Saudi Arabic Zipformer2 ASR (ONNX)
Zipformer2 Transducer models for Saudi Arabic (SAU) dialect speech recognition, trained on GigaSpeech3 Arabic data with the icefall framework.
Contents
| File | Description |
|---|---|
onnx/encoder-*.onnx | Encoder (Conv2dSubsampling + Zipformer2 + encoder projection) |
onnx/decoder-*.onnx | Prediction-network decoder |
onnx/joiner-*.onnx | Joiner output layer |
tokens.txt | BPE token table (SentencePiece unigram 500) |
unigram_500.model | SentencePiece BPE model |
Models
| Name | Description |
|---|---|
| M2_0818_fixed | textnorm-fixed retrain, epoch 30 avg 5 (recommended) |
| M2 (ebra_batch1n2_sada_haitian) | ebra batch1n2 + sada + haitian, epoch 25 avg 5 |
Usage
Decode with ONNX Runtime. See the demo Space for a complete example:
- Encoder: fbank (80-dim) → encoder_out (N, T', 512), encoder_out_lens (N,)
- Decoder: token ids (N, 2) → decoder_out (N, 512)
- Joiner: encoder_out + decoder_out → tanh → linear → logits (N, 500)
License
Apache-2.0. Trained on GigaSpeech3 Arabic data.