valtu4a/whisper-large-v3-russian-mlx

Whisper Large V3 — Russian (MLX)

Whisper Large V3 converted to MLX format for Apple Silicon (M1/M2/M3/M4).

This model is optimized for Russian speech transcription.

Usage

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="valtu4a/whisper-large-v3-russian-mlx",
    language="ru",
    word_timestamps=True,
)
print(result["text"])

Recommended Parameters

{
    "language": "ru",
    "word_timestamps": True,
    "condition_on_previous_text": False,
    "temperature": 0,
    "compression_ratio_threshold": 2.4,
    "logprob_threshold": -1.0,
    "no_speech_threshold": 0.6,
    "hallucination_silence_threshold": 2
}

Requirements

• Apple Silicon (M1 or newer)
• macOS 13.3+
• Python 3.9+

pip install mlx-whisper

Performance

Chip	RTF (Real Time Factor)
M1	~0.15
M2	~0.12
M3	~0.10

RTF < 1 means transcription runs faster than real time.

Base Model

Converted from antony66/whisper-large-v3-russian — a Whisper Large V3 fine-tuned on Russian speech data.

Acknowledgements

Special thanks to antony66 for fine-tuning and open-sourcing whisper-large-v3-russian, which served as the base for this MLX conversion.