First-shot anomalous sound detection (DCASE 2024 Task 2)
Task
- Input: normal training clips of one machine type (
waveforms (n, T) float, zero-padded, with lengths and sample_rate, 16 kHz mono in DCASE) and unlabeled test clips. There is no labelled anomaly and no target-domain training clip, so the domain shift must be inferred from the data.
- Deliverable: a 1-D finite float array, one anomaly score per test clip in input order, larger = more anomalous; only the ordering matters. Score each machine type separately with its own training clips.
- Quality: official DCASE score = harmonic mean of AUC over source-domain normals plus anomalies, AUC over target-domain normals plus anomalies, and pAUC (max FPR 0.1) over all clips; higher is better.
scilib.anomsound.official_score(y_true, domain, score) computes it when labels and domains exist.
Routes
- Log-mel route:
anomsound.log_mel_list(waveforms, lengths, sample_rate) (n_fft 1024, hop 512, 128 mels, dB; operator audio_log_mel_spectrogram), then component_scores (nn_train, nn2_pool, nn_pool, lof, band_max, maha) and rank_average. One call: anomsound.fit_predict(train_mels, eval_mels) (operator sound_anomaly_scores_log_mel) or anomsound.score_clips(train, evalset), using DEFAULT_MEMBERS. nn_train (1-nearest-neighbour distance to the training clips) is the simple baseline.
- Pretrained-embedding route:
scilib.audioenc.embed(waveforms, lengths, sample_rate, model, kind) with ast_audioset (pooled or logits) or clap_htsat (pooled or proj) (operators audio_embedding_ast, audio_embedding_clap), then anomsound.embedding_scores(train_emb, eval_emb) and rank_average(raw, ("nn2_pool",)) (operator sound_anomaly_scores_embeddings), or score_clips(train, evalset, embed=["ast_audioset", "clap_htsat"]). Assets ast_audioset, clap_htsat_unfused; beats.encode (asset beats_iter3, mean-pool its tokens) can also feed embedding_scores. Check audioenc.available() and scienceclaw_tools(operation=weights). It is a candidate to compare with the log-mel route, not a guaranteed replacement.
Pitfalls
nn2_pool, nn_pool and lof place the unlabeled test clips in the point cloud together with the training clips. That uses no labels and is fine; but never use file names, ids or metadata that encode domain or condition, and never fit on test clips with guessed labels.- A single component can be anti-correlated with the anomalies on a small test set, and AUC from a few dozen clips is noisy. A rank average of complementary clip-level scores is more stable; do not pick components by peeking at test labels.
- Clips differ in length (use
lengths; log_mel_list cuts to the true length, so padding is not scored), and the clip descriptors are z-scored with training statistics only (at least 3 training clips are needed).
- Disclose possible pretraining overlap: AST was fine-tuned on AudioSet and CLAP trained on LAION-Audio-630K; overlap with the task audio cannot be ruled out.
