A static, client-side browser demo that runs the Soprano TTS pipeline using onnxruntime-web.

This is a conversion/port of the original Soprano project: https://github.com/ekwek1/soprano

• A modern browser (Chrome, Edge, Firefox, Safari).
• You must serve this folder over HTTP (opening index.html via file:// usually breaks fetch() / module loading).
• The demo loads onnxruntime-web and @huggingface/transformers from a CDN by default (network required unless you vendor them).
• The model files are large; plan to use Git LFS or GitHub Releases if you publish them.

Place model artifacts under ./models/:

.
├─ index.html
├─ onnx-streaming.js
├─ PCMPlayerWorklet.js
├─ style.css
├─ onnx/
│  ├─ soprano_backbone_kv.onnx
│  ├─ soprano_decoder.onnx
│  └─ soprano_decoder.onnx.data
├─ tokenizer.json
├─ tokenizer_config.json
├─ special_tokens_map.json
├─ config.json
└─ generation_config.json

Notes:

• ONNX models live in onnx/ following HuggingFace convention.
• The decoder uses external weights (.onnx.data file must be present alongside the .onnx file).
• Tokenizer files are in the root directory.

Use any static file server from this directory, for example:

python -m http.server 8085

Then open http://localhost:8085.

Model paths are defined near the top of onnx-streaming.js in the MODELS object.

Sampling defaults are set in onnx-streaming.js (constructor):

• temperature
• topK
• topP
• repetitionPenalty

• "Load failed" / model never becomes Ready
  • Verify the onnx/ filenames match MODELS in onnx-streaming.js
  • Check DevTools → Network for a missing .onnx.data file (404)
  • Confirm / contains tokenizer.json (and related files)
• Performance notes
  • Both backbone and decoder run on WASM/CPU
  • Achieves real-time streaming on modern hardware

Soprano is released under Apache-2.0 in the upstream repository: https://github.com/ekwek1/soprano