yucongzh/echo-small-v2

ECHOv2

ECHOv2 is a band-splitting audio representation model for anomalous sound detection (ASD) with explicit cross-band interaction. It extends the original ECHO framework (link) with a two-level self-distillation design to model both intra-band and inter-band information.

✨ Key features

• —Band-splitting modeling: learns frequency-localized representations from spectrogram sub-bands.

• —Cross-band interaction: explicitly models dependencies across frequency regions.

• —Two-level self-distillation: jointly optimizes intra-band and inter-band learning.

• —Structured summary tokens: supports region-aware cross-band aggregation with controllable granularity.

• —Strong ASD performance: achieves consistent gains on DCASE 2020–2025 ASD benchmarks.

📊 Performance

The overall performance on the ASD benchmarks.

Embedding-based Results: [image]

Adaptation-based Results: [image]

Minimal usage

import torchaudio
from ECHOv2 import from_pretrained

model = from_pretrained()

wav, sr = torchaudio.load("/path/to/audio.wav")
if wav.shape[0] > 1:
    wav = wav.mean(dim=0, keepdim=True)

utt_feature, frame_feature = model.extract_features_from_audio(wav, sr)
print(utt_feature.shape, frame_feature.shape)

📚 Citation

@inproceedings{echo2025,
  title={ECHO: Frequency-aware Hierarchical Encoding for Variable-length Signal},
  author={Yucong Zhang and Juan Liu and Ming Li},
  booktitle={Proc. ICASSP},
  year={2026},
}