sherpa-onnx-gtcrn-ax650

GTCRN (Grouped Temporal Convolutional Recurrent Network) speech enhancement model converted to AX650 AXMODEL for on-device denoising and dereverberation via sherpa-onnx.

Model Details

• Original Model: k2-fsa/sherpa-onnx-gtcrn-dereverb-advanced-v0
• Architecture: GTCRN (Grouped Temporal Convolutional Recurrent Network)
• Task: Speech enhancement (denoising + dereverberation)
• Sample rate: 16000 Hz
• Target Chip: AX650 (NPU3)
• Quantization: U16

Files

File	Size	Description
gtcrn.bin	1.8 MB	GTCRN speech enhancement AXMODEL

Note: The model file uses .bin extension for upload compatibility. Rename to .axmodel if needed by your sherpa-onnx build, or use as-is.

Usage with sherpa-onnx

./sherpa-onnx-offline-denoiser \
  --gtcrn-model=gtcrn.bin \
  --provider=axera \
  noisy.wav enhanced.wav

Performance

Metric	Value
AXMODEL size	1.8 MB
RTF	TBD

Conversion Details

• Pulsar2 Version: 5.1
• Calibration: MinMax with 10 samples
• Input Shape: [1, 256, 257]

License

Apache 2.0