Skip to navigation
Speech to TextBenchmarks

Hindi

Pulse Hindi accuracy across public datasets.
View as Markdown

Hindi - per-dataset

WER on four Hindi datasets. Lower is better. Results are from the Pulse streaming (WebSocket) API. - means the provider was not evaluated on that dataset. For all nine Indian languages, see Indic languages.

DatasetPulseDeepgramScribeSonioxSarvam code-switchSarvamCartesia
FLEURS6.17%14.08%10.67%5.54%8.62%4.18%9.06%
IndicDiarBench28.19%36.05%71.73%24.25%30.18%25.14%29.63%
Lahaja7.92%18.51%32.21%10.92%16.44%9.02%16.74%
Svarah10.06%24.91%13.28%50.23%9.52%102.86%10.62%

Hindi - Vistaar

WER on six of the Vistaar Hindi benchmark datasets. Vistaar also includes FLEURS, which is reported once, in the table above.

DatasetPulseDeepgramScribeSonioxSarvam code-switchSarvamCartesia
Kathbath5.68%13.03%5.98%5.42%7.90%3.74%8.68%
Kathbath (hard)6.95%14.49%10.88%6.04%8.28%4.12%9.13%
Common Voice6.62%18.42%16.62%7.27%11.38%5.00%12.38%
IndicTTS3.63%7.04%22.63%2.81%5.24%2.62%5.38%
MUCS3.22%13.49%21.65%7.13%4.82%4.42%7.67%
Gram Vaani14.62%30.28%47.06%16.06%15.04%14.04%15.58%