Congratulations to the Top 8 teams on the Track 1 (Noise Event Detection) Phase 1 leaderboard. The leaderboard is NOT reset — all teams continue into Phase 2.

Track 1 test data for Phase 2 is the SAME as Phase 1 — that is why the leaderboard is NOT reset

For Track 1 (Noise Event Detection), the Phase 2 test set is identical to the Phase 1 test set. Nothing has changed — you do not need to download any new data, and your existing pipeline will continue to work as-is. Because the data is unchanged, there is no refresh/reset of the Track 1 leaderboard — all Phase 1 scores carry over and remain directly comparable with Phase 2 submissions.

Track 2 teams must submit their model and inference code with the final Phase 1 submission

To ensure the integrity and fairness of the Track 2 evaluation — and in response to concerns raised regarding submissions — we are introducing an additional verification audit as a precautionary measure. Track 2 participants are requested to provide their executable model inference code and required model files so that submitted models can be independently reproduced and verified by the organizers on the evaluation data.

Along with your final Phase 1 submission, please share a ZIP package containing:

• Model files — the trained model/checkpoint and all files required to load and run the model, in the format supported by your framework (e.g., PyTorch, NeMo, etc.).
• Requirements file — requirements.txt, environment.yml, or an equivalent dependency specification.
• Inference script — a main.py script (or equivalent executable code) that runs inference. Please name the entry-point file main.py so the main file is easy to identify.
• README — clear instructions for setting up the environment and running the inference code, including the primary contact's mobile number and email.

You do NOT need to submit the SraVaani ASR model. It is simply a Hugging Face call (ARTPARK-IISc/SraVaani-1.0) — you only need to provide the script that calls the Hugging Face model, and the audio is transcribed with the same ASR. This keeps the ZIP lightweight; only your enhancement model and inference code are required.

The inference script must take exactly two arguments:

1. Input folder path — containing all the test audio files.
2. Output folder path — where the inference results should be saved.

After inference, the enhanced audio files must be saved in the output folder with the same file names as the input audio files, along with the JSON transcript file in the same format as the submission format specified on Codabench.

Naming (important): Name the ZIP file and its top-level folder as your Team_name (e.g., Team_name.zip containing a Team_name/ folder). This lets us clearly identify multiple submissions and, if a reproduction issue arises, revert each team individually.

Welcome to Datathon@IndoML 2026 - a research-oriented data science competition held in conjunction with IndoML 2026. Building on the success of previous editions, this year's datathon challenges participants to tackle noise event detection and removal in real-world Indic speech - a critical problem for inclusive, robust Automatic Speech Recognition across Indian languages.

The competition is organised into two tightly coupled tracks: Track 1 (Detection) - detect noise events with precise timestamps, effectively utilising data annotated at different levels - and Track 2 (Removal) - suppress detected events while preserving the underlying speech. The dataset is a curated subset of the Vaani corpus, consisting of ~154.6 hours of real-world Indic audio with three levels of annotation quality.

Top-performing teams will be invited to attend IndoML 2026 and present their solutions to leading researchers and professionals from academia and industry. These teams will also receive exciting cash prizes.

Most academic speech enhancement benchmarks are built around stationary noise (white, pink, café hum) or studio-recorded mixtures. Field recordings from rural and semi-urban India contain bursty, semantically rich events — a passing motorbike, a hen, a pressure-cooker whistle, a doorbell, a TV in another room — that current denoisers either smear over or treat as speech.

Two consequences follow:

• Downstream ASR fails disproportionately on speakers from these environments, amplifying the existing performance gap between high-resource and low-resource Indic languages.
• Standard sound-event detection (SED) datasets (AudioSet, DESED, DCASE) under-represent both Indian acoustic contexts and Indian speech, making transfer learning from them brittle.

This challenge targets that gap directly. It asks participants to treat noise as a first-class, labelled, time-localised object — and then to suppress it without harming the speech.

Vaani Corpus

A large-scale, openly released Indic speech dataset spanning multiple Indian languages, collected across districts of India. Learn more at vaani.iisc.ac.in or read the paper.

The dataset consists of three types of annotated noise events, totalling ~154.6 hours of training audio across 90,637 segments. An additional 10 hours of gold-standard audio will serve as the test set, with ground-truth timestamps held back for automated evaluation. Effectively utilising data annotated at different quality levels is a key part of the challenge.

Check out last year's edition: Datathon@IndoML 2025 - Evaluating LLM-Powered AI Tutors