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

Gnani AI, a voice artificial intelligence startup based in Bengaluru, has launched its latest speech-to-text model, Prisma v2.5. This new version claims to outperform competitors in recognizing speech across various Indian languages, achieving top rankings in eight out of nine languages during rigorous tests involving real-world and noisy acoustic environments. According to the company, Prisma v2.5 has demonstrated a 15% reduction in word error rates for rural Hindi dialects and an 18% decrease for Dravidian languages, surpassing models from well-known firms like ElevenLabs, Deepgram, and Sarvam AI. These assertions are supported by benchmarks, including the Gramvaani dataset, which focuses on capturing the diverse speech patterns prevalent in semi-urban and rural India. Typically, speech recognition technology excels in ideal studio conditions but struggles with challenges such as background noise, varied accents, and the frequent switching between English and Indian languages. In response, Gnani AI has trained Prisma v2.5 using 14 million hours of unique speech data that encompasses 12 languages, taking into account regional dialects, ambient sounds, and the phenomenon of code-switching. The model is designed for use in vital sectors such as banking, financial services, insurance, and healthcare, where precise transcription is crucial for maintaining compliance and effective customer relationship management. Notably, Prisma v2.5 supports seamless code-switching between Hindi and English, Tamil and English, among other combinations, without the need for specific language tags. Bharath Shankar, co-founder and chief product and engineering officer at Gnani AI, highlighted that the model is adept at handling audio from both GSM and voice-over-internet-protocol networks. This native support ensures effective communication, even in challenging conditions. Shankar further noted that the optimization carried out after training has effectively doubled the model’s throughput compared to its predecessor, all while maintaining accuracy. Ganesh Gopalan, co-founder and CEO of Gnani AI, emphasized the importance of training voice AI systems on the nuances of Indian speech, stating, "Accents, noise, code-switching, and compressed telephony audio are not exceptions in India; they are the standard."