asymptoticlearning/IndoML2026-Track2-Arcturus

Team Arcturus: IndoML 2026 Datathon, Track 2 (Noise Event Removal) checkpoints

Final checkpoints of our Track 2 system, which removes noise events (breathing, horns, animals, alarms, music, machines, crying babies) from Vaani field recordings while preserving the speech for the mandated SraVaani-1.0 recogniser.

Final standing: 8th (leaderboard name: asymptoticchase). Phase 2 Combined 16.6113 (SI-SDR 13.8237 dB; ΔWER +2.7876 points, of which +7.8507 on synthetic and +0.4012 on natural clips).

• Code, training scripts and report: https://github.com/Abhilash1602/IndoML2026-Track2-Arcturus

• File: Arcturus_Track2_checkpoints.tar (1.28 GB) contains checkpoints/*.pt:

  • 18 fine-tuned separators: 13 MossFormerGAN-SE-16K, 3 MossFormer2-SE-48K, 2 MossFormer2-SS-16K;
  • 4 learned time-frequency fusion networks.

  Roles, sizes and sha256 checksums are in CHECKPOINTS.md.

Usage

From a clone of the GitHub repository:

bash scripts/download_checkpoints.sh https://huggingface.co/asymptoticlearning/IndoML2026-Track2-Arcturus/resolve/main/Arcturus_Track2_checkpoints.tar
python inference/infer.py <input_audio_dir> <output_dir>

The script unpacks the archive into checkpoints/ and verifies every checksum.

Licence and provenance

The separators are fine-tuned from the Apache-2.0 ClearerVoice-Studio checkpoints (alibabasglab/MossFormerGAN_SE_16K, MossFormer2_SE_48K, MossFormer2_SS_16K) and are distributed under Apache-2.0. The fusion networks are our own work. The mandated ASR, ARTPARK-IISc/SraVaani-1.0, is not included.

Team

Team Arcturus: Abhilash Jacob Mathew (team leader, HCLTech, abhilashjacob16@gmail.com) and Aman Kumar (KIET GAZIABAD, aman.kumar.cse@kiet.edu).