Ming-3.0 TTS Tiny

Native MLX support for mlx-community/Ming-3.0-tts-tiny-bf16, the 7.9B MoE language backbone with a flow-matching speech decoder. Output is 44.1 kHz mono audio. BF16, 8-bit and 4-bit MLX checkpoints use the same loader.

Precision	Repository
BF16	mlx-community/Ming-3.0-tts-tiny-bf16
8-bit	mlx-community/Ming-3.0-tts-tiny-8bit
4-bit	mlx-community/Ming-3.0-tts-tiny-4bit

Use any of these repositories with the Python and CLI examples below.

from mlx_audio.tts import load
from mlx_audio.audio_io import write

model = load("mlx-community/Ming-3.0-tts-tiny-bf16", strict=True)
result = next(model.generate(
    "The evening breeze moved through the trees and turned the clouds gold.",
    voice="US_James", seed=42,
))
write("speech.wav", result.audio, result.sample_rate)

python -m mlx_audio.tts.generate \
  --model mlx-community/Ming-3.0-tts-tiny-bf16 \
  --text "Hello from Ming three." --voice US_James

Other built-in voices include 灵小柔 and 灵楷. The multilingual cloning task supports Chinese, English, German, Spanish, French, Korean and Japanese. The remaining task controls follow the upstream training contract:

Task	Conditioning and controls
ip (default)	Built-in voice; empty speaker slot
seed / clone	ref_audio, ref_text, lang_code="en" or "zh"
multilingual	Reference audio/transcript and lang_code="de", "es", "fr", "ko" or "ja"
paralinguistic	Reference audio/transcript, EN or ZH, inline vocal-event tags in text
basic	Reference speaker and speed="慢速"/"中速"/"快速", pitch="低"/"中"/"高", volume="低"/"中"/"高"
emotion	Reference speaker and emotion="happy", "sad", "angry", "neutral", "surprised", "fearful" or "disgusted"
dialect	Reference speaker and dialect="cantonese" or "sichuanese"
voice_design	style description; empty speaker slot
freeform	Natural-language instruct; optional reference speaker
podcast	Ordered reference audio/transcript lists, text lines labeled speaker_1:, speaker_2:, etc.

Reference-speaker extraction runs natively in MLX. All listed checkpoints include the CAMPPlus weights in speaker_encoder/campplus.safetensors, so no separate model downloads are required. Neither ONNX nor ONNX Runtime is required. A precomputed spk_emb with 192 values per speaker can replace speaker extraction; cloning and podcast continuation still require the reference waveform and transcript. Array references use the model's 44.1 kHz sample rate; file paths are resampled automatically.

# Voice cloning, including acoustic continuation from the reference.
result = next(model.generate(
    "This sentence uses the reference speaker.",
    ref_audio="reference.wav", ref_text="Words spoken in the reference.",
    lang_code="en", seed=42,
))

# Freeform voice control.
result = next(model.generate(
    "Today feels like a wonderful day.",
    instruct="A cheerful young woman speaking clear English.", seed=42,
))

# Two-speaker continuation. Each speaker must appear in the target dialogue.
result = next(model.generate(
    " speaker_1:How are you?\n speaker_2:I am doing well.\n",
    task="podcast", ref_audio=["first.wav", "second.wav"],
    ref_text=["First reference transcript.", "Second reference transcript."],
    max_tokens=500,
))

Generation is offline and returns one audio result. The default limit is 250 acoustic steps (about 20 seconds); pass max_tokens or max_decode_steps for longer speech, within the 4096-position context budget. Text is passed through as supplied, so spell out numbers or abbreviations when a particular reading is required. Structured controls are mutually exclusive by task. Quantization covers the language backbone while preserving the acoustic modules and MoE routers. Runtime inference does not require PyTorch or FLA.