Most ASR models were not built for India.
Indian enterprise calls happen over GSM lines compressed to 8kHz. On factory floors and in bank branches where ambient noise is not an exception but a constant. In sentences that shift between Hindi and English mid-clause. In accents from Lucknow, Coimbatore, Patna, and Surat that no studio corpus has attempted to capture.
The consequence of this mismatch is not a higher error rate. It is operational failure. Consider a single sentence from a loan origination call.
14 million hours. 14x more than any competitor.
Gnani Prisma v2.5 is trained on 14 million hours of proprietary Indic speech — 14x more than any competing model. The corpus spans 12 languages with real dialect variation, ambient noise, and natural code-switching baked into the training distribution. This is a significant milestone towards ensuring pan-India coverage with Sovereign AI models.
Indic speech data
any competitor
real dialect variation
“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.”
“Consider a single sentence: ‘mereko loan chahiye thees karod rupaye ka. Kab tak milega.’ A model that mishears ‘thees’ as ‘theen’ produces a 10% WER on that utterance. In a loan origination call, that single error misrepresents the amount by thirty crore rupees. At 30 million calls per day, error rates are not a quality metric. They are a huge business risk.”
“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. Architectural improvements through post-training optimization have doubled throughput over the previous version without accuracy loss.”
Interactive benchmark explorer
All results from independent evaluations on publicly available datasets. Same audio input evaluated across all four models. Select any language, dataset, and metric.
27 wins out of 49 evaluations.
Across all 9 languages and all Standard WER datasets, Prisma v2.5 ranked #1 in 27 of 49 benchmark evaluations — more than ElevenLabs, Microsoft, and Sarvam combined.
Win distribution across 49 evaluations
| Matchup | Prisma wins | Avg WER delta | Best language | Worst language |
|---|
Every number, all at once.
All 9 languages × all datasets × all 4 models. Winner highlighted per cell. Hover any cell for the full four-way comparison.
When audio degrades, competing models collapse.
Average WER degradation moving from clean (kathbath) to noisy (kathbath_noisy) conditions across all 9 languages. The gap is structural, not marginal.
Not all benchmarks are equal.
Average WER across all 4 models, ranked hardest to easiest. Context for why Gramvaani results matter far more than IndicTTS when evaluating production readiness.
Prisma starts clean. The gap only widens.
Standard WER vs LLM WER for each model across 9 languages on kathbath. Prisma starts so clean that its lead increases post-correction on most languages.
South Indian languages require a different metric.
Tamil, Telugu, Kannada, and Malayalam are agglutinative. WER is structurally broken for these scripts (NAACL 2025, Jain & Bhowmick, Springer 2025). On CER, Prisma v2.5 leads across all four South Indian languages on both clean and noisy conditions.
The Hindi that 500 million people actually speak.
Gramvaani captures spontaneous, conversational Hindi from semi-urban and rural speakers — the only benchmark that approximates how Hindi is spoken by the majority of Hindi speakers in India. On this benchmark, the performance gap is the widest in the entire dataset.
Gramvaani
5.5pp worse
4.4pp worse
On standard Hindi (kathbath), Prisma v2.5 achieves 7.6% WER — the lowest of any model tested. Under noise it holds at 7.9%, a degradation of just 0.3pp where competitors lose 1–3pp.
What the WER gap costs at your call volume.
Enter your daily call volume and per-event cost assumptions. The calculator uses actual WER deltas from benchmark data to estimate misrouted calls and escalations generated by each competitor versus Prisma v2.5.