You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

Reverb-TTS: zero-shot voice cloning with open weights in 7 languages. 297M parameters; 164.9x faster than real time on an RTX 5090 at batch 32, 21.6x on an M4 Pro and 20.6x on an iPhone 17; 1.35% word error rate on MiniMax. Software Mansion AI Labs.

Reverb-TTS

GitHub · Demo · Quick start · Benchmarks

Reverb is a text-to-speech model designed for high-quality voice-cloning. Give it a few seconds of someone speaking and their transcript, and it speaks any text in that voice, with the speaker's timbre, prosody, and emotion. It's built on one of the leading open-weight TTS models.

  • Clones a voice from 3 to 20 seconds of clean speech, with no fine-tuning.
  • Seven languages: English, German, French, Spanish, Italian, Polish, and Portuguese, and a voice can speak any of them.
  • The lowest word error rate of the open-weight models we compared (1.35%).
  • Speed: 164.9× real time on an RTX 5090 at batch 32, over 20.0× on an Apple M4 Pro and iPhone 17 at batch 1.
  • On-device formats: int8 and int4 weights, and ExecuTorch programs for Apple GPUs (MLX), Android GPUs (Vulkan), and Qualcomm NPUs (QNN).

What is Reverb?

Reverb starts from OmniVoice, a 0.6B masked-diffusion TTS model that fills in a grid of audio-codec tokens over several passes, conditioned on the text and on a reference speaker's audio. We made Reverb from it in three stages:

  1. Pruning with distillation. 4 of the 28 transformer layers are removed, the feed-forward layers shrink from 3072 to 2048, KV groups from 8 to 6, and the text vocabulary from 151,676 to 62,035 tokens. The pruned model learns to match OmniVoice's predictions.
  2. Supervised retraining in Reverb's seven languages, on cleaned and selected subsets of open speech datasets such as Emilia and variants of YODAS2, such as yodas2-sidon.
  3. RLHF-style post-training: several rounds of iterative DPO with an automatic judge fitted on human preference rankings.

The result has 297M parameters, less than half of OmniVoice's 613M. Additional performance improvements let it generate speech 10× to 20× faster than the original model.

How Reverb is made: OmniVoice is pruned and distilled, retrained on seven languages, then refined with iterative DPO against a critic fitted on human preference rankings.

Quality, size and speed

We compare Reverb with open-weight voice-cloning models on the MiniMax Multilingual Test Set, restricted to Reverb's seven languages: 700 sentences, 100 per language, each spoken in the voice of one of two reference speakers of that language. Every model generates one take per sentence with its default settings. WER (Whisper large-v3) and speaker similarity (SIM-o) come from OmniVoice's evaluation scripts and are averaged over the languages.

Model Parameters: model + codec decoder WER ↓ SIM-o ↑
Reverb 297M + 22M 1.35% 0.818
OmniVoice 612M + 22M 1.64% 0.835
Audio8-TTS-Realistic-Air 601M + 131M 1.97% 0.750
Qwen3-TTS 0.6B Base 915M + 105M 11.96%¹ 0.781
Qwen3-TTS 1.7B Base 1,929M + 102M 1.52%¹ 0.781
VoxCPM2 2,290M + 46M 1.55% 0.810

The model's count is the one Hugging Face reports for its weights; the codec decoder turns codes into audio. Encoders used only to process the reference recording are not counted.
¹ Qwen3-TTS does not support Polish, so its WER is averaged over the other six languages. For Qwen3-TTS 0.6B, 4 of 600 takes didn't stop and were cut at 480 tokens (about 38 s); without them its WER is 1.78%.

Generation speed, in × real time (seconds of audio per second of compute; higher is better), on MiniMax English sentences. Timing starts from the text and an already-encoded reference voice. Every model runs its default settings.

Hardware Reverb OmniVoice Audio8 Qwen3-TTS 0.6B Qwen3-TTS 1.7B VoxCPM2 Fish S2 Pro
NVIDIA RTX 5090, batch 32 164.9× 13.2× 4.2× 3.3× 4.0× 3.0ײ 0.18ײ
Apple M4 Pro, batch 1 12.3׳ 0.84× 0.76× 0.51× 0.52× 0.69× –
iPhone 17, batch 1 20.6× 0.91× 1.03× 1.14× 0.74× – –

RTX 5090 and M4 Pro: PyTorch in bf16 on the GPU, to the waveform with codec decoding. iPhone 17: ExecuTorch MLX programs in fp16 on the GPU, to the codec codes, without the decoder.
² No batched generation: sentences run one at a time.
³ Reverb's ExecuTorch MLX program (fp16) with the Core ML codec decoder, as it ships: 21.6×.

Speed against size on an RTX 5090, batch 32: every other model is bigger and slower than Reverb-TTS

Try it

The demo Space lets you clone a voice from your own recording or try one of the preset voices.

Usage

pip install reverb-tts

Clone a voice

import soundfile as sf
from reverb_tts import TTS

tts = TTS.from_pretrained("software-mansion/Reverb-TTS")
voice = tts.voice("reference.wav", "The exact transcript of the reference recording.", language="en")
audio = tts.generate("Hello! This is my cloned voice.", voice, language="en")
sf.write("output.wav", audio, tts.sample_rate)

The reference works best as 3 to 10 seconds of clean speech from one speaker, with a transcript that matches it word for word. References over 15 seconds make worse takes.

Use a preset voice

tts.preset_voices()  # ['de_female_standard', 'de_male_standard', ..., 'pt_male_standard']
voice = tts.preset_voice("en_male_narrator_1")
audio = tts.generate("Hello! This is a preset voice speaking.", voice, language="en")

Each name starts with the voice's language: de, en, es, fr, it, pl or pt. A voice can speak any of the seven languages, though it sounds most natural in its own.

Speak another language

language in generate is the language of the text, and it can differ from the voice's:

audio = tts.generate("Dzień dobry, jak się masz?", voice, language="pl")

In the browser with ruNNtime

ruNNtime runs Reverb on the GPU in the browser through WebGPU. Install it with npm install runntime and follow its getting-started guide.

Text preparation

Numbers, dates, money, and abbreviations are read out as words for you (Verbalis prepares the text). Long texts are split at sentence ends automatically. For seeds, speed, batches, and sampler settings, see the usage guide.

Model variants

safetensors

PyTorch weights for the reverb-tts package, in five variants:

Variant Weights Activations Size
fp32 fp32 full precision 1.19 GB
fp16 fp16 full precision 594 MB
int8 (default) int8, one scale per output channel full precision 307 MB
int8a8 int8, one scale per output channel int8, one scale per token 307 MB
int4 int4, one scale per group of 32 inputs full precision 191 MB

Select one with TTS.from_pretrained(..., variant="int4"). The weights are in model/safetensors/<variant>/. On an internal test set, int8 showed no quality loss against fp32, and int4 none that we could measure. The benchmarks above use full precision.

ExecuTorch

Programs for on-device use are in model/executorch/, with the audio codec kept outside the program:

Backend Hardware Precisions
MLX Apple GPUs fp16, int8, int4
Vulkan Android GPUs int4
QNN Qualcomm NPUs, one program per chip: Snapdragon 8 Gen 1 (SM8450), 8 Gen 2 (SM8550), 8 Gen 3 (SM8650), 8 Elite (SM8750), 8 Elite Gen 5 (SM8850) int4

voices/ holds the 16 preset voices, covering all seven languages, as ready-encoded codec codes.

Limitations and responsible use

  • Reverb copies the voice in the reference, including its recording conditions: noise, room echo, and microphone coloring carry over into the output.
  • It has no style or emotion control apart from the reference: to get an angry take, use an angry reference.
  • It speaks only its seven languages. Text in other languages or scripts gives unpredictable output.
  • Clone only voices you have the right to use. Get the speaker's consent, don't impersonate real people, and don't use generated speech to deceive, defraud, or harass. Tell your listeners when they hear a synthetic voice.

License and acknowledgments

  • Code: Apache 2.0; see LICENSE.
  • Model weights: CC BY-NC 4.0, for non-commercial use with attribution.

Reverb's weights were initialized from OmniVoice by the k2-fsa team, and its target-length rule and prompt format come from OmniVoice's code. We thank its authors for releasing the model and code openly. Reverb uses the Higgs Audio v2 tokenizer as its audio codec, and Verbalis prepares its text.

Community Discord

Join the Software Mansion Community Discord to chat about Reverb or other Software Mansion libraries.

Reverb is created by Software Mansion AI Labs

Software Mansion AI Labs

We're an end-to-end AI development lab focused on local and hybrid AI. We build and fine-tune models, deploy them anywhere and turn them into agents that solve real business problems. Hire us.

Downloads last month
34
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for software-mansion/Reverb-TTS

Finetuned
Qwen/Qwen3-0.6B
Finetuned
k2-fsa/OmniVoice
Finetuned
(86)
this model