Jitesh K Prasad PRO
jiteshkprasad
·
AI & ML interests
Building udgamlabs.com
Recent Activity
posted an update about 2 hours ago
Shruti: speech-to-text for Hindi, English and Hinglish in one 29.5 MB file, with open weights (Apache-2.0).
• About 100× smaller than Whisper large-v3, and more accurate on Hindi (word error rate, same clips): Monsoon 20.46 vs 33.29, FLEURS 20.14 vs 28.88, Lahaja 22.87 vs 31.19. Whisper is better on English.
• The fewest Hinglish mistakes of the open models we measured (MUCS 2021): 35.2 vs 57.5 and 64.5.
• Runs offline in the browser, on your phone or laptop.
Built on Whistle by https://huggingface.co/Cactus-Compute. The full results table and the limits are on the card.
Demo: https://huggingface.co/spaces/udgamlabs/Shruti
Weights: https://huggingface.co/udgamlabs/Shruti liked a Space about 22 hours ago
udgamlabs/Shruti updated a Space about 22 hours ago
udgamlabs/READMEOrganizations