Operational Intelligence
CallMind
Script-compliance scoring for call audio, without manual listening.
About the Project
A GPU-accelerated pipeline that transcribes Hindi calls, separates speakers and scores each call against its campaign script, using a mix of multilingual embeddings and lexical matching. Raw audio becomes a structured compliance report for every call.
Technical Overview
Data Sources
- Hindi call recordings
- Campaign scripts
- ASR transcripts
- Ground-truth checklists
ML / NLP Stack
- Whisper ASR
- Speaker diarisation (pyannote)
- Multilingual sentence embeddings
- Hybrid LLM + lexical matching
- LLM-assisted ground-truth extraction
Key Challenges
- Noisy call audio
- Hindi accent variability
- Paraphrase vs literal script matching
- Objective compliance scoring at scale
Interested in this solution?
Let's discuss how we can adapt this approach for your business.
Get in Touch