Automated free speech to SQL transcription

Summary

Speech-to-SQL is an innovative approach designed to bridge the gap between nontechnical users and technical ones in terms of database querying. Τhe project presents an end-to-end application that transforms natural language commands into SQL queries, with the ability to export the results into flat files for further business analysis. To evaluate the effectiveness of this approach, multiple open-source commercial Large Language Models (LLMs), as well as models hosted locally (Hugging Face), were evaluated and compared. The goal was to identify the most suitable model for the given use case, and at the same time create a user friendly web application to record or upload these commands and receive the output. While no single model consistently outperformed the others across all scenarios, the findings revealed that performance was strongly influenced by the complexity of the query and the different way of writing the same query. OpenAI, Gemini, and Claude emerged as the top-performing models in terms of query prediction accuracy, while their latency measurements were found to be relatively similar. We concluded that additional steps are required before delivering a concise and productionready application. These include implementing connections to multiple database types and performing model optimization to reduce operational costs.