MiMo-Audio¶
MiMo-Audio supports offline text and speech generation on Apple Silicon using MLX. Both the Base and Instruct checkpoints share one implementation, with a separate audio tokenizer supporting all 20 residual codebooks.
| Model | Tasks |
|---|---|
| MiMo-Audio-7B-Instruct | English/Chinese TTS, voice/style prompting, ASR, audio understanding, spoken and text dialogue, multi-turn history |
| MiMo-Audio-7B-Base | Few-shot speech transformations and raw text/audio completion |
| MiMo-Audio-Tokenizer | 24 kHz mono audio encoding and reconstruction |
from mlx_audio.sts import load
from mlx_audio.audio_io import write
model = load("XiaomiMiMo/MiMo-Audio-7B-Instruct", strict=True)
result = model.generate(task="tts", text="Hello from MLX Audio!")
write("hello.wav", result.audio, result.sample_rate)
transcription = model.generate("speech.wav", task="asr")
print(transcription.text)
python -m mlx_audio.sts.generate \
--model XiaomiMiMo/MiMo-Audio-7B-Instruct \
--task spoken_dialogue --audio question.wav --output-path answer.wav
Generation is offline and processes one example per call. Streaming and server integration are deferred. The separate tokenizer loads lazily for audio tasks; use audio_tokenizer_path in generate() or --audio-tokenizer-path on the CLI for a local copy. Arrays require their sample rate. Output audio is 24 kHz mono.
See the complete model guide for task prompts, Base examples, result fields, and history, the codec guide for standalone use, and the conversion guide for quantization.