Reverie Launches India-Focused STT Model, Reports 1.5x Faste…
Analytics India Magazine (Mohit Pandey)
Reverie Language Technologies, a veteran in Indian-language AI, marked its 16th anniversary with the launch of a new Speech-to-Text (STT) model built to decode India’s multilingual chaos, from Hinglish to the several dialects.
The model, which logged 30 lakh API calls over the last year, is tuned to India’s real speech, full of code-switching, mixed languages, and regional quirks that global systems often fail to catch.
In independent tests against Deepgram for voice agent use cases, Reverie’s model scored about 4.2% higher in accuracy and 1.5x faster in response times, putting it among the most capable systems for Indian users.
The company says the system’s strength lies in how it handles what others miss. It can pick up numbers whether said in English (“twenty-three”), Hindi (“तेईस”), or a mix which is crucial for banks and call centres. It correctly identifies names from across India’s linguistic map, accounting for spelling and pronunciation differences that often throw global models off.
It also recognises geographic names, from Tier 1 cities to small towns without losing context.
Beyond Hinglish, Reverie has also released a family of STT models covering Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam, Assamese, Odia, and Punjabi. Each is trained separately on large datasets of regional voices and accents, creating purpose-built systems that reflect how people actually speak in those languages.
“Our R&D has always focused on India-specific language challenges,” Pranjal Nayak, R&D head at Reverie, said. “This Hinglish model is a direct outcome of that — it understands how Indians say numbers, how we mix English with Hindi, and how accents vary even within the same sentence. It makes AI agents sound less robotic and more human.”
The company has already seen adoption across industries. A major financial services firm deployed Reverie’s STT engine to process over 15,000 multilingual debt collection calls, running 100 concurrent threads with high accuracy on numbers and payments.
What sets the model apart is not just its accuracy, but its cultural understanding. Reverie trained it on live conversational data to capture emotional tone, phrasing, and natural language switches that happen mid-sentence — from English to Hindi to regional languages. The engine preserves meaning instead of just spitting out literal text.
The model is now live on Reverie’s API platform for enterprises, available on both cloud and on-prem setups. It supports add-ons like domain-specific language packs, numeric and name disambiguation, and hot-word boosting — all configurable through the same API.
The post Reverie Launches India-Focused STT Model, Reports 1.5x Faster Than Deepgram appeared first on Analytics India Magazine.
Generated by RSStT. The copyright belongs to the original author.