FluidVoice-Labs/fluid-speech-mini-coreml

Fluid Speech Mini (Core ML)

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 Core ML build, for iPhone and Mac: a 253 MB package that runs on the Apple Neural Engine.

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	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

The package accepts the same inputs and produces the same outputs as the parakeet-tdt-0.6b-v3-coreml package (16 kHz audio, 15-second windows, 8,192-piece vocabulary, blank = 8192), so it is a drop-in replacement wherever that runs. Requires iOS 18 / macOS 15 or later.

On a Mac: 25 ms per 15-second window on the Neural Engine, about 277 MB of memory, flat across an 84-minute recording, which transcribes in 12 seconds.

Use it

Swift, with FluidAudio

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

import FluidAudio

let models = try await AsrModels.load(
    from: URL(fileURLWithPath: "fluid-speech-mini-coreml"),
    version: .fluidParakeetMini
)
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); the upstream FluidInference/FluidAudio package also loads it as a Parakeet v3 model.

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. The audio preprocessor is from FluidInference's Parakeet v3 Core ML package, also CC BY 4.0.