Devanagari out, Hinglish in
Transcripts are written in Devanagari script (हिन्दी), with Hindi punctuation and sentence structure. Real Indian speech, though, is rarely pure Hindi — Hinglish code-switching is the norm, and the model handles it the way Hindi media does: English words and phrases inside Hindi sentences are usually transcribed in Latin script inline (आज हम discuss करेंगे next quarter के targets), matching how Hinglish is actually written.
Predominantly English speech with Hindi phrases mixed in works better through the auto-detect uploader on the home page, which will treat it as English.
Accuracy tier and the Urdu question
Hindi sits in the model's solid mid-high tier: clear speech — YouTube content, lectures, interviews — produces very usable transcripts, a step below the top European languages mostly due to code-switching density and speaker variety. Spoken Hindi and Urdu overlap heavily; on this page the output is Devanagari, so Urdu-register speech will be written in Devanagari, not Nastaliq. For Urdu-script output, that's a separate language code (ur) via the home page.
Frequently asked questions
Is the output in Devanagari or Latin letters?
Devanagari for the Hindi, Latin for embedded English — the way written Hinglish actually looks. You won't get an all-romanized transcript. If you need romanized Hindi, transliterate the exported TXT with any Devanagari→Latin tool afterwards.
How well does Hinglish code-switching work?
Well — it's the defining feature of modern Hindi audio and the training data reflects that. Mid-sentence English ("basically", "presentation", full clauses) is transcribed inline in Latin script. Where it wobbles is rapid-fire switching every few words; expect the occasional wrong-script rendering to fix in the editor.
How accurate is Hindi transcription overall?
Solidly usable: on clear audio (podcasts, lectures, broadcast) you get a transcript that needs light editing, a tier below English/Spanish. Regional accents from non-native Hindi speakers and heavy background noise pull accuracy down faster than they do for English.
What about Urdu — same spoken language, different script?
Spoken Hindustani is largely shared, and the model will transcribe Urdu-register speech here — but in Devanagari, since this page pins Hindi. For Urdu output in Nastaliq/Arabic script, upload via the home page and pin Urdu (ur) in the options instead.
Hindi audio to English text?
This page produces the Hindi transcript — the accurate step. Export TXT and machine-translate for English; hi→en translation on clean text is strong. One-shot speech translation loses more, especially on code-switched content. In-app translation is planned.
Do Bollywood songs or bhajans transcribe?
Speech is the design target; singing is best-effort. Clear solo vocals often come out mostly right, but film songs with full orchestration will be patchy. Dialogue scenes, interviews, and vlogs transcribe like normal speech.