Gnani.ai Releases Vachana STT Trained on 1M Hours of Voice Data

As voice becomes a primary interface for customer engagement, automation, and analytics in India, speech recognition is emerging as foundational infrastructure rather than an application-layer add-on. Bengaluru-based Gnani.ai has entered this arena with the launch of Vachana STT, a foundational Indic speech-to-text model trained on over 1 million hours of real-world voice data, released under the IndiaAI Mission.

Vachana STT is the first public release in Gnani.ai’s upcoming VoiceOS, a unified voice intelligence stack that spans speech recognition, synthesis, understanding, and orchestration. The company positions the model as core infrastructure, designed for production environments rather than stitched together from multiple APIs.

From Language Coverage to Production Reality

India’s speech AI challenges extend beyond linguistic diversity. Audio quality varies widely across channels, accents shift within regions, and real-world conversations are rarely clean or controlled.

Vachana STT is trained on proprietary multilingual datasets spanning 1,056 domains, allowing it to operate across noisy, omnichannel environments without additional fine-tuning. According to Gnani.ai, this enables consistent performance across use cases where speech accuracy directly impacts automation outcomes, compliance monitoring, and customer experience.

The model forms a critical layer of VoiceOS, which aims to offer a sovereign, end-to-end voice infrastructure built from first principles rather than assembled components.

Benchmarking Indic Speech at Scale

Across evaluations on publicly available datasets and live omnichannel audio, Vachana STT reports 30–40 per cent lower word error rates (WER) on low-resource Indian languages and 10–20 per cent lower WER on the top eight languages spoken in India, compared with leading providers.

Languages covered include Hindi, Bengali, Gujarati, Marathi, Punjabi, Tamil, Telugu, Kannada, Malayalam, Odia, Assamese, and several others. Gnani.ai says detailed benchmarking and comparative evaluation reports are available to enterprises on request.

The emphasis on low-resource languages signals a shift away from selective localisation toward broader foundational coverage.

Built for High-Volume Enterprise Workloads

Vachana STT is already deployed across BFSI, telecom, customer support, and large-scale voice automation systems, processing approximately 10 million calls per day. The platform supports both real-time and batch transcription and integrates via enterprise-grade APIs.

Latency remains a critical requirement in these environments. Gnani.ai reports p95 latency of around 200 milliseconds, even under high concurrency, enabling use cases such as agent assist, speech analytics, and compliance workflows.

The model is optimised to handle compressed audio ranging from 8 kbps to 64 kbps, making it suitable for telephony as well as digital channels with variable network quality.

Vachana STT’s release follows Gnani.ai’s selection under the IndiaAI Mission, where the Government of India has identified a limited group of startups to build sovereign foundational AI models from India.

Rather than focusing on application-layer experimentation, the selection underscores an emphasis on core AI infrastructure, particularly in areas where data, language, and deployment conditions are uniquely Indian.

“Speech recognition in India is not a localisation problem. It is a foundational systems problem,” said Ganesh Gopalan, co-founder and CEO of Gnani.ai.

“Vachana STT is built as core infrastructure, trained on how India actually speaks, and designed to operate across channels, not just telephony. Being selected under the IndiaAI Mission reinforces our belief that foundational AI models must be built from India, with production reality at the centre.”

Availability and Access

Vachana STT is available immediately via API access for enterprise customers. Early adopters receive 100,000 free minutes of usage. Enterprises can request benchmarking data, technical evaluations, or API access directly from Gnani.ai.