Gnani.ai Launches New Speech-to-Text Model, Claims to Outperform Global Rivals

Indian voice AI startup Gnani.ai has launched Prisma v2.5, a speech-to-text model that the company claims delivers the highest speech recognition accuracy for Indian languages across multiple independent benchmarks. 

The release confirms AIM’s earlier exclusive report that Gnani.ai was developing a new speech-to-text model for enterprise use, among five new AI models.

The company said Prisma v2.5 is available through its API and is targeting enterprise deployments across sectors, including banking, financial services, insurance, healthcare, retail, and telecom.

The Bengaluru-based startup said the model ranked first in 27 of 48 evaluations and led in eight of nine Indian languages tested on datasets including Kathbath, Kathbath Noisy, Gramvaani, CommonVoice, FLEURS, IndicTTS, MUCS 2021, and the IITM Evaluation Set.

According to the company, Prisma v2.5 achieved 15% lower word error rates (WER) for rural Hindi dialects and 18% lower WER in noisy Dravidian language environments compared with competing global and Indic speech recognition models. 

It achieved the lowest WER on the Gramvaani benchmark, which captures conversational speech from semi-urban and rural Hindi speakers, recording a 22% WER compared to 27.5% for ElevenLabs, 26.4% for Microsoft, 23.4% for Sarvam, and 34% for Deepgram, per the company.

The model was trained on 14 million hours of proprietary speech data spanning 12 Indian languages, which Gnani.ai claimed is 14 times more than any competing model has publicly disclosed.

The company said the model supports word-level code-switching between language pairs such as Hindi-English and Tamil-English without requiring language tagging. It also includes native support for GSM and VoIP audio codecs, enabling transcription of low-quality telephony audio commonly used in tier-2 and tier-3 cities. 

Gnani.ai added that post-training optimisations doubled inference throughput compared with the previous version without reducing accuracy.

“Most ASR models are built for ideal studio conditions. Indian calls happen over compressed network lines, in at least two languages inside a single sentence, in accents no studio corpus has ever captured. Gnani Prisma v2.5 is built for that reality,” said Ganesh Gopalan, CEO and Co-founder of Gnani.ai, in a statement.

The company said the model addresses transcription errors involving numerals, alphanumeric strings, named entities, and short utterances, which are common in enterprise call centre applications.