Translations sound too literal/textbook-like
I integrated Sarvam’s Translation Model into one of my projects a couple of months ago. The API and inference quality are generally good.
However, my client keeps raising the same concern: the translations sound too literal.
In most Indian languages, people don’t speak in purely “local” words — they usually mix the local language with some English or Hindi. But the model produces translations that are too formal or pure, which makes some words hard for ordinary people to understand. So for projects where the downstream subtasks relating to speech generation, it kinda becomes useless.
I understand this isn’t something I can fix by just changing a model hyperparameter. What I’m looking for is any quick fix to make the translations sound more spoken and natural, rather than like literature or textbook language.