FluidVoice-Labs/fluid-speech-pro-coreml

🎙️ Fluid Speech Pro (Core ML): Fast Multilingual Speech-to-Text on the Mac

| | | |

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	Apple Silicon Macs, with parakeet-mlx	Yes
GGUF	C++ with parakeet.cpp, CPU or GPU, no Python	No
Core ML (this one)	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, the most accurate open model under 1B parameters, and #2 among open models on noisy speech.

This is the Core ML build, for apps on Apple silicon Macs. It is the same model, small extra block included, with 8-bit weights: a 652 MB download that runs on the Apple Neural Engine and transcribes 16.6 minutes of audio in about 4 seconds.

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)
• 25 European languages, detected automatically
• Punctuation and capitalization included
• On-device: runs on the Neural Engine, 652 MB download, about 250 MB of memory
• Fast: a 16.6-minute recording in about 4 seconds on a Mac
• 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).

The leaderboard figures are the NeMo model's results. How this Core ML build compares with it is under Accuracy of this build.

Quick start (Swift, FluidAudio)

hf download FluidVoice-Labs/fluid-speech-pro-coreml --local-dir fluid-speech-pro-coreml

import FluidAudio

let models = try await AsrModels.load(
    from: URL(fileURLWithPath: "fluid-speech-pro-coreml"),
    version: .fluidParakeetPro
)
let asr = AsrManager()
try await asr.initialize(models: models)
let result = try await asr.transcribe(URL(fileURLWithPath: "recording.wav"))
print(result.text)

A complete command-line example is in examples/swift/. FluidAudio is at https://github.com/altic-dev/FluidAudio (branch main).

Clean mode

Clean mode (it drops filler words like "um" and repeated words like "I I think") is coming to this build in a later update, as a switch in this same model. It is already available in the NeMo version.

Files

Files	Download
Preprocessor, Encoder, Decoder, JointDecision (.mlmodelc folders), vocabulary, config.json	652 MB

The encoder's weights are 8-bit (int8, one scale per output channel); the decoder and joint stay 16-bit. The audio preprocessor computes NeMo's features exactly, and the encoder expects its frame counts, so keep the files together.

Model 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

Accuracy of this build

The same 400 clips from 8 public English test sets (AMI, Earnings-22, GigaSpeech, LibriSpeech clean and other, Monsoon, SPGISpeech, VoxPopuli) as the ONNX version, compared with the NeMo model. Core ML runs go through FluidAudio on the Neural Engine. All runs decode greedily.

Run	Same text as NeMo	WER
NeMo (reference)	400 / 400	3.53
Core ML, 8-bit	335 / 400	3.43

• WER is over the 350 clips from the 7 sets with per-clip references (Earnings-22's references are whole calls).
• On clips up to 15 seconds, 319 of 345 come out word for word the same. FluidAudio reads audio in 15-second windows and splits longer clips, while NeMo reads each clip whole; that accounts for most of the other differences.
• The Neural Engine computes in 16-bit floats, so a few words also change on shorter clips.
• This build decodes greedily. The leaderboard numbers above use beam search with 4 beams, which makes about 0.6% fewer errors.

Languages

Fluid Speech Pro supports 25 languages: English, Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian. It got better than Parakeet TDT 0.6B v3 in every one of the 24 languages beyond English: on FLEURS, 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. Per-language numbers are on the main model page.

Speed and memory

On an Apple silicon Mac (M5 Max), with FluidAudio:

• 25 ms per 15-second window on the Neural Engine.
• A 16.6-minute recording in about 4.2 seconds, with about 250 MB peak process memory.
• The first load prepares the model for the Neural Engine and takes about 8 seconds; later loads take 0.2 seconds.

Good to know

• Input: 16 kHz mono audio. FluidAudio converts other formats.
• Needs macOS 15 or later on Apple silicon. Tested on Mac only so far.
• Same inputs and outputs as the parakeet-tdt-0.6b-v3-coreml package: 15-second windows, 8,192-piece vocabulary, blank = 8192.
• Word timestamps, beam search and clean mode are in the NeMo version.

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.

License and attribution

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}
}

The FluidVoice team