Voice Tables: Turning Speech into Structured Spreadsheet Rows

Voice Tables is a tool that lets users speak naturally and receive structured spreadsheet rows without typing. The system uses a two‑stage pipeline: Whisper transcribes the audio into plain text, then a custom prompt chain extracts fields into JSON. Whisper is configured with model whisper‑1, auto‑detect language, temperature 0, and text response format to keep output deterministic. The extraction prompt is column‑aware, listing each column name, type, and optional description so the LLM knows exactly what to pull. Validation normalises numbers, dates, and coerces types, flagging missing required fields. A retry fallback to the larger Whisper model handles edge cases like background noise or accented speech. The approach demonstrates how column descriptions dramatically improve extraction accuracy and how a simple two‑stage pipeline can power voice‑to‑data workflows.