sherpa-onnx-wespeaker-resnet34-ax650

WeSpeaker ResNet34 speaker embedding model converted to AX650 AXMODEL for on-device speaker verification via sherpa-onnx.

Model Details

• Original Model: WeSpeaker voxceleb_resnet34
• Architecture: ResNet34 (CNN-based speaker embedding)
• Task: Speaker verification / speaker embedding extraction
• Input: 80-dim FBank features, up to 200 frames (~2s @ 16kHz)
• Output: 256-dim speaker embedding (L2-normalized)
• Target Chip: AX650 (NPU3)
• Quantization: U16 (MinMax calibration)

Files

File	Size	Description
wespeaker-resnet34.axmodel	6.6 MB	AX650 speaker embedding model

Performance (AX650)

Metric	Value
Cosine Similarity vs ONNX	0.9973
Max Absolute Error	0.030
MSE	0.00011
MACs	4.5 G
ONNX Size (original)	25 MB
AXMODEL Size	6.6 MB
Compression Ratio	3.8x

Usage with sherpa-onnx

./sherpa-onnx-offline-speaker-identification \
  --speaker-model=wespeaker-resnet34.axmodel \
  --provider=axera \
  test.wav

Conversion Details

• Pulsar2 Version: 6.0 (48520c11)
• Input Shape: [1, 200, 80]
• Calibration: 10 random FBank samples

License

Apache 2.0