Audio8-TTS-Realistic-0.6b-bf16 (MLX)

MLX conversion of Edge0/Audio8-TTS-Realistic-Air (revision 925bbd4) for Apple Silicon. Upstream published this model as Audio8-TTS-Realistic-0.6b and renamed it Realistic-Air on 2026-11-02; the weights are unchanged, so this conversion is the same model under its new name. It is a 0.6B-parameter multilingual text-to-speech model with zero-shot voice cloning and a bundled 44.1 kHz neural codec.

Realistic is an RL-tuned successor to Audio8-TTS-Preview-0.6b. The architecture, codec, tokenizer and tensor layout are the same; only the language-model weights differ. Anything that loads the Preview bundle loads this one.

• Language model: bf16 (as published upstream)
• Codec: fp32, converted from codec.pth to safetensors with weight-norm folded and conv weights in MLX (channels-last) layout. Byte-identical to the Preview bundle's codec.safetensors.
• Architecture: arktts, a DualAR transformer. A 24-layer slow AR predicts one semantic token per frame, and a 4-layer fast AR predicts 10 codec codebooks per frame.

Conversion parity vs the PyTorch reference (fp32, CPU, transformers 4.57.3):

• unit and block outputs within 1e-3
• reference-audio codec encoding is 100% code-exact
• greedy generation is token-exact over all 129 frames of the validation utterance
• decoded waveform max-abs 7.1e-6

Usage (mlx-audio)

Requires mlx-audio ≥ 0.4.7, the first release with the arktts model family.

pip install -U "mlx-audio>=0.4.7"

import soundfile as sf
from mlx_audio.tts.utils import load

model = load("mlx-community/Audio8-TTS-Realistic-0.6b-bf16")

# zero-shot voice cloning: the reference transcript must match the reference audio
for result in model.generate(
    text="Welcome to Audio8 TTS, running on Apple Silicon.",
    ref_audio="reference.wav",
    ref_text="Transcript of the reference clip.",
    temperature=0.7, top_p=0.9, top_k=50,
):
    sf.write("output.wav", result.audio, result.sample_rate)

Without ref_audio, the model synthesizes with a default voice.

For Swift, the mlx-audio8-tts-swift package loads this bundle as Audio8RealisticPackage.

Against the Preview checkpoint

These numbers come from our own A/B of this conversion against the Preview conversion. Both cloned two English reference voices, with paired seeds, 2 draws per line, scored by Whisper-large-v3-turbo and CAM++.

	Preview-0.6b	Realistic-0.6b
error rate, en / zh / de / es / fr / ja (macro)	0.0122	0.0126
speaker similarity (CAM++)	0.7212	0.7211
pause time per line	0.62 s	1.52 s

On those six languages the two models score the same on accuracy and voice similarity. Realistic pauses more: its takes run about 13% longer, and all of the extra time is silence.

Supported languages

• Upstream's original 11: Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Spanish.
• New in Realistic, measured at 1–7% error: Finnish, Indonesian, Malay, Portuguese, Russian, Turkish.
• Experimental: Hindi, Norwegian and Swedish work but are uneven, at 10–25% error.
• Not usable: Arabic, Thai and Vietnamese. Thai falls into repetition loops.

These figures come from three lines per language, two voices and two seeds. Most of the new-language lines were not checked by native speakers.

License

Apache-2.0, following the upstream model. The upstream fine-tunes the Apache-2.0 Audio8-TTS-Preview-0.6b, whose slow AR derives from Qwen2.5-0.5B (Apache-2.0). The bundled codec weights are the Preview's and carry the same license.