Back to blog

Blog

OdishaVox: A Dedicated Odia Speech Corpus and Recognition Model

@openodia
speechmodelsdatasets

What happened

OdishaVox was created as a focused spin-off from BharatVox — a dedicated Odia speech corpus and automatic speech recognition (ASR) model. The project includes curated Odia audio data and a fine-tuned Whisper model achieving competitive word error rates.

Why it matters

General multilingual models often underperform on low-resource languages like Odia. OdishaVox proved that focused, community-built datasets can produce models that rival or exceed commercial offerings for specific languages.

Technical details

The model is based on OpenAI's Whisper architecture, fine-tuned on hundreds of hours of Odia speech. It supports both transcription and translation tasks, and can run on consumer hardware with 4GB+ VRAM.

Links