What I built

Connected local Whisper transcription, model scoring and searchable transcript turns in a FastAPI application, with two authored demonstration cases.

Architecture sketch

  appointment audio ──► Whisper (local) ──► reference speaker labels
                                                  │
                                                  ▼
  transcript turns ◄── quote matching ◄── local LLM scoring
         │
         └── embeddings / retrieval ──► FastAPI product flow

Data and model roles

Audio → local Whisper → reference-labelled speakers → local model scoring → transcript quote matching → searchable transcript turns. Speaker labels come from the authored demo timeline; the workflow does not establish autonomous diarization.

Implementation scope and verification

Real local Whisper, Ollama scoring, transcript-turn embeddings/retrieval and FastAPI product flow. Two authored demo cases. Speaker attribution uses reference timeline labels; quote matching is exact/fuzzy transcript matching and can retain unverified evidence.

End-to-end applied AI case with speech, structured output, evidence checks and explicit prototype limits. No established scoring accuracy or production uptime.

About the evidence

Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.

Recorded demonstration

Original public recording of authored fictional appointments and reference speaker labels. The player shows recorded output and runs no new transcription, scoring or model inference.

Open the original recording →