Qwen3-TTS gets a native C++ streaming port with voice cloning and 5090 benchmarks
offgridai · reddit · 2026-07-27
A Reddit user ported Qwen3-TTS to native C++ and added incremental streaming plus voice-cloning support.
What they built
- Apache 2.0 C++ streaming library for Qwen3-TTS
- Demo harness with transcription via sherpa-onnx
- Support for 0.6B and 1.7B models, CustomVoice, and VoiceDesign
- RTX 4090/5090 CUDA builds, async transformer/vocoder execution, callback-only integration, and unit tests
Reported performance on an RTX 5090
- New clone creation: about 2.5s from 48s of reference audio
- Cold start from clone: about 1.85s
- First 350 ms of audio: about 400 ms
- Streaming speed: about 1.2× real time
The author says they are using it for a gaming project, but it could also help with local assistants and accessibility tools, and they are looking for RTX 4090 testing and feedback.
More from coding & agent
- An AI agent that turns daily reports into hourly status updates — huangyun_122 · 2026-07-27
- A developer ships code on vacation using only a phone and Codex — SupPandaHugger · 2026-07-27
- Opus 5 code demo improves and the source is now public — bdsqlsz · 2026-07-27
- AI-assisted coding is starting to feel like a multiple-choice exam — dfinke · 2026-07-27
- People doubt Matt Shumer’s game prompt, then run it and build custom games — mattshumer_ · 2026-07-27
- Claude Code Max user spends about $600 in tokens building a game with Matt Shumer’s prompt — mattshumer_ · 2026-07-27