sherpa-onnx-ced-mini-ax650

CED (CNN-based Encoder-Decoder) audio tagging model converted to AX650 AXMODEL for on-device inference via sherpa-onnx. Detects 527 AudioSet event classes.

Model Details

• Original Model: k2-fsa/sherpa-onnx-ced-mini-audio-tagging-2024-04-19
• Architecture: CED (CNN-based Encoder-Decoder) - ConvNeXt style
• Task: Audio tagging (527 AudioSet event classes)
• Input: 64-dim mel-spectrogram features, up to 1001 frames (~10s @ 16kHz)
• Output: 527-class event probabilities
• Target Chip: AX650 (NPU3)
• Quantization: U16 (MinMax calibration)

Files

File	Size	Description
ced-mini.axmodel	11.2 MB	AX650 compiled model
class_labels_indices.csv	-	527 AudioSet class labels
test.wav	-	Sample test audio

Performance (AX650)

Metric	Value
RTF	0.007
Top-1 Accuracy vs ONNX	100%
Top-5 Accuracy vs ONNX	60%
Max Absolute Error vs ONNX	0.035
ONNX Size (original)	37 MB
AXMODEL Size	11.2 MB
Compression Ratio	3.3x

Tested on AX650 NPU3 with 3-second 16kHz mono audio.

Usage with sherpa-onnx

./sherpa-onnx-offline-audio-tagging \
  --ced-model=ced-mini.axmodel \
  --labels=class_labels_indices.csv \
  --provider=axera \
  test.wav

Or via Python:

import sherpa_onnx

config = sherpa_onnx.AudioTaggingConfig()
config.model.ced = "ced-mini.axmodel"
config.model.provider = "axera"
config.labels = "class_labels_indices.csv"

tagger = sherpa_onnx.AudioTagging(config)
stream = tagger.create_stream()
stream.accept_waveform(16000, samples)
results = tagger.compute(stream)

Conversion Details

• Pulsar2 Version: 6.0 (48520c11)
• Input Shape: [1, 64, 1001]
• Calibration: 10 random mel-spectrogram samples

License

Apache 2.0