zeroweight-ai/ZeroTTS

What it solves

ZeroTTS makes it easy to generate natural-sounding Vietnamese speech from text, even on a normal laptop without a GPU. It solves two common problems: getting a TTS that reads Vietnamese correctly (including tones, dates, numbers, and mixed English words) and doing it fast enough to stream audio as it's generated. The project also allows you to use a specific voice by providing a short sample of it, though in this open-source release the voice cloning step itself is done through their platform—the package can only load voices you've already obtained.

How it works

ZeroTTS uses a neural network that takes text and a "voice latent" (a small numeric representation of a speaker's voice) and produces audio. The model is trained on Vietnamese speech and is optimized to run on CPU using ONNX Runtime, which is why it can run faster than real time. It streams audio in small chunks, so the first sound arrives in about 70 ms. For Vietnamese, it has a built-in text normalizer that expands dates, numbers, and abbreviations into spoken words, though you can switch that off if you're feeding it English text.

The package includes eight pre-built voices, and you can also add your own by dropping in a .npz file (or a zip of such files) that you get from their platform. The voice encoder that creates these latents from audio is not published, so you can't train your own voices entirely within the package.

Who it's for

ZeroTTS is aimed at developers and hobbyists who need Vietnamese TTS in applications like chatbots, content creation, or accessibility tools. It's especially useful if you want TTS that runs locally without cloud services or GPUs, and you care about low latency and accuracy. The browser demo also shows it can run entirely client-side.

Highlights

• State-of-the-art accuracy: about 4× fewer word errors than the next best open Vietnamese TTS on their benchmark.
• Real-time performance on CPU: about 2× faster than real time, with the first audio chunk in ~70 ms.
• Zero-shot voice cloning: voices can be cloned from 3–30 seconds of reference audio, though the cloning tool itself is only available via their platform.
• Pure Python package with minimal dependencies (no PyTorch needed for generation).
• Includes a normalizer for Vietnamese text that expands dates, times, fractions, and acronyms.
• Optional int8 quantization to roughly halve synthesis time on memory-limited CPUs.
• Web UI and browser demo included.