Original title: FluidVoice-Labs/fluid-speech-pro
Speech recognition model uploaded by the FluidVoice team on 10 30, based on NVIDIA Parakeet TDT 0.6B v3. The model card states it has 0.63B parameters and supports 25 European languages (including English, German, French, Spanish, Russian, Ukrainian, etc., but excluding Chinese). It automatically detects the language without prompts and outputs text with punctuation and capitalization. Five versions are provided: NeMo / Transformers, ONNX, MLX, GGUF, and Core ML. An optional "clean mode" removes repeated words and filler sounds like "um" and "uh." Licensed under CC BY 4.0.
Why it matters Small size, commercially usable license, and additional ONNX, MLX, GGUF, and Core ML versions allow it to run on CPUs and Macs, making it suitable for subtitles, meeting notes, and dictation
Rising score 46 · Likes 49 · Downloads 8951
Reasons
- The model card claims it ranks 2 among open models on the Open ASR Leaderboard, with an average word error rate of 6.06, making it the most accurate open model under 10 billion parameters; on the noisy speech dataset URGENT 2024, the word error rate is 9.83
- The authors claim approximately 3,700x real-time speed on a single GPU, transcribing one hour of audio in about one second; ONNX, MLX, and GGUF versions can run on CPUs and Apple silicon
- Compared to the base model Parakeet TDT 0.6B v3: It performs better across all 7 languages in the leaderboard's multilingual track, with an average reduction of 12% errors; on FLEURS, it reduces errors by approximately 20% on average for the 24 non-English languages
Usage
- Transformers 5.22+: pipeline("automatic-speech-recognition", model="FluidVoice-Labs/fluid-speech-pro"), pass 16 kHz mono audio
- For timestamps or clean mode, use NeMo: pip install "nemo_toolkit[asr]", download fluid_speech.py from this repository via hf download, then run fluid_speech.load("FluidVoice-Labs/fluid-speech-pro", map_location="cuda", clean_mode=True); transcribe([...], timestamps=True) returns word-level timing
- To decode similarly to the leaderboard: set decoding.strategy to malsd_batch and beam_size to 4 (authors claim an average reduction of approximately 0.6% errors); use fluid-speech-pro-onnx for CPU, fluid-speech-pro-mlx for Mac, and fluid-speech-pro-gguf for C++
Risks
- Does not support Chinese; rankings and speed are data provided by the authors in the model card, and the authors also advise testing with your own audio, accents, and languages first
- The Transformers version removes a small module and uses only greedy decoding; authors claim results are nearly identical to the NeMo version; the GGUF version lacks clean mode, and the Core ML version is not yet supported
- Clean mode rewrites the original speech; do not enable it if verbatim transcription is required; use under CC BY 4.0 requires attribution to "Fluid Speech Pro by FluidVoice." This entry has not been run