Software Mansion's Reverb-TTS Clones Voices in 7 Languages Running on iPhones
Software Mansion AI Labs releases Reverb-TTS, a compact zero-shot voice cloning model that tops the Open TTS Leaderboard and runs locally on phones.
Takeaways
- Software Mansion AI Labs releases Reverb-TTS, a zero-shot voice cloning model in 7 languages
- 319M parameters, 1.03% WER on MiniMax, beating much larger open-weight models
- Runs at 164.9x real time on RTX 5090 and over 20x on iPhone 17
- Built by pruning and distilling OmniVoice, then iterative DPO with an automatic judge
- Ships as int4/int8 ExecuTorch programs for Apple MLX, Android Vulkan, Qualcomm NPU
- Weights CC BY-NC 4.0 (non-commercial); code Apache 2.0; GitHub
Software Mansion AI Labs has released Reverb-TTS, an open-weight text-to-speech model that clones a voice from a short recording. Its 319-million-parameter benchmark entry posts the lowest word error rate in the Open TTS Leaderboard’s MiniMax comparison, while reported batch-one performance exceeds 20 times real time on an iPhone 17 and Apple M4 Pro.
Voice-cloning systems condition generated speech on a reference speaker’s recording. Reverb accepts 3 to 20 seconds of speech without per-speaker fine-tuning, then synthesizes English, German, French, Spanish, Italian, Polish, or Portuguese using the cloned voice. The downloadable weights support local deployment, although their noncommercial license restricts paid products.
Low error rates, close voice matches
The MiniMax test set contains 100 sentences per language, each generated from a reference recording. Word error rate, or WER, measures how accurately an automatic speech recognizer recovers the requested text; lower scores are better. SIM-o compares embeddings from generated and reference speech to estimate speaker similarity; higher scores are better.
| Model | Parameters | WER | SIM-o |
|---|---|---|---|
| Reverb | 319M | 1.03% | 0.817 |
| OmniVoice | 634M | 1.13% | 0.833 |
| VoxCPM2 | 2.34B | 1.20% | 0.812 |
| Fish S2 Pro | 4.95B | 1.35% | 0.761 |
| MOSS-TTS | 8.49B | 1.84% | 0.774 |
| Higgs TTS 3 | 4.65B | 6.59% | 0.754 |
The benchmark uses Qwen3-ASR to calculate WER and WavLM to calculate SIM-o. Reverb leads the listed models on WER, while OmniVoice records the highest speaker-similarity score at 0.833. These automated metrics provide reproducible comparisons, but they do not fully measure naturalness, prosody, emotional delivery, or listener preference.
The project describes Reverb’s core network as having 297 million parameters, while the leaderboard counts 319 million for the full benchmark entry. Developers estimating memory use should rely on the published artifact sizes for the selected precision and runtime.