v0.1.1 — PyPI page fixes

Packaging fix only — no code changes. The library, the CLI and the weights
are byte-for-byte what 0.1.0 shipped, so there is no reason to upgrade except a
readable project page.

Fixed

• Broken banner and images on the PyPI project page. PyPI renders the
  README without any repository context, so relative paths like
  docs/assets/banner.png resolved against pypi.org and 404'd. Every image
  and in-repo link is now an absolute URL.
• Blank sample players on PyPI. PyPI sanitizes the README against a tag
  allowlist that has no <video> in it, so the three sample clips rendered as
  empty table cells. Each player now carries a fallback .mp3 link inside it —
  GitHub still shows the video players, PyPI shows playable audio links.

Install

pip install --upgrade zerotts
zerotts say "Xin chào" --voice maichi -o out.wav

Full changelog: v0.1.0...v0.1.1

v0.1.0 - first public release

First public release of ZeroTTS — Vietnamese zero-shot text-to-speech that runs
on ONNX Runtime. No PyTorch, no GPU, faster than real time on a laptop CPU.

Why it's worth a look

4× fewer word errors than the next-best open Vietnamese system, and the
fastest of the four on CPU. Measured on
ZeroBench-TTS,
every system reading the same normalized text:

	ZeroTTS	OmniVoice	XTTS-v2-vietnamse	viXTTS
WER ↓	0.56 %	2.12 %	7.27 %	8.61 %
Naturalness (UTMOS) ↑	2.91	2.75	2.49	2.34
Dead air ↓	0.029 s	0.386 s	0.568 s	0.215 s
RTF, CPU ↓	0.50×	6.12×	0.71×	0.73×
Time to first audio, CPU ↓	~70 ms	~34 s	~6.1 s	~5.1 s
Size	202 M params, 0.86 GB fp32, CPU	3.1 GB, GPU	1.9 GB, GPU	1.9 GB, GPU

The gap is widest in latency: the two XTTS fine-tunes also beat real time, but
need seconds before the first sample. ZeroTTS streams its first audio frame in
about 70 ms.

What's in it

• Real-time streaming on CPU — RTF 0.50×, first chunk in ~70 ms, 8 threads,
  no GPU anywhere in the path.
• Vietnamese text normalization built in — dates, times, numbers, %, @
  and abbreviations expanded to Vietnamese words before synthesis, so
  31/12/2026 and ZeroTTS are read the way a person would read them.
• Code-switched English handled inside Vietnamese sentences.
• Eight voice packs shipped with the weights: maichi, baotrang,
  kimoanh, hamy, giahuy, huuduc, quangminh, tiendat.
• Long-form chunking — sentence-aware splitting with the audio joined back
  into one continuous take.
• Three ways to run it — Python API, the zerotts CLI, a Gradio web UI
  (pip install "zerotts[webui]"), plus a fully in-browser WebAssembly demo
  under js/ that downloads the weights once and synthesizes locally.

Install

pip install zerotts
zerotts say "Xin chào" --voice maichi -o out.wav

from zerotts import ZeroTTS

tts = ZeroTTS.from_pretrained("zeroweight-ai/ZeroTTS")
audio = tts.synthesize("Xin chào các bạn, mình là ZeroTTS.", voice="maichi")
tts.save_audio(audio, "out.wav")

Weights are pulled from
🤗 zeroweight-ai/ZeroTTS on first
use. Python 3.9+, MIT licensed.

Known limits

• Voice cloning from your own audio is not in this release. The voice
  encoder is not published — a voice here is a precomputed latent array shipped
  with the weights, so you get the eight packs above and no way to add a ninth.
  See docs/VOICES.md, or
  zeroweight.ai if you need latents for your own voice.
• CFG above 1.0 doubles inference cost — it runs a second unconditional branch
  per frame. The default of 1.0 is what the benchmark numbers were produced
  with.
• Unconditioned mode (no voice pack) does not keep a consistent speaker between
  segments; it is a quality check, not a usable voice.

Speech codec: MOSS-Audio-Tokenizer-Nano
by the OpenMOSS team (Apache-2.0) — see NOTICE.