Get per-speaker transcript segments with WhisperX or Deepgram

Steps

1. Install peepshow + WhisperX (local) OR set a cloud API key

   For local diarisation, `pip install whisperx`. Cloud paths use existing API keys.

   npm install -g peepshow
   pip install whisperx
   export HF_TOKEN=hf_...  # required for pyannote model download

2. Run with --diarise

   WhisperX path used if `whisperx` is on PATH; cloud providers honour `--transcribe` selection.

   peepshow ./meeting.mp4 --diarise

3. Pick a cloud provider explicitly

   Pass through to Deepgram's `?diarize=true` or AssemblyAI's `speaker_labels:true`.

   peepshow ./meeting.mp4 --diarise --transcribe deepgram
   peepshow ./meeting.mp4 --diarise --transcribe assemblyai

4. Cap speaker count

   Helps the diariser converge faster.

   peepshow ./call.mp4 --diarise --max-speakers 3

Why it works

Diarisation is the missing piece of meeting / interview transcription. WhisperX (local) ships pyannote.audio under the hood — needs a Hugging Face token to download the pyannote models, then runs entirely offline. Deepgram and AssemblyAI both ship server-side diarisation with a single flag. peepshow's `--diarise` routes through the existing transcription provider chain, so you opt-in once and the right path fires based on what's available.

When it helps

• Meetings / standups / interviews — know who said what without re-listening.
• Podcast post-production — auto-tag segments by host.
• Court / legal interviews where speaker attribution matters.
• Customer-support call review — separate agent vs caller.

Pitfalls

• WhisperX needs a Hugging Face token (`HF_TOKEN`) to download pyannote models on first run.
• Cloud providers don't always agree on speaker IDs — WhisperX uses `SPEAKER_00`, Deepgram uses numeric, AssemblyAI uses `A`/`B`/`C`. peepshow normalises to the provider's native scheme.
• Diarisation accuracy drops on overlapping speech and very short utterances.

Works with these LLMs

Pairs with these sinks