dialogue-transcriber

Transcribe conversations and find out who said what.

Point it at an interview, panel discussion, meeting recording, or YouTube URL and get back a transcript where every line is attributed to a speaker — plus a web UI to inspect the speaker clusters, listen to any segment, and fix labels by hand.

How it works

audio  ──►  transcribe  ──►  segment  ──►  extract_clips  ──►  embed  ──►  cluster
              (Whisper)        (sentence-       (ffmpeg)       (TitaNet)    (UMAP +
                               level)                                       KMeans +
                                                                            silhouette)

Whisper produces word-level timestamps; words are grouped into sentence segments; each segment's audio is embedded with NVIDIA NeMo TitaNet; the embeddings are clustered on a UMAP projection; and the transcript comes out labeled Speaker 1, Speaker 2, … Every stage is cached on content hash, so re-runs and config tweaks are cheap.

Quickstart

ffmpeg and ffprobe must be on PATH (brew install ffmpeg on macOS).

# No install needed:
uvx --from "dialogue-transcriber[all]" transcriber transcribe interview.mp3

# Or install the tool:
uv tool install "dialogue-transcriber[all]"

transcriber transcribe interview.mp3 --participants 2
transcriber transcribe "https://www.youtube.com/watch?v=..." --backend openai
transcriber serve interview.mp3        # review UI on http://127.0.0.1:8000

The default backend runs faster-whisper locally; --backend openai uses the OpenAI Whisper API instead (requires OPENAI_API_KEY, much faster on machines without a GPU). The key can be exported in the environment or kept in a .env file in your project — the CLI loads .env from the working directory (or nearest parent), and exported variables always take precedence over the file.

Where does data go?

• Pipeline cache: ./.transcriber-cache/ in the directory you run from (override with --work-dir) — chunks, per-segment clips, embeddings, YouTube downloads, and the web UI's job state. Safe to delete; it will be rebuilt.
• Transcripts: written next to the input audio (interview.txt), or wherever --output points; --output - prints to stdout.
• Model weights (local backend): downloaded once into ~/.cache (Hugging Face / NeMo). The Whisper large-v3 download is ~3 GB, so the first local run takes a while.

Nothing leaves your machine with the default local backend; --backend openai sends audio to the OpenAI API.

Picking your extras

[all] is the easy button. For smaller installs:

uv pip install dialogue-transcriber              # core only
uv pip install "dialogue-transcriber[local]"     # + faster-whisper backend
uv pip install "dialogue-transcriber[openai]"    # + OpenAI Whisper API backend
uv pip install "dialogue-transcriber[cluster]"   # + scikit-learn / UMAP
uv pip install "dialogue-transcriber[embed]"     # + NeMo TitaNet speaker embedder
uv pip install "dialogue-transcriber[api]"       # + FastAPI backend (powers the web UI)
uv pip install "dialogue-transcriber[youtube]"   # + yt-dlp downloader
uv pip install "dialogue-transcriber[oip]"       # + MCP server for OIP consumers

CLI

# Full pipeline; writes a speaker-labeled transcript next to the audio
transcriber transcribe path/to/audio.mp3

# Speakers, language, format
transcriber transcribe interview.mp3 --participants 3 --language sv --format vtt

# Machine-readable output on stdout (see "For AI agents" below)
transcriber transcribe interview.mp3 --format json --output -

# Pull audio from YouTube
transcriber download "https://www.youtube.com/watch?v=..."

# Pipeline + web UI
transcriber serve interview.mp3 --participants 3

Formats: txt (merged speaker turns), vtt, srt, json. Pass --context "names, jargon" to prime Whisper with vocabulary it should expect. --output - streams the transcript to stdout and the summary to stderr, so the output pipes cleanly.

Web UI

transcriber serve runs a FastAPI backend and serves the bundled React frontend. You get:

• a UMAP scatter where each dot is one segment, colored by cluster — lasso a cluster to bulk-rename it;
• a continuous waveform with one region per segment — click or scrub to play anything;
• a Gantt-style speaker timeline;
• a virtualized transcript with full-text search;
• inline-renameable speaker chips (renames persist server-side);
• TXT / VTT / SRT export;
• keyboard navigation (↑/↓ segments, Space play/pause, / search).

Multiple jobs can run side by side; add more via the sidebar.

serve picks its backend automatically: openai when an OPENAI_API_KEY is available (environment or .env), otherwise local. Pass --backend to choose explicitly. (A legacy single-job Dash UI is still available as transcriber ui.)

For AI agents

This project is built to be driven by agents as well as humans.

Claude Code skill — the repo doubles as a plugin marketplace. Install the skill and Claude Code will know how to transcribe and diarize audio on demand:

/plugin marketplace add Novia-RDI-Seafaring/transcriber
/plugin install dialogue-transcriber@dialogue-transcriber

Structured output — --format json --output - emits a stable shape on stdout:

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

MCP / OIP — the package is an Open Ingestion Protocol producer, so transcripts can be ingested by any OIP-aware consumer (e.g. Anchor) with no consumer-side changes:

transcriber oip install --data-dir ~/transcripts     # register the producer
transcriber oip ingest audio.mp3 --data-dir ~/transcripts
transcriber oip serve                                # MCP server (also: transcriber-mcp)

Tool namespace: transcribe. Region kind: transcript_segment. source_ref.kind: audio-timestamp.

Library use

from transcriber.config import ClusterConfig, PipelineConfig, TranscribeConfig
from transcriber.pipeline import run_pipeline
from transcriber.render import render_txt

cfg = PipelineConfig(
    transcribe=TranscribeConfig(backend="local", language="en"),
    cluster=ClusterConfig(participants=2),
)
result = run_pipeline("interview.mp3", config=cfg)
print(render_txt(result.segments))

PipelineResult.segments is a list of SpeakerSegment records with the sentence text, time range, the on-disk clip, and the assigned speaker. PipelineResult.cluster.projection is the 2-D UMAP for plotting.

Backends

Concern	Default	Override via
Transcribe	faster-whisper large-v3	--backend openai
Embed	nvidia/speakerverification_en_titanet_large	pass embedder= to run_pipeline
Cluster	UMAP(2) + KMeans + silhouette	pass a ClusterConfig
YouTube	yt-dlp	replace YouTubeDownloader

All backends are Protocols — see transcriber/transcribe/base.py and transcriber/embed/base.py. Tests use in-memory fakes, so the heavy models are not required to run the suite.

Development

See CONTRIBUTING.md for guidelines and CHANGELOG.md for release history.

git clone https://github.com/Novia-RDI-Seafaring/transcriber
cd transcriber
uv venv
uv pip install -e ".[dev,cluster,api,openai,embed,youtube]"
(cd web && pnpm install && pnpm build)   # so `transcriber serve` can serve the UI

pytest                  # core + clustering + api tests
pytest -m "not slow"    # skip heavy/network tests
ruff check src tests

For frontend work: cd web && pnpm dev (http://127.0.0.1:5173, proxies /api to :8000) with transcriber serve … --port 8000 in another shell.

Releases: publishing a GitHub release triggers .github/workflows/release.yml, which builds the frontend, bundles it into the wheel, and publishes to PyPI via trusted publishing.

License

Apache-2.0 — see LICENSE.