Research
Research: Swecha Gonthuka Telugu ASR
An overview of the research behind the Swecha Gonthuka Telugu speech recognition system. A full paper covering methodology, findings, and evaluation design is in preparation.
Overview
Telugu is spoken by over 80 million people but remains underrepresented in open ASR research. Existing multilingual models often treat Telugu as a low-resource tail language, resulting in weak recognition across native speakers and dialectical variation. Swecha Gonthuka starts from a community-collected, Telugu-first dataset rather than adapting a generic multilingual model. The aim is a recogniser that performs across the range of speakers in the training distribution — not only standardised studio speech.
Research Paper
Releasing SoonA formal write-up is in preparation. It will cover:
- Methodology
- Model Training
- Evaluation Design
- Findings
- Limitations
- Future Directions
Use the Model
Model weights and usage documentation are available on Hugging Face.
View on Hugging FaceDocumenting the Path to Linguistic AI
Sharing our technical findings, methodologies, and insights as we build for the South Indian AI ecosystem.
Our Approach to Research
At ViswamAI, we believe that building meaningful technology requires rigorous documentation and an open exchange of ideas. As we develop solutions tailored for low-resource languages, we are committed to analyzing our processes, tracking our benchmarks, and publishing our learnings.
- • Methodologies for ethical, community-driven data collection.
- • Technical Insights gained from training models for regional dialects and cultural context.
Future Publications
Releasing SoonWe are in the early stages of our research journey. As our projects mature and yield verified results, this space will serve as the central repository for our technical reports, preprints, and peer-reviewed publications.
- Collect data - steps
- Task decision - ? used using publication - validated by
- Train-dev- test
- Model
- Comparision/ Bench marking