code-switching is the use of two or more languages in the same conversation or utterance. in text to speech (tts), a code-switched line asks the voice model to change pronunciation rules while keeping the same speaker, pace, and emotion.

how does code-switching work?

a language can change between sentences, at a clause boundary, or around one word. the tts model has to work out which language fits each part of the text, then apply the sounds and stress patterns that belong to it.

the difficult part is the boundary. the voice should not pause as if two recordings were joined together. its pitch and pace should continue through the switch and the speaker should remain unchanged on both sides.

script is a separate choice. hindi may appear in devanagari or as romanised hindi in latin letters. transliteration changes the script. translation changes the language and meaning.

code-switching example

rumik muga accepts hindi and english in one line. a voice agent could send:

Aapka order number "4820" hai. Delivery date Friday hai

Aapka order number "4820" hai. Delivery date Friday hai

the line means, "your order number is 4820. the delivery date is friday."

Aapka and hai are romanised hindi. Order number, delivery date, and Friday are english. muga moves between the two languages inside each sentence while keeping the same voice.

the quotation marks around 4820 tell silk to read the order number one digit at a time. they control number handling rather than code-switching, but both behaviours have to work in the same line.

listen at the language boundary. the english words should sound clear without creating a new voice or breaking the rhythm of the hindi clause.

code-switching vs code-mixing

linguists do not use code-switching and code-mixing in one fixed way. some use both terms for the same behaviour. others call a change between clauses or sentences code-switching and reserve code-mixing for words from two languages inside one clause.

the muga example is code-mixed under that narrower definition because hindi and english share the same clauses. hinglish can include this kind of mixing as well as full switches between hindi and english sentences.

why does code-switching matter for voice models?

callers notice a bad language switch even when they understand every word. a pause before the english phrase makes the reply sound stitched together. a change in voice makes one speaker sound like two, while a drop in emotion can alter the point of the sentence.

rumik silk handles hindi-english text through muga and mulberry. muga accepts devanagari and romanised hindi alongside english. mulberry accepts hindi in devanagari and english in latin text within the same request.

we built indicemo to measure whether a tts model carries the requested emotion or speaking style through a language change. in the published comparison, rumik-oss 1 scored 2.92 out of 5 overall. that result applies to rumik-oss 1, not muga or mulberry.

three audio-capable automated judges scored the 98 tasks retained in the final comparison. indicemo measures expression quality across language changes, not transcript correctness, native accent accuracy, human mean opinion score, or overall tts quality.

test full replies from the real call flow. listen for the same speaker, clear pronunciation, natural timing, and steady emotion on both sides of every language switch.