mlx-community/Zonos-v0.1-transformer-bf16

Zonos-v0.1-transformer (MLX, bf16)

Apple-Silicon MLX port of Zyphra/Zonos-v0.1-transformer — an autoregressive transformer TTS over DAC codec tokens at 44.1 kHz, with zero-shot voice cloning, an 8-D emotion vector, and prosody controls (pitch, fmax, speaking-rate). Self-contained: no PyTorch, no multi-repo fetch.

Code: https://github.com/xocialize/zonos-mlx

Verification

Every component is numerically parity-tested against the PyTorch reference (CPU fp32):

Component	Metric
AR backbone (GQA transformer)	rel err ~7e-7
Conditioning + espeak G2P	rel err ~2e-7
DAC decode	SI-SDR 108 dB
DAC encode	exact integer code match
Speaker encoder (ResNet293)	rel err ~1.2e-5
Duration (speaking_rate)	monotonic
Emotion (8-D vector)	content + energy change, speaker stable

Usage

import mlx.core as mx, soundfile as sf
from zonos_mlx.pipeline_mlx import ZonosPipeline

tts = ZonosPipeline.from_pretrained_mlx("mlx-community/Zonos-v0.1-transformer-bf16")

# voice cloning from a 16 kHz mono reference clip:
ref = mx.array(...)                       # (N,) float32 @ 16 kHz
spk = tts.make_speaker_embedding(ref)
wav = tts.generate("Hello from MLX.", speaker=spk, speaking_rate=15.0,
                   emotion=[0.0,0,0,0,0,0,0.1,0.9])  # happy..neutral 8-D
sf.write("out.wav", mx.array(wav)[0], tts.sampling_rate)

License & attribution

Apache-2.0, inherited from the upstream Zonos model and code (Zyphra). Bundled third-party components, all redistributable:

• DAC (Descript Audio Codec) — descript/dac_44khz, via HF transformers DacModel.
• Speaker embedding — ResNet293 (SimAM/ASP) + LDA from Zyphra/Zonos-v0.1-speaker-embedding.
• espeak-ng phonemizer frontend (GPL-3.0 tool, invoked at runtime — not redistributed here).

Please cite Zyphra's Zonos for the model weights and architecture.