Can Whisper transcribe Bengali?

It accepts Bengali, but the paper reports 104.1% word error rate on FLEURS for large-v2, with only 1.3 hours of Bengali training audio. Test an Indian-language speech model, or your own audio, before relying on it.

Which offline model should I try first for Bengali speech?

AI4Bharat’s IndicConformer or IndicWhisper. In the IndicVoices paper IndicASR scored 15.9% on natural Bengali speech, and IndicWhisper averages 20.1% on its Bengali test sets. Both are MIT-licensed; IndicConformer’s Hugging Face page asks for contact details first.

Do Llama models support Bengali?

Not officially. Meta’s Llama 3.3 and Llama 4 cards list Hindi but not Bengali. Qwen3 lists Bengali, and gpt-oss reports a Bengali MMMLU score.

Can offline AI translate Bengali documents?

For short passages the scores are respectable: 49.7 chrF++ from English to Bengali and 58.4 from Bengali to English on conversational sentences. Have a Bengali reader check anything official.

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 Bengali?
বাংলা across transcription, chat and translation: where open models hold up and where they fail.

Short answer. Text and translation, yes; transcription only with an Indian-language model. Whisper as released is the weak point: its paper lists 1.3 hours of Bengali training audio and reports 104.1% word error rate on FLEURS for large-v2, which means more errors than words. AI4Bharat’s models do far better: on natural speech in the IndicVoices paper, its IndicASR scored 15.9% against 93.2% for Whisper. For chat, Qwen3 lists Bengali and gpt-oss’s card reports 80.9% on Bengali MMMLU (120b, high reasoning). For translation, IndicTrans2 scored 49.7 chrF++ from English to Bengali on conversational sentences. ANVE has not tested Bengali yet. Coverage and figures are as of October 2026.

What open models cover in Bengali

Task	Models	What the source says
Speech to text: Whisper	Whisper large-v2, large-v3	The Whisper paper counts 1.3 hours of Bengali speech with Bengali transcripts and 1,988 hours paired with English text. Large-v2 scored 104.1% word error rate on FLEURS and 103.4% on Common Voice 9. The large-v3 card gives no Bengali figure.
Speech to text: Indian-language models	IndicConformer, IndicWhisper, Meta MMS	IndicConformer lists Bengali (MIT licence). IndicWhisper averages 20.1% on Bengali test sets, 20.9% on FLEURS. IndicVoices, natural speech: IndicASR 15.9%, MMS 44.4%, Whisper 93.2%. The MMS checkpoint is CC-BY-NC, which bars commercial use.
Chat, summaries, drafting	Qwen3, gpt-oss, Gemma, Sarvam	Qwen3’s launch post lists Bengali. gpt-oss: Bengali MMMLU 80.9% (120b) and 77.1% (20b), at high reasoning effort. The Llama 3.3 and 4 cards do not list Bengali. MILU (2024 test, Indian exam questions; open base models, 5-shot): Llama 3.1 70B 67.37%, Gemma 2 27B 63.47%. Sarvam-1 and Sarvam-M name Bengali among their 10 Indian languages.
Translation	IndicTrans2, NLLB-200	IndicTrans2 paper (2023), conversational set (IN22-Conv), chrF++: English to Bengali 49.7, Bengali to English 58.4; NLLB 1.2B distilled 47.1 and 55.4. On web and Wikipedia sentences (IN22-Gen): 51.8 and 63.2. NLLB-200 is CC-BY-NC; IndicTrans2’s checkpoints are MIT.

Bengali script, West Bengal and Assam

Bengali (বাংলা) has its own script, and FLORES-200 gives Bengali and Assamese the same script code. The 2011 Census counted 9,72,37,669 people with Bengali as their mother tongue, 8.03% of India and the second largest group. West Bengal had 7,86,98,852, Assam 90,24,324, Jharkhand 32,13,423 and Tripura 24,14,774. The Census counts people in India only; Bengali is spoken beyond India as well, which these figures do not cover.

If your staff mix English into Bengali sentences, or type Bengali in Roman letters, test both. AI4Bharat’s IndicLID identifies Bengali in native and romanised text, but no source here reports how well speech models transcribe mixed Bengali and English.

English spoken by people whose mother tongue is Bengali is its own test. In the Svarah paper, Whisper large scored 7.6% on their English, against 5.3% for native Hindi speakers.

What to test before you rely on it

Do not assume the speech model works. A word error rate above 100% means the transcript has more wrong words than the reference has words, so test any Whisper-based tool on your own Bengali audio.
Record spontaneous speech from speakers in West Bengal, Assam and Tripura; the Census shows large Bengali populations in each.
Check Bengali digits and mixed numbers (০ to ৯ against 0 to 9), rupee amounts and dates.
Check that conjunct letters and names are spelt as your documents spell them.
Translate a notice in both directions and ask a Bengali reader to mark the errors. chrF++ scores say how close the output is to a reference, not whether it is acceptable.

Where ANVE stands

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

ANVE runs open models, so the table above is the honest starting point, including the gap between Whisper as released and the Indian-language models.

Models change, and so will the evidence on this page. ANVEAI installs updates on site today, signed update packages applied offline by USB are not on ANVE yet, and ANVE does not update itself over the internet, so a newer Bengali model would be a deliberate update that someone installs. The security page has the details.

A demo call is free. If Bengali decides your purchase, bring recordings and ask for them to be tried before you commit.

Questions people ask

Can Whisper transcribe Bengali?: It accepts Bengali, but the paper reports 104.1% word error rate on FLEURS for large-v2, with only 1.3 hours of Bengali training audio. Test an Indian-language speech model, or your own audio, before relying on it.
Which offline model should I try first for Bengali speech?: AI4Bharat’s IndicConformer or IndicWhisper. In the IndicVoices paper IndicASR scored 15.9% on natural Bengali speech, and IndicWhisper averages 20.1% on its Bengali test sets. Both are MIT-licensed; IndicConformer’s Hugging Face page asks for contact details first.
Do Llama models support Bengali?: Not officially. Meta’s Llama 3.3 and Llama 4 cards list Hindi but not Bengali. Qwen3 lists Bengali, and gpt-oss reports a Bengali MMMLU score.
Can offline AI translate Bengali documents?: For short passages the scores are respectable: 49.7 chrF++ from English to Bengali and 58.4 from Bengali to English on conversational sentences. Have a Bengali reader check anything official.

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.

Can offline AI handle Bengali?বাংলা across transcription, chat and translation: where open models hold up and where they fail.