sherpa-onnx-zipformer-ctc-en-ax650

Zipformer CTC English ASR model converted to AX650 AXMODEL for on-device inference via sherpa-onnx.

Model Details

• Original Model: csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02
• Architecture: Zipformer2 CTC
• Task: English speech recognition (ASR)
• Sample rate: 16000 Hz
• Feature dim: 80-dim mel filterbank
• Frame shift: 10ms
• Subsampling factor: 4
• Vocabulary size: 500
• Target Chip: AX650 (NPU3)
• Quantization: U16 (int16)

Files

File	Size	Description
zipformer-ctc.axmodel	73 MB	AX650 compiled model
tokens.txt	5 KB	500 output tokens

Input / Output

• Input: x [1, 512, 80], x_lens [1] — 512 frames (~5.12s @ 16kHz)
• Output: log_probs [1, 126, 500] — CTC log probabilities

Usage with sherpa-onnx

./sherpa-onnx-offline \
  --zipformer-ctc-model=zipformer-ctc.axmodel \
  --tokens=tokens.txt \
  --provider=axera \
  audio.wav

Known Limitations

• Fixed input frame length (T=512); shorter inputs must be padded
• Output log_probs length determined by subsampling_factor=4

Conversion Details

• Pulsar2 Version: 6.0
• Input Shape: x [1, 512, 80], x_lens [1]
• Calibration: MinMax with 10 random mel features
• Quantization: U16

License

Apache 2.0