Agharia Kasimali — ASR Models (Annam.ai)

This space acts as a centralized verification archive for the automatic speech recognition (ASR) modules developed during my research internship at Annam.ai (Government of India & IIT Ropar Initiative) from August to October 2025.

The work focuses on multilingual ASR for 10+ Indian languages, analyzing language-specific acoustic characteristics (Indo-Aryan, Dravidian families) and speaker variability across regions and demographics, along with fairness and scalability considerations.

🎙️ ASR Models & Artifacts

Click on any link below to review model demos, fine-tuning experiments, and pipeline components directly in your browser:

1. Multilingual ASR Demos

NEW-ASR-VOX — Multilingual ASR demo with VoxLingua107-based models.
NEW-ASR-VOXLINGUA — Cross-lingual ASR evaluation across Indian languages.
ASRLID — Joint ASR + Language Identification prototype.
Whisper Malayalam Fine-tuning — Low-resource ASR adaptation for Malayalam.

2. Fine-tuning & Experiment Logs

FINETUNINNG — ASR fine-tuning experiments and training logs.
FINETUNNING WHISPER — Whisper-based fine-tuning runs for Indian languages.
XLS-R1B — XLS-R 1B parameter model experiments for multilingual ASR.

3. Pipeline Components & Utilities

Whole Pipeline — End-to-end ASR inference pipeline (audio → text).
INDICTRANS2 — IndicTrans2 integration for ASR output translation.
IndicXlitF — Romanized-to-native script conversion for ASR outputs.

🔒 Note: This archive is intended for research verification and demonstration purposes. Model weights, datasets, and internal configurations follow the respective institutional and licensing guidelines.