Oriserve open-sources India-focused AI speech model fine-tuned on Whisper

Oriserve has released Whisper–Hindi2Hinglish-Apex, an open-source automatic speech recognition (ASR) model fine-tuned on OpenAI’s Whisper and adapted for Hindi, Hinglish and Indian-accented English. The model is now available on Hugging Face.

The launch targets a critical gap in AI speech systems for India, where global ASR models typically show accuracy drops on code-mixed speech, strong regional accents and noisy telephonic audio. While Whisper remains a widely used multilingual ASR model, its performance declines on non-standardised and hybrid Indian datasets.

Whisper–Hindi2Hinglish-Apex retains Whisper’s architecture but is trained on more than 1,000 hours of conversational Indian audio, including call-centre recordings and mixed Hindi–English speech. The fine-tuning is intended to improve accuracy in enterprise conditions, where accent diversity and low-quality audio are common.

The model contains about 800 million parameters. Oriserve says it offers:

•faster inference than larger Whisper variants,

•a 42% improvement over Whisper’s baseline on internal benchmarks, and

•stronger handling of accented, hybrid and noisy audio.

“India needs speech models trained on its own linguistic data, not just adapted global datasets,” said Anurag Jain, co-founder, Oriserve. “Open-sourcing this model enables developers to build AI systems aligned with real Indian audio environments.”

Co-founder Maaz Ansari said the effort is aimed at reducing dependence on proprietary cloud ASR systems and enabling on-premise or hybrid deployments across sectors such as BFSI, telecom, healthcare and education.

This is Oriserve’s third release in its open-source AI series. The company plans to extend fine-tuned Whisper variants to Marathi, Gujarati, Tamil, Telugu, Kannada, Malayalam, Bengali and Punjabi as part of its larger multilingual AI roadmap.