My family switches language mid-sentence without noticing. English for the practical part, Urdu for the part that means something, back to English for the joke. Nobody announces the change. It is not code-switching in the academic sense, it is just how the sentence came out.
Bollywood has written like this for forty years. And yet nearly every tool built for songwriting assumes one language per song, chosen up front from a dropdown.
What goes wrong, specifically
The accent problem
Devanagari or Urdu script in the lyric field usually produces something with an English speaker's accent laid over it. Certain sounds are consistently wrong. The retroflex consonants come out flat, and the aspirated ones lose their breath entirely, which is the difference between two different words.
The workaround is romanised phonetic spelling, tuned against the model rather than against correctness. You end up spelling things wrong on purpose because wrong sounds right, which is a strange way to spend a Tuesday.
The averaging problem
Ask for Hindi and you tend to get a generic subcontinental sound: some tabla, some strings, a vaguely devotional reverb. It flattens sixty years of wildly different traditions into one texture.
Naming things precisely fixes most of it. Not Indian instruments but harmonium, dholak, sarangi. Not Bollywood but late-nineties playback with the vocal forward and strings behind. Specificity is the whole game.
The register problem
This one is subtle and it is the one I care about most. Urdu poetry carries a formality that has no easy English equivalent. There is a way of addressing someone you love that is respectful and intimate at once, and it does not translate. English either sounds cold or sounds like a greetings card.
Models default to the greetings card. Left alone they produce Urdu that scans as a translation of an English pop lyric, which is exactly the thing everybody's grandmother can hear from across a room.
The rhyme problem
English rhyme is mostly about the end of the word. Urdu poetry uses radif, a whole word or phrase repeated at the end of every line, with the rhyme sitting just before it. It is a different machine entirely, and it is what gives a ghazal its particular circling feeling.
Models trained overwhelmingly on English lyrics do not reach for it. They will rhyme the last syllable and stop, which produces something that scans correctly and sounds, to anyone raised on the form, oddly flat. If you want radif you have to write it yourself and then protect it, because every rewrite pass will try to smooth it away.
What actually works
After a year of this, the process that survives contact with reality:
- 1Write the Urdu or Hindi lines first, by hand, in roman script. They carry the weight, so they should not be a translation of anything.
- 2Let the model draft the English around them. This is the direction it is good in.
- 3Read the whole thing out loud. Bilingual lines that look fine on a page routinely trip the tongue, and a singer will find that before an audience does.
- 4Respell phonetically for the model, then keep a correctly spelled version for the customer. What produces the right sound is not what belongs on a lyric sheet.
- 5Name instruments, never traditions. Harmonium beats Indian instruments every time.
Why any of this matters commercially
There are, at a conservative guess, several hundred million people who speak like this at home. Their weddings run for days and are built entirely around music. And almost every tool aimed at them was designed for a monolingual English market and then translated at the interface, which is not the same thing at all.
That gap is most of why Sureela exists. Not because the technology is different, but because someone has to care whether the aspirated consonant landed, and no dropdown is going to care on your behalf.
Why we ask which language before anything else
Our order form asks for English, Hindi or Hinglish early on, before the style or the feel. That is deliberate. It changes the vocabulary, the instrumentation, the rhyme structure and the vocal we go looking for. Asked at the end, it is a filter. Asked at the start, it is the shape of the song.
Most people choose Hinglish. Which makes sense: it is the only one of the three that sounds like how anybody actually talks.
You do not write a bilingual song by writing it twice. You write it the way the family already speaks.