ONNX Magpie Runtime Bundle
This repository contains ONNX Runtime artifacts for a local MagpieTTS conversion.
It is intended for use with the onnx-magpie runtime and does not require NeMo
or PyTorch at inference time.
Files
magpie.pipeline.json: runtime pipeline manifest.magpie_text_context.onnx: text/speaker context graph.magpie_decoder_prefix.onnx: autoregressive decoder prefix graph.magpie_codec_decoder.onnx: codec-token to waveform graph.magpie.pipeline.kv-cache.json: optional explicit-KV decoder pipeline manifest.tokenizers/: dependency-free tokenizer artifacts.onnx-magpie-hub.json: download manifest consumed by the runtime.
Runtime
python scripts/run_onnx_magpie.py --hf-repo mweinbach1/onnx-magpie --text "Hello world." --wav-output speech.wav
The original MagpieTTS checkpoint is governed by NVIDIA's model license. Verify that your use of these converted artifacts complies with the upstream terms.