Transcribing Code-Switched Indian Content: Hindi-English and Beyond

Code-switching means alternating between languages within a conversation, sentence, or even word. It is not translation, borrowing, or error - it is a systematic linguistic behaviour with its own grammar rules. Hindi-English code-switching follows patterns: Hindi provides the grammatical frame, English supplies technical nouns and verbs.

A sentence like 'main kal meeting attend karunga' follows Hindi grammar (main... karunga) while inserting English words (meeting, attend). This is distinct from speaking English with Hindi words inserted, which follows English grammar.

Hindi mode transcribes into Devanagari script. When English words appear, a good system writes them in Latin script within the Devanagari text: 'मैंने presentation दी'. This preserves readability for Hindi readers who expect Devanagari. But some systems attempt to transliterate English into Devanagari, producing 'प्रेज़ेन्टेशन' - technically readable but visually jarring.

Hinglish mode outputs everything in Roman script: 'maine presentation di'. This works well for casual content, social media captions, and creators whose audience reads Roman-script Hindi. It avoids script-boundary issues entirely but loses Devanagari for readers who prefer it.

Tamil-English switching follows different rules than Hindi-English. Tamil agglutinates English roots with Tamil suffixes: 'meeting-ku ponom' (we went to the meeting). The English word receives Tamil grammatical markers. This makes word segmentation harder for AI because 'meetingku' may be treated as one word.

Telugu-English switching frequently places English adjectives before Telugu nouns or uses English verbs with Telugu auxiliary constructions. Bengali-English switching in Kolkata urban speech can reach 40% English vocabulary while maintaining Bengali grammar throughout.

Step 1: Listen to 30 seconds of the content and estimate the English percentage. Step 2: If above 30% English, use Hinglish or English mode. If below 20%, use the primary language mode. Between 20-30%, try both and compare. Step 3: After transcription, scan for garbled words at language boundaries - these indicate the model struggled with a switch.

For multi-speaker content where different speakers have different mixing levels, transcribe in the mode that fits the dominant speaker. Fix the minority speaker's segments manually.

The hardest case is intra-word mixing: English roots with Hindi/Tamil suffixes, or Hindi words with English plurals. 'Files-on ko sort karo' contains 'files' (English) with 'on' (Hindi postposition). AI may output 'file son' or 'files on' depending on how it segments.

Accept that some manual correction is unavoidable for heavily mixed content. The goal is choosing the mode that minimizes total corrections, not eliminating them.