Skip to content

Speech-to-Text (STT) Models

MLX Audio provides a range of speech-to-text models optimized for Apple Silicon, from lightweight English-only models to large multilingual systems with translation capabilities.

Model Comparison

Model Provider Parameters Languages Streaming Timestamps Repo
Whisper OpenAI Various 99+ -- Segment + Word mlx-community/whisper-large-v3-turbo-asr-fp16
Distil-Whisper HuggingFace Various EN -- Segment distil-whisper/distil-large-v3
Parakeet NVIDIA 0.6B EN (v2), 25 EU (v3) Yes Sentence + Word mlx-community/parakeet-tdt-0.6b-v3
Voxtral Realtime Mistral 4B Multiple Yes -- 4bit, fp16
Qwen3-ASR Alibaba 0.6B / 1.7B ZH, EN, JA, KO + more Yes Segment mlx-community/Qwen3-ASR-1.7B-8bit
Qwen3-ForcedAligner Alibaba 0.6B ZH, EN, JA, KO + more -- Word-level mlx-community/Qwen3-ForcedAligner-0.6B-8bit
VibeVoice-ASR Microsoft 9B Multiple Yes Segment mlx-community/VibeVoice-ASR-bf16
Voxtral Mistral 3B Multiple -- -- mlx-community/Voxtral-Mini-3B-2507-bf16
Cohere Transcribe Cohere 2B 14 languages -- Segment CohereLabs/cohere-transcribe-03-2026
Qwen2-Audio Alibaba 7B Multiple -- -- mlx-community/Qwen2-Audio-7B-Instruct-4bit
Canary NVIDIA ~1B 25 EU + RU, UK -- -- README
SenseVoice Alibaba DAMO ~234M 50+ -- -- mlx-community/SenseVoiceSmall
FireRedASR2 Xiaohongshu ~1.18B ZH, EN -- -- mlx-community/FireRedASR2-AED-mlx
Granite Speech IBM ~1B EN, FR, DE, ES, PT, JA Yes -- README
Moonshine Useful Sensors 27M / 61M EN -- -- README
MMS Meta 1B 1000+ -- -- README

Cohere quantized local checkpoints

Cohere Transcribe can be loaded from local MLX-converted checkpoints, including 8-bit and 4-bit variants. To generate these variants explicitly, use --q-bits 8 or --q-bits 4 together with --quantize.

  • 8-bit: python -m mlx_audio.convert --model-domain stt --quantize --q-bits 8
  • 4-bit: python -m mlx_audio.convert --model-domain stt --quantize --q-bits 4

You can also pass --q-group-size to control the quantization group size when needed.

Unified API

All STT models share the same loading interface:

from mlx_audio.stt import load

model = load("mlx-community/whisper-large-v3-turbo-asr-fp16")
result = model.generate("audio.wav")
print(result.text)

mlx_audio.stt.generate \
  --model mlx-community/whisper-large-v3-turbo-asr-fp16 \
  --audio audio.wav \
  --verbose

Choosing a model

  • Best multilingual coverage: Whisper (99+ languages) or MMS (1000+ languages)
  • Best accuracy for English: Parakeet v2 or Whisper large-v3-turbo
  • Best for European languages: Parakeet v3 (25 languages) or Canary
  • Lowest latency / streaming: Voxtral Realtime (4bit variant)
  • Smallest footprint: Moonshine tiny (27M parameters)
  • Speaker diarization built-in: VibeVoice-ASR
  • Word-level alignment: Qwen3-ForcedAligner
  • Emotion / event detection: SenseVoice