Soprano ONNX (KV Cache)

This repository hosts ONNX exports of the Soprano 1.1 80M model with KV caching.

Contents

• onnx/soprano_backbone_kv_fp32.onnx, soprano_backbone_kv_fp16.onnx, soprano_backbone_kv_int8.onnx (backbone with past_key_values)
• onnx/soprano_decoder_fp32.onnx + onnx/soprano_decoder_fp32.onnx.data (vocoder decoder)
• onnx/soprano_decoder_int8.onnx (vocoder decoder)
• / (tokenizer assets)

Inference & demo

See the streaming inference code here: https://github.com/KevinAHM/soprano-web-onnx

Not compatible with WebGPU via onnxruntime-web as of January 2026.

Upstream

Original project: https://github.com/ekwek1/soprano