ANVE: the private AI office, on the record. See what that means

ANVE

Offline AI in your language · updated 2026-10-16

Can offline AI handle Sindhi?
सिन्धी and سنڌي in speech, text and translation: two scripts, thin evidence.

Short answer. Weakly for speech, workably for translation. Whisper’s language list includes Sindhi, but its paper has no Sindhi transcription data, only 46 hours of Sindhi speech paired with English text, and large-v2 scored 156.5% word error rate on FLEURS. AI4Bharat’s IndicConformer lists Sindhi and Meta’s MMS lists it for speech recognition; in the IndicVoices paper IndicASR scored 29.6% on natural speech after training on only 10 hours. Sindhi is written in Arabic script and in Devanagari, and IndicTrans2 handles both, though NLLB 1.2B scored higher from English into Arabic-script Sindhi (48.9 against 44.9 for IndicTrans2). Qwen3 lists Sindhi. ANVE has not tested Sindhi yet. Coverage and figures are as of October 2026.

01

What open models cover in Sindhi

TaskModelsWhat the source says
Speech to text: WhisperWhisper large-v2, large-v3Sindhi (sd) is in Whisper’s tokenizer, but the paper lists no Sindhi speech with Sindhi transcripts: its only Sindhi data is 46 hours of Sindhi speech paired with English text. Large-v2 scored 156.5% word error rate on FLEURS, the highest figure the paper reports for an Indian language.
Speech to text: Indian-language and Meta modelsIndicConformer, Meta MMSIndicConformer lists Sindhi (MIT licence); IndicWhisper leaves it out of its 12 languages. IndicVoices, natural speech: IndicASR 29.6%, trained on just 10 hours of Sindhi; Whisper 92.6%; no MMS figure. Meta’s table marks Sindhi for speech recognition, not synthesis; the checkpoint is CC-BY-NC.
Chat, summaries, draftingQwen3, Sarvam, IndicBERTQwen3’s launch post lists Sindhi. The Llama 3.3 and 4 cards, the gpt-oss card and MILU cover no Sindhi, so no chat figure is available. Sarvam 30B and 105B list Sindhi (sd) in their card metadata, and AI4Bharat’s IndicBERT v2 includes it.
TranslationIndicTrans2, NLLB-200, IndicTrans3-betaFLORES-200, Arabic script, chrF++: English to Sindhi 44.9 for IndicTrans2 and 48.9 for NLLB 1.2B distilled; Sindhi to English 65.1 and 64.0. In Devanagari, IndicTrans2 scored 36.6 and 57.3 on web sentences (IN22-Gen) and 30.3 and 49.6 on conversational ones. NLLB-200 lists Arabic script only (snd_Arab). IndicTrans3-beta gives Sindhi preliminary support and warns that quality may vary.
02

Arabic script, Devanagari and the Kachchhi question

Sindhi is written in Arabic script (سنڌي) and in Devanagari (सिन्धी). IndicTrans2 lists both, FLORES-200 and NLLB-200 list Arabic script only, and Meta’s MMS table has one Sindhi entry without naming a script. The 2011 Census counted 27,72,264 people with Sindhi as their mother tongue, 0.23% of India and the seventeenth largest group. Gujarat had 11,84,024, Maharashtra 7,23,748, Rajasthan 3,86,569, Madhya Pradesh 2,45,161 and Chhattisgarh 93,424. The Home Ministry says Sindhi joined the Eighth Schedule in 1967.

The Census groups Kachchhi (10,30,602 speakers) and Bhatia (22,409) under Sindhi. Sindhi proper is 16,79,246, about three in five of the group. If your clients speak Kachchhi, none of the sources tests it on its own, and it is not clear which Sindhi model would follow it.

For English meetings, the Svarah paper found Whisper large at 7.3% word error rate on speakers whose mother tongue is Sindhi, against 5.3% for native Hindi speakers.

03

What to test before you rely on it

  • Decide the script first, Arabic or Devanagari, and check that the tool can produce it. The cards for the Sindhi speech models do not say which script their output uses.
  • Record ten minutes of natural speech. IndicASR’s 29.6% came from a model trained on only 10 hours, so expect variation between speakers.
  • Do not start with Whisper: 156.5% on read sentences is the highest FLEURS error rate its paper reports for any language.
  • Compare IndicTrans2 and NLLB-200 on the same English-to-Sindhi notice in Arabic script. In the paper NLLB scored higher in that direction.
  • Ask a Sindhi reader to mark errors, especially in names and in Devanagari spellings of Arabic-script words.
04

Where ANVE stands

ANVE’s meeting-minutes maker handles Hindi, English and Hinglish today. Sindhi has not been tested on ANVE yet, so there is no ANVE result to report and nothing here promises support.

If no local model serves Sindhi well enough, a cloud model is possible only when an administrator adds a provider. Its answers are labelled, and every model call is logged with the user, app, model and whether it ran locally or in the cloud. A policy that blocks confidential data from cloud models is not on ANVE yet. The security page has the details.

A demo call is free. If Sindhi decides your purchase, say which script you need and ask for your own recordings to be tried before you commit.

05

Questions people ask

Does Whisper support Sindhi?
It has a Sindhi language tag, but the paper lists no Sindhi speech with Sindhi transcripts, only 46 hours paired with English text, and large-v2 scored 156.5% word error rate on FLEURS.
Which script does offline Sindhi AI use?
Both appear. AI4Bharat’s IndicTrans2 handles Sindhi in Arabic script and in Devanagari; FLORES-200 and NLLB-200 cover Arabic script only. The speech model cards do not state a script, so test it.
Is there an offline Sindhi speech model?
AI4Bharat’s IndicConformer lists Sindhi (MIT licence), and Meta’s MMS lists Sindhi speech recognition under a non-commercial licence. In the IndicVoices paper IndicASR scored 29.6% on natural Sindhi speech.
Can offline AI translate Sindhi?
Yes. On FLORES-200 IndicTrans2 scored 44.9 chrF++ from English to Arabic-script Sindhi and 65.1 from Sindhi to English, against 48.9 and 64.0 for NLLB 1.2B distilled. Have a Sindhi reader check anything official.
06

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

07

Sources