Gnani AI launches Prisma v2.5, claims better accuracy than Sarvam, ElevenLabs on Indian speech

Bengaluru-based voice artificial intelligence startup Gnani AI has launched Prisma v2.5, its latest speech-to-text model. The company claimed that Prisma v2.5 ranked first in eight of nine Indian languages across real-world and acoustically noisy speech-recognition benchmarks. 

It said the model recorded 15% lower word error rates for rural Hindi dialects and an 18% reduction across Dravidian languages compared with competing models from ElevenLabs, Deepgram and Sarvam AI.

The claims are based on benchmarks cited by Gnani AI, including Gramvaani, a dataset designed to capture speech patterns from semi-urban and rural India.

Speech recognition systems typically perform well on studio-quality recordings but can struggle with compressed telephone audio, background noise, regional pronunciation and conversations that switch between English and Indian languages.

Gnani AI said Prisma v2.5 was trained on 14 million hours of proprietary speech data spanning 12 languages. The training corpus includes regional dialects, ambient noise and code-switching, according to the company. 

The model is aimed at enterprises in sectors such as banking, financial services and insurance, healthcare and insurance, where transcription errors involving numbers, names or technical terms can affect compliance records, customer relationship management systems and agent-assistance tools.

The company said Prisma v2.5 supports code-switching between Hindi and English, Tamil and English, and other regional language combinations without requiring language tags. It also supports audio transmitted through GSM and voice-over-internet-protocol networks.

“CODEC handling for GSM and VoIP is native. Code-switching across Hindi-English, Tamil-English, and regional-English pairs works at the word level without language tagging,” said Bharath Shankar, co-founder and chief product and engineering officer at Gnani AI.

Shankar said post-training optimisation had doubled the model’s throughput compared with the previous version without reducing accuracy.

Ganesh Gopalan, co-founder and chief executive officer of Gnani AI, said voice AI deployments in India frequently encounter problems because models are not trained on the way Indians speak in real-world conditions.

“Accents, noise, code-switching, compressed telephony audio, these are not edge cases in India; they are the norm,” Gopalan said.