PenxPen/emotion2vec-plus-base-onnx

emotion2vec+ base, ONNX (fp32)

An ONNX export of emotion2vec+ base, the speech emotion model by the emotion2vec authors (Alibaba DAMO Academy / FunASR). It is published here so the open-source caption app VC Captions can download it on first use. The app's code isn't public yet.

Files

• emotion2vec_plus_base.onnx: the model, fp32. 16 kHz mono float waveform in; probabilities over 9 classes out.
• labels.txt: the class names, one per line, in output order.
• SHA256SUMS: checksums. The app checks each file against sums pinned in its source.

How it was made

Made by tools/emotion2vec/export.py in the app's CI ("emotion2vec export" workflow). The export follows FunASR's utterance path: normalised waveform → encoder → mean over time → linear head → softmax. The fp32 output matches PyTorch and FunASR to about 1e-14. int8 versions were tried and not published: their error depended on the CPU.

Licence and credit

The weights come from emotion2vec+ base and are shared under the FunASR Model Open Source License Agreement 1.1 (MODEL_LICENSE), which allows use, copying, modification and sharing with attribution to the source and authors. All credit for the model goes to the emotion2vec authors:

• emotion2vec: https://github.com/ddlBoJack/emotion2vec
• FunASR: https://github.com/modelscope/FunASR

This repository only changes the format (PyTorch to ONNX).