Skip to content

Audio8 TTS

Audio8-TTS-Realistic-0.6b generates mono 44.1 kHz speech in 23 languages on Apple Silicon, with optional zero-shot voice cloning.

Ready-to-use MLX checkpoints include the codec and require no PyTorch:

Precision Repository
BF16 nativ-community/Audio8-TTS-Realistic-0.6b-bf16
4-bit nativ-community/Audio8-TTS-Realistic-0.6b-4bit
8-bit nativ-community/Audio8-TTS-Realistic-0.6b-8bit

Use any of these repository IDs in the examples below.

Generate speech

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

model = load("nativ-community/Audio8-TTS-Realistic-0.6b-bf16")
result = next(model.generate(
    text="Welcome to Audio8, running on Apple Silicon.",
    temperature=0.8,
    top_p=0.95,
    max_tokens=1024,
))
write("speech.wav", result.audio, result.sample_rate)

python -m mlx_audio.tts.generate \
  --model nativ-community/Audio8-TTS-Realistic-0.6b-bf16 \
  --text "Welcome to Audio8, running on Apple Silicon." \
  --temperature 0.8 --top_p 0.95 --max_tokens 1024

Clone a voice

result = next(model.generate(
    text="This sentence uses the reference speaker.",
    ref_audio="reference.wav",
    ref_text="The exact words spoken in the reference recording.",
))
write("cloned.wav", result.audio, result.sample_rate)

For the CLI, add --ref_audio reference.wav --ref_text "The exact reference transcript.". Reference paths are resampled automatically; Python audio arrays must already be mono at 44.1 kHz. Language follows the input text. Realistic supports Arabic, Cantonese, Chinese, Dutch, English, Finnish, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Thai, Turkish, and Vietnamese. Quality varies by language.

Generation produces one complete waveform per call. Text tokens, reference audio frames, and generated frames share a 2,048-position context budget. max_tokens limits audio frames, each approximately 46 milliseconds long. Long input text or reference recordings leave less room for output; split long passages into shorter utterances. See the model README for generation options and the supported Preview models.

Realistic-0.6b is released under Apache-2.0.