sajidjavid/CodeAMU-IndoML2026-Track1

CodeAMU — IndoML 2026 Datathon, Track 1 (Noise Event Detection): detector checkpoints

The 24 frame-level noise-event detectors of team CodeAMU's Track 1 system, 1st place on the final (Phase 2) leaderboard: Combined 1.6173 (Event F1 0.7895 + Segment Dice 0.8278).

Code, inference pipeline, training configurations and technical report: https://github.com/sajidjavid222/CodeAMU_Track1 (python scripts/download_checkpoints.py there downloads this repository into checkpoints/).

Authors: Shah Ahmed Shakir Abu Asim Khan (Aligarh Muslim University), Sajid Javid (IIIT Delhi), Osama Zaheer (Binary Semantics), Ahraz Shamim (Vecmocon Technologies). Contact: sajidj@iiitd.ac.in

Contents

detectors/<name>/best.pt — PyTorch checkpoint (model state dict, training args, decode_params, frame_sec); detectors/<name>/best.json — validation score of the checkpoint. Load with inference/infer.py of the GitHub repository (it rebuilds the architecture from args; hr_* detectors need SED_TIME_POOL=1).

Every detector outputs per-frame noise presence, onset and offset probabilities (plus 7 category and clip-level tag heads) on a 20 ms grid (10 ms for hr_*). The final system routes each clip to a natural or a synthetic ensemble of these detectors and post-processes the result (see the report).

Detector	Encoder	Frame	Training data	Initialised from	Scored on	Combined
beats_full_gold	BEATs + Whisper medium	20 ms	all gold + all validation (13,612)	rb_beats_gold	official validation (2,501)	1.5984 ¹
bg_lg3_syn_gold	Whisper large-v3	20 ms	all 11,111 gold clips	bg_lg3_syn	official validation (2,501)	1.4186
bg_lg3_syn	Whisper large-v3	20 ms	all tiers + 11,762 synthetic clips	—	official validation (2,501)	1.4242
bg_sml_syn	Whisper small	20 ms	all tiers + 11,762 synthetic clips	—	official validation (2,501)	1.4022
cb_big	CRNN (128-mel)	20 ms	all tiers + 11,762 synthetic clips	—	official validation (2,501)	1.4136
cb_s7_gold	CRNN (128-mel)	20 ms	all 11,111 gold clips	cb_s7	official validation (2,501)	1.4128
cb_s7	CRNN (128-mel)	20 ms	all tiers + 11,762 synthetic clips	—	official validation (2,501)	1.4076
gv_lg3_A	Whisper large-v3	20 ms	gold + validation half B	bg_lg3_syn	validation half A (1,259)	1.4371
gv_lg3_B	Whisper large-v3	20 ms	gold + validation half A	bg_lg3_syn	validation half B (1,242)	1.4468
gv_med_A	Whisper medium	20 ms	gold + validation half B	ov_med_syn	validation half A (1,259)	1.4369
gv_med_B	Whisper medium	20 ms	gold + validation half A	ov_med_syn	validation half B (1,242)	1.4351
gv_sml_A	Whisper small	20 ms	gold + validation half B	bg_sml_syn	validation half A (1,259)	1.4220
gv_sml_B	Whisper small	20 ms	gold + validation half A	bg_sml_syn	validation half B (1,242)	1.4210
hr_med_syn_gold	Whisper medium	10 ms	all 11,111 gold clips	hr_med_syn	official validation (2,501)	1.3810
hr_med_syn	Whisper medium	10 ms	all tiers + 11,762 synthetic clips	—	official validation (2,501)	1.3973
lg2_full_gold	Whisper large-v2	20 ms	all gold + all validation (13,612)	v9_lg2_gold	official validation (2,501)	1.6190 ¹
lg3_full_gold	Whisper large-v3	20 ms	all gold + all validation (13,612)	v9_lg3_s21	official validation (2,501)	1.6547 ¹
ov_med_syn	Whisper medium	20 ms	all tiers + 11,762 synthetic clips	—	official validation (2,501)	1.4062
rb_beats_gold	BEATs + Whisper medium	20 ms	7,112 gold clips (3,999 held out)	rb_beats	gold hold-out (3,999)	1.0840
v2_crnn_gold	CRNN (64-mel)	20 ms	10,000 gold clips (hold-out excluded)	v2_crnn	gold hold-out (1,111)	1.0697
v4_wavlm_gold	WavLM Base+	20 ms	10,000 gold clips (hold-out excluded)	v4_wavlm	gold hold-out (1,111)	1.0607
v8_whis_lg_gold	Whisper large-v3	20 ms	10,000 gold clips (hold-out excluded)	v8_whis_lg	gold hold-out (1,111)	1.1102
v9_lg2_gold	Whisper large-v2	20 ms	10,000 gold clips (hold-out excluded)	v9_lg2	gold hold-out (1,111)	1.1037
v9_lg3_s21	Whisper large-v3	20 ms	all tiers minus a 1,111-clip gold hold-out	—	gold hold-out (1,111)	1.1119

¹ in-sample: the validation clips are part of the training data of the *_full_gold models. Scores on different sets are not comparable (the gold hold-outs contain natural clips only and are harder).

Pretrained encoders

Encoder	Source (revision)
Whisper small / medium / large-v2 / large-v3	openai/whisper-* (973afd2 / abdf7c3 / ae46427 / 06f233f)
WavLM Base+	microsoft/wavlm-base-plus (4c66d48)
BEATs iter3+ AS2M (cpt2), frozen	WeiChihChen/BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2 (172e407), mirror of Microsoft UniLM BEATs

Training data

ARTPARK-IISc/Vaani-Noise-Event-Dataset (90,635 clips; gold / silver / bronze annotation tiers), synthetic clips built from its gold tier, and the official IndoML 2026 Track 1 validation release (2,501 clips).