ZeroTTS is a Vietnamese text-to-speech package from ZeroWeight AI that ships both code and weights under MIT and runs its entire inference path on ONNX Runtime with no PyTorch involved: pip install zerotts pulls numpy, onnxruntime, tokenizers, huggingface_hub, soundfile and scipy, and the roughly 900 MB fp32 weight set downloads from the Hugging Face Hub on first use. The model is 202M parameters and targets CPU rather than GPU — the project measures a 0.50x realtime factor and a first audio chunk at about 70 ms on CPU with eight inference threads, figures produced on its own ZeroBench-TTS dataset with PhoWhisper-large as the ASR scorer. On that benchmark the README reports 1.03% word error rate against 4.13% for OmniVoice and 16.42% and 18.40% for two XTTS-v2 Vietnamese fine-tunes, plus a 2.91 UTMOS naturalness score and 0.029 seconds of excess silence; it also records OmniVoice ahead on speaker similarity at 0.950 versus ZeroTTS's 0.936. Every system in the table reads raw text with no normalisation frontend, and the documented generation defaults are the exact settings the scores were produced with. The limitation worth reading first is the headline capability itself: the README states plainly that voice cloning is not available in this release, because the voice encoder that turns a reference clip into speaker latents is not published. Eight preset voice packs ship with the weights and any latent array can be passed straight to synthesize(), but creating a voice from your own audio means uploading a clip to the company's hosted platform and installing the resulting zip. Around that core sit a streaming Python API emitting 48 kHz float32 chunks, a CLI, a Gradio web UI, a client-side browser demo running the same graphs through onnxruntime-web, a separate Vietnamese normaliser for dates, times and numbers, and chunking helpers, since the model is trained on utterances rather than paragraphs. The weights are shipped unquantized fp32, which the authors call a deliberate quality-over-size choice aimed at desktop broadband rather than mobile data, and the package is on PyPI as zerotts, with v0.1.4 tagged on 2026-09-14. The bundled MOSS-Audio-Tokenizer-Nano decoder is Apache-2.0 and the benchmark dataset is CC-BY-NC-4.0.