narteybrown/signal-bench-ad-v1

Signal-Bench AD Test Set v1

What this is

Log-mel-spectrogram features extracted from the DCASE 2020 Task 2 ToyADMOS ToyCar test partition, packaged for signal-bench's Phase 5 anomaly-detection benchmarks.

This dataset contains processed features only, not raw audio. DCASE 2020 Task 2 audio is research-use-only and prohibits redistribution of raw audio. Users who need raw audio must obtain it directly from Zenodo (record 3678171) via signal-bench's scripts/datasets/stage_dcase.py.

Source attribution

Raw audio: DCASE 2020 Task 2 development dataset, ToyCar machine type.

• URL: https://zenodo.org/records/3678171
• Source paper: Koizumi et al., "ToyADMOS: A Dataset of Miniature-machine Operating Sounds for Anomalous Sound Detection," WASPAA 2019.
• License: Research use only, no redistribution of raw audio.

Feature extraction parameters match MLPerf Tiny v1.2's AD reference preprocessing:

• Pinned commit: 5dae3296bd899ed58a65311a8e6fd91d83f664ab
• Upstream baseline: benchmark/training/anomaly_detection/baseline.yaml + common.py::file_to_vector_array

Processing

Each 10-second 16 kHz audio clip is converted to a 2D feature matrix following the MLPerf Tiny AD reference pipeline:

1. librosa.feature.melspectrogram(n_fft=1024, hop_length=512, n_mels=128, power=2.0) → power-spectrum mel filterbank.
2. 20.0 / power * np.log10(mel + sys.float_info.epsilon) → dB-scaled log (resolves to 10 * log10(...) for power=2.0).
3. [:, 50:250] → central 200-frame slice (matches upstream's "take central part only" step, skipping the first ~1.6 s and last ~2 s of the 10 s clip to avoid edge effects).
4. 5-frame context concatenation: each output row stitches 5 consecutive log-mel columns into a 640-dim vector (128 mels × 5 frames).
5. Output per clip: (196, 640) float32 — 196 overlapping context-concatenated vectors.

Each row in the dataset corresponds to one input clip; the features column is a 2D array of shape (196, 640) per clip. Anomaly detection at inference time evaluates each 640-dim vector independently against the reference autoencoder and aggregates per-clip anomaly scores.

Subset structure

• Samples: 2,459 clips (1,400 normal + 1,059 anomaly across machine IDs 01–04).
• Classes: 2 (binary: 0 = normal, 1 = anomaly).
• Feature shape per clip: (196, 640) float32.
• Columns: features (2D array), label (int), filename (str, source WAV name for provenance only).

Intended use

Benchmark evaluation for edge AI anomaly detection. Designed to pair with MLPerf Tiny's reference autoencoder (ad01_int8.tflite, 640-dim input, dense AE).

from datasets import load_dataset
ds = load_dataset("narteybrown/signal-bench-ad-v1", split="test")
print(ds)
# Each row: features (196 x 640 float array) + label (0 or 1) + filename

MCU subset

A 100-sample class-balanced subset (50 normal + 50 anomaly, one random frame per clip, seed=42) is committed to the signal-bench repo at data/mcu_subsets/ad/subset_v1.npz. Subset arrays: inputs (100, 640) float32 + labels (100,) int64 + sources (100,) string.

Balanced 50/50 is a deliberate stratification choice for MCU benchmarking (per-sample latency/energy measurement on balanced classes), deviating from the source distribution (1,400 normal / 1,059 anomaly). To regenerate: uv run python scripts/datasets/prepare_ad.py.

Citation

@misc{signalbench-ad-v1,
  author = {Brown, Daniel},
  title = {Signal-Bench AD Test Set v1},
  year = {2026},
  publisher = {Agoo AI},
  howpublished = {\url{https://huggingface.co/datasets/narteybrown/signal-bench-ad-v1}},
}

@inproceedings{Koizumi_WASPAA2019_01,
  author = {Koizumi, Yuma and Saito, Shoichiro and Uematsu, Hisashi and Harada, Noboru and Imoto, Keisuke},
  title = {{ToyADMOS}: A Dataset of Miniature-machine Operating Sounds for Anomalous Sound Detection},
  booktitle = {Proceedings of {IEEE} Workshop on Applications of Signal Processing to Audio and Acoustics ({WASPAA})},
  year = {2019},
  pages = {308--312},
}

@inproceedings{Koizumi_DCASE2020_01,
  author = {Koizumi, Yuma and Kawaguchi, Yohei and Imoto, Keisuke and others},
  title = {Description and Discussion on {DCASE}2020 Challenge Task2: Unsupervised Anomalous Sound Detection for Machine Condition Monitoring},
  booktitle = {Proceedings of the DCASE 2020 Workshop},
  year = {2020},
  pages = {81--85},
}

Limitations

ToyADMOS recordings were collected from miniature toy machines (not real industrial equipment), with environmental noise mixed in post-recording. Benchmark results using this set measure inference performance on a standardized reference workload; they do not generalize to real-world industrial anomaly detection.

The MCU subset (100 frames, 50/50 balanced) is small enough that per-class anomaly-score statistics have wide confidence intervals; it is intended for latency and energy measurement, not AUC reporting.

License

DCASE 2020 Task 2 audio is licensed for research use only and prohibits redistribution of raw audio. This dataset publishes derived features (log-mel-spectrograms with 5-frame context concatenation), which are derivative works rather than the raw audio itself. Users requiring raw audio must obtain it directly from Zenodo record 3678171.