Impression

Impression is a web app for semantic search over spoken-word audio. You search by meaning, not keywords. Ask something like "how to set goals" and it returns the exact moments a speaker discussed that idea, even if they never used those words, as timestamped, playable clips that seek straight to the moment.

It started from running with podcasts: I could never recall the one thing someone said last week without scrubbing back through hours of audio by hand. Impression turns that listening history into something I can actually query and revisit.

Ingestion runs offline -- import a podcast or YouTube episode, transcribe it with faster-whisper, chunk it into timestamped windows, and embed each chunk with sentence-transformers (MiniLM, 384-dim). At search time the query is embedded with the same model, matched against the vector store by cosine similarity, and returned as ranked clips. Playback streams directly from S3 via presigned URLs.

Retrieval quality is tracked with a hand-labeled eval harness (recall@5 ≈ 64% over 15 queries) rather than vibes.