Leveraging Generative AI for Assessing the Completeness of VoIP Call Recordings

Ensuring the completeness of VoIP recordings is a paramount problem for companies that use the recordings as proof. Call recordings have an essential role in verifying crucial interactions from financial transactions like wire transfer, insurance-related conversations to healthcare and logistics coordination. Recording companies usually have extremely strict contracts with customers that have penalties in case of incomplete recordings. In contact centers, the numbers of calls recorded each day run to thousands and it’s practically impossible to check the completeness of every call. In an attempt to solve this problem, we developed an on-premises system that check the recording accuracy and safeguarding the information. Our system integrates Whisper X for audio transcription with fine-tuned Llama 3.2 1B language model and optimized with Unsloth. The model has been designed to have a minimal hardware requirement, thus providing a cost-effective, scalable solution to verify call completeness. This paper provides the framework and details of the implementation and assessment with the working system and its capability of fulfilling contractual requirements while reducing the risks of incomplete recordings.