Synthetic transcript dataset

14 fictional loan-sales calls with deliberate noise: Hinglish, Romanized Hindi, spelling and ASR errors, repeated words, broken punctuation, number-format variation, abbreviations, code-switching and diarization errors. Each has gold annotations for speakers, dialogue acts, entities, intent and evidence. The same calls are reused across the labs and the capstone.

PRECOMPUTED EXAMPLE

Personal loan — clear interest, job change

HinglishnoisyAgent Ainbound4m 32ssegment: salariedHinglishSpelling errorsNumber format variationCode-switchingRomanized Hindi

1Agenthello sir kaise help karuGreetingOpening

2Customersir mujhe ek personal lon^LOAN chahiyeRequestDiscovery

3Agentji sir kitne ka chahiyeQuestionDiscovery

4Customeraround five lac^AMOUNTAnswerDiscovery

5Agentokay sir aap salaried^EMPLOYMENT hoQuestionQualification

6Customerhaan but abhi job change kr raha huAnswerQualification

7Agentkoi baat nahi sir, tenure kitna chahiye aapkoQuestionQualification

8Customer3 saal^TENURE... teen saal theek rahegaAnswerQualification

9Agentok sir main aapko documents ki list bhejta huInformClosing

10Customerhaan bhej do, main kal^DATE tak apply kar dungaConfirmationClosing

Gold output

{
  "intent": "interested",
  "product": "personal_loan",
  "amount": 500000,
  "tenure_months": 36,
  "rate": null,
  "employment": "salaried",
  "evidence_turns": [
    2,
    10
  ],
  "objections": [],
  "needs_review": true
}

Annotator notes

Explicit request (turn 1) and commitment to apply (turn 9). Flag for review: the customer is mid job-change, which affects eligibility, not intent.

Amount normalized: ₹5,00,000