Introduction

FluidVoice, the team behind the open-source FluidVoice dictation app, has released Fluid Speech Pro. It's an open-weight speech recognition (ASR) model with 0.63 billion parameters, licensed CC BY 4.0 and free for commercial use. The model card was finished on Hugging Face on November 13, with Transformers weights and the full results. The

followed at 02:24 Malaysia time on November 14 (18:24 UTC on November 13).

The headline is size versus accuracy. Fluid Speech Pro averages 6.06% word error rate on Hugging Face's Open ASR Leaderboard. FluidVoice says that put it #2 among open models at launch, behind only a 2B model, and ahead of NVIDIA's 30B Nemotron 3.5 Transcribe. Eric Bezzam from Hugging Face's audio team

that it's one of three open models to pass Nemotron 3.5 Transcribe in English ASR in the past two weeks.

What Shipped

  • Model: Fluid Speech Pro 1.0, from FluidVoice Labs (blog post).
  • Base: fine-tuned from NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra, both CC BY 4.0. It keeps Parakeet's FastConformer encoder and TDT (token-and-duration transducer) decoder, and adds one small extra block early in the encoder.
  • Languages: 25 European languages, detected automatically with no prompt: English, Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian.
  • Input and output: 16 kHz mono audio in. Text with punctuation and capitalization out. Word- and segment-level timestamps are available with NeMo.
  • Clean mode: an optional switch that drops filler words ("um", "uh") and repeated words for notes, captions and dictation. It's off by default, so you get a word-for-word transcript unless you ask. It works in the NeMo, ONNX and MLX versions.
  • Speed: about 3,690× real time on one GPU, per FluidVoice. That's roughly an hour of audio in a second.
  • License: CC BY 4.0. Credit it as "Fluid Speech Pro by FluidVoice".

The Numbers

All figures are FluidVoice's own, from the model card and blog, measured on the Open ASR Leaderboard's test sets. Word error rate (WER): lower is better.

  • Average WER, all nine leaderboard sets: 6.06%. The next open models listed are Nemotron 3.5 Transcribe (30B) at 6.13% and Qwen3-ASR (1.7B) at 6.15%. Musci-ASR (2B) leads at 5.70%.
  • With paid models included: #10 of 66 at launch, per the blog.
  • Under 1B parameters: the most accurate open model, with 10% fewer errors than the next one.
  • Noisy speech (URGENT 2024): 9.83% WER. FluidVoice says that's #2 among open models, and the only open model under 1B below 10%.
  • Per test set: VoxPopuli 1.70%, SPGISpeech 2.31% (in clean mode), Voice Arena Monsoon (Indian English) 3.71%, Earnings-22 5.42%, GigaSpeech 7.51%, AMI 7.55%.
  • Versus its Parakeet base: 12% fewer errors on the leaderboard's multilingual track, and about 20% fewer on average across the 24 non-English languages on FLEURS. The biggest gains are in Estonian (28%), Latvian (25%), Italian and Lithuanian (23% each).

Where to Run It Today

  • Transformers (5.22 or newer): the quickest way to try it.

from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="FluidVoice-Labs/fluid-speech-pro")
print(asr("audio.wav")["text"])

  • NVIDIA NeMo: the reference checkpoint, loaded through fluid_speech.py from the repo. Use it for timestamps, clean mode (clean_mode=True) and the leaderboard-style 4-beam decoding.
  • ONNX: for CPUs, with onnx-asr or sherpa-onnx. Full and int8 builds.
  • MLX: for Apple Silicon Macs, with parakeet-mlx. Full and 8-bit builds.
  • GGUF: for parakeet.cpp, on CPU or GPU, with no Python. Full and 8-bit builds.
  • Core ML: for Swift apps on Apple's Neural Engine, through FluidAudio.
  • No code: FluidVoice says the model is already live in its free FluidVoice dictation app.

Caveats

  • European languages only. All 25 languages are European. There's no Chinese, Japanese, Korean, Arabic, Hindi or Malay. For those, look at Qwen3-ASR or the Nemotron models.
  • Self-reported rankings. The rank and WER figures come from FluidVoice's submission. Leaderboards move fast: Bezzam counted three new open models passing Nemotron 3.5 in two weeks.
  • Transformers is slightly simplified. The Transformers weights leave out the extra block and decode greedily. FluidVoice says they give nearly the same results, but timestamps and clean mode need NeMo, ONNX or MLX.
  • Clean mode edits your words. Leave it off for legal, medical or any word-for-word record.
  • Early checkpoint already on HF. A short-card version of the repo has been public since October 30, and some community ports are based on it. The full 1.0 release with Transformers weights and the official ONNX, MLX, GGUF and Core ML builds landed on November 13.
  • The usual ASR weak spots. The card warns that heavy noise, overlapping speakers, unfamiliar accents and rare names still cause more errors. Test it on your own audio first.

Conclusion

Fluid Speech Pro is a small model with big-model accuracy. At 0.6B parameters it sits next to 2B and 30B models on the Open ASR Leaderboard. It's commercially usable under CC BY 4.0, and it ships in every runtime you'd want, from Transformers on a GPU to GGUF on a CPU to Core ML on Apple devices. If you transcribe podcasts, cut captions or build local dictation in a European language, it's a strong default to test against Parakeet and Whisper this weekend.

—Anabel ♡