IndoML 2026 Track 1: Multi-Backbone SED Ensemble with Snippet-Bank Matching (Team TDS)
Checkpoints and probability caches of our Datathon@IndoML 2026 Track 1: Noise Event Detection system. Authors: Shivam Dwivedi, Animish Yadav, Shubhi Jain (IIIT Delhi).
- Code, training and inference instructions: https://github.com/dwivediishivam/indoml26_track1
- Final Codabench submission 995227 (rank 6): Event F1 0.7711, Segment Dice 0.8232, combined 1.5943.
System
The system has three parts:
- Twelve frame-level SED members, each a pretrained encoder plus a mel-CNN feeding a 2-layer BiGRU, with 24 tracks per 20 ms frame (presence, 7 category tracks, union and per-category onset/offset). They are fused in the logit domain.
- A snippet bank: 206 noise snippets isolated from the validation synthetic pairs, detected in test audio by FFT normalised cross-correlation with chunk-lag consistency and a leave-fold-out calibrated logistic scorer.
- A routing decoder (bank, category-aware hysteresis decoder, or prior), tuned with test-mixture weighting.
| Member | Encoder | checkpoints/ file |
SHA-256 |
|---|---|---|---|
s9000 |
WavLM-base+ SED | s9000_weights_9000.pt |
a3acdf3711e314eeb203c346c184255b6ec462ef91b5b3e8115714b8531300d4 |
s1600 |
WavLM-base+ SED | s1600_weights_1600.pt |
4ad8a99a114034e420fb5667b00c7a9f005607cc5865a2ddc6fcba669aeb3ff8 |
m1_atst |
ATST-F strong + mel-CNN | m1_atst_student_step0018000.pt |
22593e67928cd982e39e7449c9902253af850132800e4daa1bcedc5fe8e3b0fe |
m4_breath |
WavLM-base+ + mel-CNN (breath specialist) | m4_breath_student_step0012000.pt |
87a98690a0afc47e6314ca6990ef76e7399fc1e782e17f67c954f003ca443028 |
m1g_atst_gold |
ATST-F strong + mel-CNN | m1g_atst_gold_ema_step0009000.pt |
20b800abd61ba3af6c7cd08e8c0f47017b68b9aa9be04224908673f4b9f886cc |
m1v_atst_val |
ATST-F strong + mel-CNN | m1v_atst_val_ema_step0009000.pt |
dcc28aa5d5b18f26a3864a84975d791ad72a5ec11d58fe1b0bd117c14ec14f82 |
m2_beats |
BEATs strong + mel-CNN | m2_beats_ema_step0007776.pt |
bd8b426f68a468186966c916f7f3fd01742305d6ce5051d6262f8fdea25d23de |
m3_wavlm_large |
WavLM-large + mel-CNN | m3_wavlm_large_ema_step0009000.pt |
01ea580186e65c705f5c0140482305b33fb745ece0bf657662d572abaab676bd |
m5_w2vbert |
w2v-BERT 2.0 + mel-CNN | m5_w2vbert_student_step0010000.pt |
75e770db43df5ca1c6535190bcfbd1665c6a3c0f8307a9f792b6671d601f7733 |
m6_atst_colab, m7_atst_s77, m8_atst_layermix |
ATST-F strong + mel-CNN | test-set probabilities in caches/ |
see the cache .json sidecars |
caches/<member>/{val,blind}.npz hold float16 sigmoid probabilities at 20 ms for all twelve members on the
validation and official test sets, with .json sidecars recording the checkpoint, step, window (10 s), hop (5 s),
floor (0.3) and TTA shifts. Together with the released bank detections in the GitHub repository, they regenerate
the submitted predictions of 995227 byte for byte. configs/clean_plus_a30.json is the final fusion and decoder
configuration.
Usage
git clone https://github.com/dwivediishivam/indoml26_track1.git && cd indoml26_track1
pip install -r requirements_inference.txt
python checkpoints/download_checkpoints.py --caches # regenerate the final submission
python main.py TEST_AUDIO_DIR OUT_DIR --mode replay
python checkpoints/download_checkpoints.py --all # weights for inference on new audio
python main.py INPUT_DIR OUT_DIR
- Input: 16 kHz mono WAVs.
- Output:
predictions.jsonl, one line per clip with{"clip_id", "events": [{"onset", "offset"}]}in seconds, plus a Codabench ZIP.
Pretrained dependencies and licence
Encoders:
- ATST-F strong and BEATs strong (PretrainedSED release v0.0.1, https://github.com/fschmid56/PretrainedSED)
microsoft/wavlm-base-plus,microsoft/wavlm-large,facebook/w2v-bert-2.0
Training data: Vaani Noise Event Dataset (Gold/Silver/Bronze tiers, CC BY 4.0) and the IndoML 2026 validation set.
Released for non-commercial research use under CC BY-NC 4.0.