dialogue-transcriber

Dialogue Transcriber

transcriber runs a pipeline: Whisper word-level transcription → sentence segments → per-segment audio clips → TitaNet speaker embeddings → UMAP + KMeans clustering → a transcript where every line is attributed to Speaker 1, Speaker 2, …

Setup

Check availability first; prefer an existing install:

transcriber --version || uv tool install "dialogue-transcriber[all]"
ffmpeg -version >/dev/null || echo "ffmpeg missing — install it (e.g. brew install ffmpeg)"

One-off runs also work without installing: uvx --from "dialogue-transcriber[all]" transcriber …

Pick a transcription backend:

• --backend local (default) — faster-whisper on this machine. No API key, but downloads a large model on first use and is slow without a GPU.
• --backend openai — OpenAI Whisper API. Fast and light; requires OPENAI_API_KEY, read from the environment or from a .env file in the working directory (or a parent — exported variables win over the file). Prefer this when the key is available and the audio is not sensitive; ask the user if unsure.

Transcribe (the main task)

# Local file or YouTube URL; writes <audio>.txt next to the input
transcriber transcribe interview.mp3
transcriber transcribe "https://www.youtube.com/watch?v=..." --backend openai

# Useful options
#   --participants 3      number of speakers, if known (improves clustering)
#   --language sv         audio language (default en)
#   --format txt|vtt|srt|json
#   --context "..."       names/jargon hint passed to Whisper
#   --output PATH | -     output file, or '-' for stdout

Pass --participants whenever the user tells you (or the content implies) how many people are speaking — clustering is noticeably better with it.

For your own consumption, use JSON on stdout — parse it instead of scraping the human-readable formats:

transcriber transcribe interview.mp3 --format json --output -

{
  "speakers": ["Speaker 1", "Speaker 2"],
  "n_segments": 42,
  "duration": 512.3,
  "segments": [
    {"speaker": "Speaker 1", "start": 0.0, "end": 4.2, "text": "..."}
  ]
}