Recall

Why it exists

A recording preserves a conversation but leaves the hard part to the listener: finding the useful moment and knowing who said what. Recall is my local call-recording and processing project, scoped to recording, transcription, speaker identification and analysis. Keeping the processing local gives me control over personal files and a practical place to test open speech models on my own hardware.

Approach

The working import path turns existing audio or video into speaker-labeled transcripts organized by session. It combines Python, Whisper transcription and pyannote speaker diarization, the process of separating speech by speaker. The work covers both the speech pipeline and the application around it: importing files, managing settings and sessions, and making the workflow behave consistently.

Recall is also the demanding application my Local AI Harness is tested against. A settings defect with a failing baseline test became a bounded task for a local model.

What I learned

Dedicated speech tools beat asking a general language model to infer who said what. Full-call testing matters alongside small regression checks, because application behavior and speaker assumptions can fail in ways a compact fixture never exposes. Recall is working software with bugs and polish work remaining; next is an issue inventory and a release threshold for dependable regular use.