Reverb-TTS: zero-shot voice cloning with open weights in 7 languages. 297M parameters; 232.3x faster than real time on an H200 and 164.9x on an RTX 5090 at batch 32, 21.6x on an M4 Pro and 20.6x on an iPhone 17; 0.96% word error rate on MiniMax. Software Mansion AI Labs.

Reverb-TTS

GitHub Β· Demo Β· Quick start Β· 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 on the Open TTS Leaderboard's MiniMax test set (0.96%).
  • Speed: 232.3Γ— real time on an NVIDIA H200 and 164.9Γ— on an RTX 5090 at batch 32, over 20Γ— 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 10Γ— to 20Γ— 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 well-known open-weight models on the Open TTS Leaderboard's MiniMax test set, in seven languages: 100 sentences per language, one take each, every model cloning a reference recording. WER (Qwen3-ASR) and speaker similarity (SIM-o, WavLM) are averaged over the languages.

Model Parameters WER ↓ SIM-o ↑
Reverb 319M 1.03% 0.817
OmniVoice 634M 1.13% 0.833
VoxCPM2 2,336M 1.20% 0.812
Fish S2 Pro 4.95B 1.35% 0.761
Audio8-TTS-Realistic 732M 1.46% 0.764
MOSS-TTS 8.49B 1.84% 0.774
Chatterbox 0.80B 2.76% 0.715
Higgs TTS 3 4.65B 6.59% 0.754

Generation speed, in Γ— real time (seconds of audio per second of compute; higher is better), with each model's default settings. The H200 row is the Open TTS Leaderboard's run over all seven MiniMax languages; the other rows are ours, on MiniMax English sentences, timed from the text and an already-encoded reference voice.

Hardware Reverb OmniVoice Audio8 Qwen3-TTS 0.6B Qwen3-TTS 1.7B VoxCPM2 Fish S2 Pro
NVIDIA H200, batch 32 232.3Γ— 45.5Γ— 11.2Γ— – – 1.9Γ— 0.40Γ—Β²
NVIDIA RTX 5090, batch 32 164.9Γ— 13.2Γ— 4.2Γ— 3.3Γ— 4.0Γ— 3.0Γ—Β² 0.18Γ—Β²
Apple M4 Pro, batch 1 12.3Γ—Β³ 0.84Γ— 0.76Γ— 0.51Γ— 0.52Γ— 0.69Γ— –
iPhone 17, batch 1 20.6Γ— 0.91Γ— 1.03Γ— 1.14Γ— 0.74Γ— – –

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. H200: the Open TTS Leaderboard's default voice, MiniMax test set, 7 languages.
Β² No batched generation: sentences run one at a time.
Β³ Reverb's ExecuTorch MLX program (fp16) with the Core ML codec decoder, as it ships: 21.6Γ—.

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. It runs Reverb on your GPU in the browser through ruNNtime, so it needs a browser with WebGPU, such as Chrome or Edge.

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("DzieΕ„ dobry, jak siΔ™ masz?", voice, language="pl")

On your Mac with Native

Run it on your Mac with Nativ, from Neywa Labs. You can read more in a blogpost here.

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 text-to-speech guide.

Text preprocessing

Reverb reads text in spoken form: it was trained on text with numbers, dates, money, units, symbols, and abbreviations written out as words, so normalizing the text is a required step before generating.

  • With the reverb-tts package, it's done for you: generate and voice run Verbalis on the text and the transcript by default.
  • With the ExecuTorch programs or ruNNtime, the text reaches the model exactly as you give it. Normalize it yourself first; we strongly recommend Verbalis, with C++, TypeScript, and Python APIs for all seven languages.

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

The programs take text exactly as given: normalize it first, ideally with Verbalis's C++ API (see Text preprocessing).

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
46
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for software-mansion/Reverb-TTS

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

Space using software-mansion/Reverb-TTS 1

Collection including software-mansion/Reverb-TTS