Introduction

Udgam Labs just added Shruti Gujarati to its on-device Shruti speech-to-text family. The product page dates the drop 4 November 2026, and the Hugging Face card udgamlabs/shruti-gujarati was created 2026-11-04T15:25Z. One 29.7 MB W4A8 file (shruti-gujarati.cact, ~55.9M params) turns Gujarati + English speech into text offline — phone, laptop, or browser WASM — with no server.

Gujarati comes out in Gujarati script; English loanwords stay in Latin letters the way people actually type (દસ વાગ્યે party plot પર મળીએ). That code-mix behavior is the maker hook — same family pattern as the Telugu / Bengali / Tamil / Marathi cards we covered earlier today.

What Shipped

  • Vendor / product: Shruti Gujarati — Udgam Labs (sibling of Hindi, Marathi, Tamil, Bengali, Telugu)
  • Modality: speech-to-text / ASR (on-device)
  • Open vs closed: open weights, Apache-2.0 (built on Shruti Hindi → Cactus Whistle + needle engine)
  • Runnable today: pip install cactus-needle + hf download udgamlabs/shruti-gujarati shruti-gujarati.cact; CLI ./needle --model shruti-gujarati.cact --audio clip.wav; live page at udgamlabs.com/models/shruti
  • Concrete fact: on DiarBench code-mix (English words scored in English), Shruti Gujarati hits WER 28.5 vs SraVaani-1.0 32.4 (30× the file size) and IndicConformer Gujarati 44.9 (5×) — measured 2 November 2026 on the same clips/scorer
  • Device note: ~10.3× realtime and ~64 MB RAM on a 2021 Galaxy F62 (native engine); LibriSpeech test-clean 6.17 WER for English; 0 words invented on 100 silent clips

Run It

pip install cactus-needle
hf download udgamlabs/shruti-gujarati shruti-gujarati.cact config.json --local-dir .

from needle import Whistle

model = Whistle("shruti-gujarati.cact")
print(model.transcribe("clip.wav"))  # Gujarati / English auto-detect

Set NEEDLE_TELEMETRY=0 if you want the Python package quiet. Native binaries for Mac/Linux/Windows/Android/iOS + WASM live under Cactus-Compute/needle3.

Where It Beats (and Loses)

  • Code-mixed conversation win: fewer errors than SraVaani-1.0, IndicConformer Gujarati, Vosk Gujarati, and the Whisper Gujarati fine-tunes on DiarBench with English-in-English scoring; ~50.4% of English words stay in Latin (those larger Gujarati models write 0%).
  • Read Gujarati still trails: IndicConformer leads FLEURS (18.8 vs 27.9) and Kathbath (14.5 vs 20.7); SraVaani is ahead on both read sets too.
  • Limits: no punctuation; numbers as words; loud noise can loop phrases; about half of English words still land in Gujarati script on some clips.

Original Source

Conclusion

If you need privacy-first Gujarati+English ASR that fits in an app download and keeps Hinglish-style Latin English inside the transcript, Shruti Gujarati is the day-0 open file to pull — and it closes another Indic language gap in the same 29.7 MB needle format as the rest of the Shruti kit.

—Anabel ♡