Vietnamese TTS that clones voices on a laptop CPU

What it does VieNeu-TTS generates Vietnamese speech at 24 kHz, handles English-Vietnamese code-switching, and clones any voice from 3–5 seconds of reference audio. It ships as a Python SDK with three speed tiers: a GPU-backed standard mode, a CPU-optimized variant, and a lightweight “Turbo” mode for edge devices and real-time chatbots.

The interesting bit The project treats voice cloning as a deployment problem, not just a research demo. It packages models in GGUF/ONNX for CPU inference, offers a Docker server with built-in tunneling, and even specifies a formal voices.json preset format so third-party voices remain compatible across SDK versions. The bilingual pronunciation is handled by a separate sea-g2p module the author also maintains.

Key highlights

• Runs fully offline; GPU optional, CPU sufficient for Turbo mode
• Zero-shot cloning needs no transcript in Turbo mode, though Standard mode asks for ref_text
• “Podcast & Conversation” mode with multi-speaker dialogue and emotional presets (natural, storytelling)
• Server deployment via LMDeploy with remote SDK access so clients load only a lightweight codec
• 10,000+ hours of bilingual training data claimed (no independent verification cited)

Caveats

• Turbo mode trades quality for speed and can glitch on very short phrases (< 5 words)
• Standard voice cloning requires reference text for higher accuracy; the “no transcript” ease is Turbo-only
• Full quality mode demands CUDA ≥ 12.8 or Apple Silicon MPS; the CPU path is deliberately degraded

Verdict Worth a look if you’re building Vietnamese voice interfaces, audiobook pipelines, or localized AI assistants where cloud latency or privacy rules out SaaS TTS. Skip if you need production-grade cloning without quality trade-offs or if your workload is primarily non-Vietnamese.