Extracting email IDs accurately from Erroneous Voice call transcripts: Tackling an Unsolved Problem.

Voice call transcripts have spelling errors. Eg: 'a.b@....' being transcripted as ‘a dot b at the rate...'. Such errors can be fatal in case of email IDs, which need accuracy. Numerous businesses worldwide, which use AI based voice calling, etc., face this issue. Working on a novel approach to tackling this problem. There are no voice based email datasets online. So, created my own datasets using various AI-tools. Fine-tuning LLama3 and Ph3 to extract email IDs from the conversation transcripts, modify email IDs based on user feedback.Using Vocode, WhisperAI along with HuggigSound models for the Speech to Text part. This work can prove to be a leader in the space of AI-Based Voice Calls, solving a very pressing problem by decreasing the need of human intervention. The solution can be expanded to other areas like phone calls, addresses, names, etc.