Research Papers

Research Papers

Peer-reviewed publications and preprints from the Viswam AI research team, covering speech recognition, language modeling, datasets, and sustainable AI.

Swecha Gonthuka: A Community-Driven Telugu ASR Dataset

2025

Viswam AI TeamarXiv preprint

We present Swecha Gonthuka, a large-scale community-driven Telugu speech dataset collected through grassroots contributions. The dataset contains over 15 million audio samples across diverse dialects, demographics, and recording conditions. We benchmark several ASR architectures and establish strong baselines for Telugu speech recognition.

Low-Resource Language Modeling for Telugu: Challenges and Approaches

2025

Viswam AI TeamIn preparation

We investigate language modeling approaches for Telugu, a morphologically rich low-resource language. We compare tokenization strategies, model architectures, and training techniques, demonstrating that character-level and subword-level models with language-specific preprocessing achieve competitive perplexity despite limited data.