Roadmap

Things we want to add, roughly in order. None of them are in the package yet.

Voice cloning without WebGPU#

Chatterbox's quantized graphs use an ONNX operator the WebAssembly build of ONNX Runtime lacks. A WebAssembly-friendly export would let cloning run in browsers without a GPU. Multilingual cloning is also on the list.

More models#

• Image embeddings (SigLIP) for embed with image input.
• Larger VLMs and text models as WebGPU memory limits rise.
• Faster TTS for low-end devices.
• More paint models once one-step models fit in browser memory.

Grammar-constrained decoding#

schema currently validates and retries. Constrained decoding would guarantee valid JSON on the first try.