Build a Speaker-Recognition App with Claude Code

An internal speaker-recognition application can be built quickly using AI coding agents like Claude Code to overcome the limitations of anonymous transcription. The goal is to replace generic labels like "Speaker 1" with actual names in meeting transcripts. The process uses a transcription service like Soniox to get an initial transcript with anonymous speaker labels and timestamps. Audio samples for each unique speaker are then extracted, embedded, and compared against a stored database of known, labeled speakers using cosine similarity to find a match. If a speaker is new or the match confidence is low, the user is prompted to provide a label, which is then added to the database for future automatic recognition.