Automating the Conducting of Surveys Using Large Language Models

Conducting electronic surveys in remote areas can be challenging due to the lack of the required Internet infrastructure. Traditional phone surveys are typically used in such cases since mobile phones have become pervasive. However, this process can be time-consuming since a human is required to conduct the session and they must then upload responses to a database. We propose using Large Language Models (LLMs) to process an audio recording of the phone session to extract the responses and store them in a database. In our proof of concept humans were used to conduct the survey but further automation is planned whereby the phone session itself is carried out by a robot. We use OpenAI’s Whisper for the speech-to-text process and we then pass the text to a Large Language Model (GPT-4) which is prompted to extract the responses. The responses are then uploaded to a database. Finally we use an LLM to provide answers to questions about the survey responses. For multiple choice questions we obtained an accuracy score of 97%.