Is Shruti state of the art?Pending benchmarks.Announced Oct 19, 2026; no independent results yet.
Udgam Labs' speech-to-text model for Hindi, English and Hinglish in one 29.5 MB file that runs offline on a phone, laptop or in a browser tab. Built on Cactus Compute's Cactus Whistle: Udgam fine-tuned all 55M weights on 2,732 hours of openly licensed Indian speech and added 714 Devanagari vocabulary pieces. Apache-2.0.
The details so far
- Lab
- Udgam Labs
- Announced
- Oct 19, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 55M (29.5 MB file, 4-bit weights / 8-bit activations)
- Base model
- Cactus-Compute/whistle (Cactus Whistle, Apache-2.0); fine-tuned derivative by Udgam Labs, not trained from scratch
- Training data
- 2,732 hours of openly licensed Indian speech (IndicVoices, Vaani, Shrutilipi, Common Voice 17 and others)
- Languages
- Hindi in Devanagari, English, and Hinglish with English words kept in Latin script
- Runtime
- One .cact file for Cactus's open-source needle engine (Mac, Linux, Windows, Android, iOS, WebAssembly); no server needed
- Input
- 16 kHz mono audio; one pass up to 30 s, longer audio in 30 s windows
- License
- Apache-2.0 (weights), keeping Cactus's notice
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: word error rate on Hindi of 20.46 (Monsoon), 19.92 (FLEURS) and 26.28 (Lahaja), against Whisper large-v3's 33.29, 28.88 and 31.19 in Udgam's harness; Udgam reports Whisper large-v3 clearly ahead on English.
Lab claim: 35.21 WER on MUCS 2021 Hinglish, the fewest errors among the open models Udgam measured; the larger Hindi-only Koyal-hi-120M and IndicConformer are more accurate on pure Hindi.
Lab claim: about 1 second per sentence and 80 MB peak memory offline on a Samsung Galaxy F62.
Sources
LABShruti release post (Udgam Labs blog)udgamlabs.com/blog/shruti · Oct 19, 2026LABShruti model page (Udgam Labs)udgamlabs.com/models/shruti · Oct 19, 2026REPOudgamlabs/Shruti model card (Hugging Face)huggingface.co/udgamlabs/ShrutiREPOudgamlabs/shruti (GitHub)github.com/udgamlabs/shrutiTracking continues until an independent board lists it.