AI BriefingSopro V2, Lightweight TTS Released
·2026.08.27 23:38
Key point
The 120-million-parameter open-source TTS model Sopro V2 delivers SOTA-level voice cloning performance and enables fast inference even on CPUs.
Details
The open-source TTS model Sopro V2, with 120 million parameters, has been released. It is licensed under Apache-2.0 and supports zero-shot voice cloning using only 5–20 seconds of reference audio.
Performance and Technical Features
- On the Seed-TTS-eval benchmark, it achieved an English WER of 1.51–1.65, demonstrating competitive performance against models 3–14 times larger, such as F5-TTS (1.83) and CosyVoice 3 (2.02).
- Optimized for CPU inference, it generates the first audio in approximately 300ms (0.21 RTF) on an M3 CPU, and supports fast streaming at 0.07 RTF on H100.
- It supports English, European Portuguese, French, and German, and is the first open TTS model to natively support European Portuguese.
Deployment and Usage
- You can run a local web UI with a single command:
uvx --from sopro soprotts serve. - Model weights and code are available via GitHub, Hugging Face, and PyPI, and a browser-based demo is also provided.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Qwen3.5-based Small Decision Model 'Kev' Released with TypeSafe API Compatibility and Performance MetricsCua Releases Small Specialized Model 'CUA-S1' and Open-Source Ecosystem for Computer-Use Agents