ONNX export of
inesc-id/EBranch-w2vBERT2-EP
(ESPnet2 E-Branchformer with a w2v-BERT 2.0 frontend, trained on ~425 h of European Portuguese
from the CAMOES project) for
onnx-asr with the
espnet-aed model type (requires the espnet model-family patches, see
TigreGotico/onnx-asr PR #4).
This repo holds the full-model AED variant: encoder plus transformer attention decoder with greedy decoding. Best quality; adds a ~127 MB (40 MB int8) decoder.
The original model uses partial rotary position embeddings (dim=32 of 64, interleaved pairs,
self-attention only); the rotary frequency table is taken verbatim from the released checkpoint.
fp32 CTC output is bit-exact with the original PyTorch model (8/8 test clips identical).
fp32 and int8-quantized graphs are included.
1import onnx_asr
2model = onnx_asr.load_model("espnet-aed", "path/to/this/repo") # or quantization="int8"
3print(model.recognize("audio_16khz.wav"))
License: CC-BY-4.0 (inherited from the source model; attribution: INESC-ID, CAMOES project,
arXiv:2508.19721).
