sherpa-onnx-paraformer-ax650

Paraformer-large-Chuan Chinese ASR model converted to AX650 AXMODEL for on-device inference via sherpa-onnx.

Model Details

• Original Model: csukuangfj/WSChuan-ASR
• Architecture: Paraformer (SAN-M encoder + CIF predictor + SAN-M decoder)
• Task: Chinese speech recognition (ASR)
• Sample rate: 16000 Hz
• Input: 5-second fixed length audio chunks
• Vocabulary: 8404 tokens
• LFR: window=7, shift=6
• Target Chip: AX650 (NPU3)

Files

File	Precision	Size	Description
encoder-5-seconds.axmodel	INT8	165 MB	Encoder [1,83,560]→[1,83,512]
predictor-5-seconds.axmodel	INT8	831 KB	Predictor [1,83,512]→[1,83]
decoder-5-seconds.axmodel	INT8	60 MB	Decoder [1,83,512]×3→[1,83,8404]

Performance

Metric	Value
RTF	0.007
Pulsar2 Version	6.0 (commit 48520c11)

Usage with sherpa-onnx

./sherpa-onnx-offline \
  --tokens=tokens.txt \
  --paraformer=encoder-5-seconds.axmodel,predictor-5-seconds.axmodel,decoder-5-seconds.axmodel \
  --provider=axera \
  audio.wav

Known Limitations

• INT8 quantization may cause accuracy loss; recognition output may be empty or partially missing
• Fixed 5-second input length; longer audio requires external chunking
• For higher accuracy, use U16 encoder + FP32 predictor + FP32 decoder variants

Conversion Details

• Pulsar2 Version: 6.0 (48520c11)
• Calibration: MinMax with 10 samples
• Quantization: INT8

License

Apache 2.0