Write the eval before the fix: building a WhatsApp transcriber with LLM-TDD (waldo.vanderlore.de)

A new tool called WhatYouSay has been developed to transcribe WhatsApp voice messages, allowing users to read them instead of listening, thereby addressing the common issue of inconvenient voice notes. This server-based solution transcribes, cleans up, and posts the transcript directly in the chat, enhancing communication efficiency and offering users greater control over their privacy compared to existing services. The development process utilized an innovative approach called LLM-TDD (Large Language Model Test-Driven Development), where the checks are established first to guide the LLM in generating the code, ensuring the final product meets defined quality standards. What makes this project significant for the AI/ML community is its emphasis on privacy and the detailed error-handling mechanisms in its pipeline, which operate across multiple languages—English, German, and Dutch. The system fine-tunes the transcription process, implementing language-specific rules to minimize errors while maintaining colloquial language and personal nuances. With the ability to provide immediate, contextually accurate transcripts while preserving users’ privacy by not storing any transcribed text, WhatYouSay sets a notable precedent for future applications in personal communication technologies. The open-source nature of the project, hosted on GitHub, invites further exploration and development, making it an intriguing case study for AI practitioners.