Transcribing Hinglish: what happens to English words in Hindi audio

What we ran

The recording is the Hindi sample on our samples page: 27 seconds, four newspaper-headline-style sentences read aloud by one volunteer from Mozilla Common Voice 22.0 (CC0). Three of them carry English names or terms inside Hindi grammar (a hat-trick against Spain, Ronaldo, a record, the India Today Conclave, OBC), 8 English-origin words in 39 reference words. That is the commonest form of Hinglish in writing, but it is read speech, not a conversation, and it is one speaker and one clip. We did not find an openly licensed recording of natural, spontaneous Hinglish that we could use, so we did not invent one.

Each run is the engine's own call (the small model, the one the free preview and regular accounts use, int8 on CPU, voice-activity filter on, beam 5), run directly on the file; the runs with Hindi and with automatic detection return exactly what the sample page shows. Dated 2026-10-27. Error rates come from the same code as the WER calculator, against the sentences the speaker was asked to read, after lowercasing and removing punctuation.

What came back

The English-origin words are the eight in the reference that come from English or are foreign names: Spain, hat-trick, Ronaldo, record, India, Today, Conclave and OBC, counted when they appear in the output spelled exactly as in the reference. Other Hindi words were also wrong (the error rate counts all of them), so the rate does not measure English words alone.

With the language set to Hindi

This is the setting the free preview uses on our Hindi pages. The English words arrived as plausible Devanagari spellings that are not the standard ones:

The sentences in full:

Whether हेट्रिक counts as an error depends on your style sheet: Hindi has no single spelling for every English word, and the reference is one volunteer's text, not the only correct one. What does not change is that these words will not be the spelling you would search for, so plan to find and fix them.

With the language set to English

Choosing English because the speaker uses English words does not give you a Hindi transcript with English left alone. The engine wrote English sentences for the whole clip, and they are a translation of the Hindi, not a transcript of it (the first three lines):

• Don't forget to come here at 7 am tomorrow.
• Ronaldo made a new record with a hat-trick against Spain.
• In India Today Conclave, Deepika told Kamyaabika Mantra,

The Hindi words that were kept came out as guessed spellings ("Kamyaabika Mantra"), and the meaning of the last line was lost. Our app has a separate option for translating to English; this is what you get when the language setting does it by accident. Use English only when the speech is in English.

With vocabulary hints

The app's Vocabulary hints field (under Advanced options, in a signed-in account) nudges the engine toward spellings you give it. We tried it three ways. With the five English-origin terms typed in Devanagari the word error rate fell from 51.3% to 35.9% and 6 of 8 English-origin words matched the reference ("India Today Conclave", Ronaldo and hat-trick in the spelling we typed); रिकॉर्ड and OBC were still wrong, and other words changed too, one for the worse (बजे, "o'clock", became बचे).

The same five terms in Latin letters were worse, not better: the names did appear in Latin letters, but the first two sentences disappeared from the output and the remaining third sentence was stamped as starting at the beginning of the clip, though it is spoken near the 11-second mark. And three unrelated hints changed nothing in the score (still 51.3%), though they changed the spelling of some words, so an improvement comes from the right words, not from using the field.

Two cautions. The hints here were typed from the reference sentences, which is the best case: in real use you will know some of the names in advance but not all. And one clip proves little. Try hints on a short piece of your own recording first and compare, as the Latin-letters run shows they can also hurt.

What to do with a Hinglish recording

• ·Set the language to Hindi when Hindi carries the grammar. Do not set English because some words are English.
• ·Expect Devanagari. The Hindi setting never produced Latin letters for Hindi words in any of our runs. If you need Roman Hinglish ("kal yahan saat baje aana"), transliterate the Devanagari afterwards with a tool built for it; we do not offer one and have not tested any.
• ·Decide how English words are written (Devanagari or Latin) before you proofread, and apply it throughout. Search the transcript for the words you expect (account holders can search all transcripts) and fix them in one pass.
• ·Try Vocabulary hints in Devanagari for the handful of names and terms you know will recur, on a short piece first. Keep the list short.
• ·If the recording is mostly English with some Hindi, we did not measure that direction for Hinglish. For Arabic and English we did, and the finding there was that the language setting decides which language everything is written in, so test a short piece first. See Arabic and English in one recording.
• ·Check names, numbers and quotes against the audio. This is a draft, and the free preview runs the small model; a larger model only runs on a self-hosted install.

What this test cannot tell you

• ·It is one 27-second clip of read speech by one speaker. Conversation, speed, accents, background noise and longer stretches of English will all behave differently.
• ·The error rate is against one volunteer's text, after lowercasing and removing punctuation. It counts every word, not only the English ones.
• ·Results from a neural model can change when the model or its settings change. Dated 2026-10-27; the script that produced the raw data is in our repository (scripts/codeswitch/).
• ·We did not try other engines. This is about our engine, with the small model.

Data

The raw output of all 16 runs, with the metrics, is one JSON file: code-switching.json (CC BY 4.0). The recording's source and licence are on the Hindi sample page.