 0.6B v3	Fluid Speech Pro	Fewer errors
Estonian	13.28	9.51	28%
Latvian	21.76	16.39	25%
Italian	2.48	1.91	23%
Lithuanian	21.04	16.21	23%
Maltese	19.23	15.05	22%
Hungarian	13.58	10.65	22%
Slovak	9.18	7.20	22%
Finnish	11.19	8.82	21%
Czech	11.06	8.83	20%
Portuguese	4.47	3.57	20%
Slovenian	22.48	17.97	20%
Ukrainian	5.95	4.77	20%
Greek	21.16	17.10	19%
Romanian	11.38	9.28	18%
Spanish	3.22	2.66	17%
Swedish	13.39	11.10	17%
Polish	6.84	5.70	17%
German	4.18	3.52	16%
Danish	17.20	14.57	15%
Bulgarian	11.94	10.19	15%
Croatian	11.40	9.73	15%
Dutch	6.39	5.51	14%
Russian	4.92	4.35	12%
French	4.75	4.65	2%
Average of all 24	11.35	9.13	20%

Word error rate (%) on FLEURS test sets, greedy decoding. A future version will be trained on more multilingual data.

Limitations

Please check the model on your own audio, accents and languages before you rely on it.
Like every speech model, it makes more mistakes on heavy noise, overlapping speakers, strong accents it has rarely heard, and rare names and terms.
Clean mode edits the text. Don't use it where you need an exact record.

License

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

Built on NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra, both CC BY 4.0.

Citation

@misc{fluidspeechpro2026,
  title  = {Fluid Speech Pro},
  author = {{FluidVoice}},
  year   = {2026},
  url    = {https://huggingface.co/FluidVoice-Labs/fluid-speech-pro}
}