Reverb-TTS-bf16

An MLX Audio conversion of Software Mansion's Reverb TTS for Apple silicon. It generates 24 kHz mono speech and supports preset voices and voice cloning from a WAV file with its transcript.

The Reverb weights use BF16. The Reverb weights occupy 594.5 MB. The bundled Higgs Audio v2 codec retains its original FP32 weights and occupies 805.7 MB.

Usage

pip install mlx-audio

In the examples, use nativ-community/Reverb-TTS-bf16 after this bundle is published, or substitute the path to the local model directory.

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

model = load("nativ-community/Reverb-TTS-bf16")
result = next(model.generate("Hello from Reverb on Apple silicon.", lang_code="en"))
write("speech.wav", result.audio, result.sample_rate)
python -m mlx_audio.tts.generate \
  --model nativ-community/Reverb-TTS-bf16 \
  --text "Hello from Reverb on Apple silicon." \
  --voice standard_male \
  --lang_code en

The included presets are standard_male (default), angry_male, and sensual_female. For voice cloning, pass ref_audio="reference.wav" and ref_text="The reference transcript." to model.generate().

Licenses and attribution

Software Mansion's Reverb weights are licensed under CC BY-NC 4.0. This conversion repackages the weights for MLX. See the upstream model card for acknowledgments.

Built with Higgs Materials licensed from Boson AI USA, Inc., Copyright Boson AI USA, Inc., All Rights Reserved and Meta Llama 3 licensed under the Meta Llama 3 Community License, Copyright Meta Platforms, Inc., All Right Reserved. Built with Meta Llama 3.

The bundled codec weights are byte-identical to Boson AI's Higgs Audio v2 tokenizer and were sourced from the Transformers port. See the Boson Higgs Audio 2 Community License, Meta Llama 3 Community License, and NOTICE for the codec.

Downloads last month
27
Safetensors
Model size
0.3B params
Tensor type
BF16
ยท
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for nativ-community/Reverb-TTS-bf16

Finetuned
Qwen/Qwen3-0.6B
Finetuned
k2-fsa/OmniVoice
Finetuned
(2)
this model