1779 train / 43 validation calls from a production collections voice agent (Hindi/Hinglish), processed into the same schema as MeghanaKap/speech-to-speech-data. sides can be modelled without speaker separation. Real customers are heard in these recordings, so access is gated and reviewed. Kept: the customer's first name (the model must learn to say it), the amounts discussed, and the last 4 digits of the loan the agent reads aloud. Removed: full loan/account numbers, phone numbers (muted in the audio, tagged in the text, and stripped from the word timings), conversation and client ids, the source recording URL, and any digit run of 5 or more in the reference script. slotdict / systemprompt…