deep-transcribe 0.1.10

High-quality transcription, formatting, and analysis of videos and podcasts

deep-transcribe

High-quality transcription, formatting, and analysis of videos and podcasts.

Deep Transcribe accepts YouTube and other media URLs or local audio and video files. It uses Deepgram Nova-3 with the current batch diarizer, then can identify speakers, format paragraphs and timestamps, add sections and summaries, research key passages, capture video frames, and export browser-ready HTML.

LLM processing uses configurable kash model roles. New workspaces use the current Anthropic profile by default, and an equivalent OpenAI profile is included.

Requirements

Install uv and ffmpeg. Deep Transcribe requires Python 3.13, which uv fetches automatically.

Set DEEPGRAM_API_KEY and one LLM provider key in the process environment or a .env file in the current directory or one of its parents:

• ANTHROPIC_API_KEY for the default Anthropic profile
• OPENAI_API_KEY for the OpenAI profile

Do not commit API keys.

Zero-Install CLI

Run the pinned release without installing it globally:

uvx --from deep-transcribe==0.1.9 deep-transcribe --help

For repeated human use, a persistent tool install is also available:

uv tool install deep-transcribe
deep-transcribe --help

Cross-Agent Skill

Install the repository skill through the cross-agent skills installer:

npx skills add jlevy/deep-transcribe

The skill uses a local deep-transcribe executable when available and otherwise uses the pinned zero-install runner. It routes agents to CLI help rather than carrying a second copy of the command manual.

Self-Documenting CLI

Start with the top-level command directory, then open the help page for the relevant task:

deep-transcribe --help
deep-transcribe transcribe --help
deep-transcribe models --help
deep-transcribe mcp --help
deep-transcribe logs --help

The command pages document all presets, individual processing stages, Deepgram language and model selection, source metadata and speaker hints, caching and rerun behavior, JSON output, model profiles, MCP tools, and examples. Existing flag-only invocations remain supported for backward compatibility.

Model Provider

Inspect the exact current Anthropic and OpenAI role mappings before selecting one:

deep-transcribe models
deep-transcribe models --set anthropic
deep-transcribe models --set openai

The selection is saved in the chosen workspace. Pass --workspace to models and transcribe when using a location other than ./transcriptions.

Quick Demo

This short two-person hotel dialogue exercises transcription and diarization:

deep-transcribe transcribe --basic \
    "https://www.youtube.com/watch?v=wyqfYJX23lg"

Use deep-transcribe transcribe --help to choose formatted, annotated, deep, or custom processing.

Source Context and Corrections

Add context directly or use one reusable YAML file for URLs and raw media files:

title: Acme product interview
description: Alice Chen interviews Bob Diaz about SignalFlow.
additional_context: Alice is the host and Bob is the product lead.
key_terms:
  - Alice Chen
  - Bob Diaz
  - SignalFlow
speaker_hints:
  "0": Alice Chen
  "1": Bob Diaz

deep-transcribe transcribe --annotated --metadata interview.yml ./interview.mp3

Update the metadata and rerun the same command to correct speaker names, descriptions, or other semantic context. Those corrections reuse the cached Deepgram transcript. Changing key_terms deliberately requests a new transcript because the terms affect speech recognition. Use --rerun-processing to force every downstream stage while preserving the raw transcript; --rerun intentionally requests fresh speech-to-text processing. Run deep-transcribe transcribe --help for the equivalent concise flags. If Deepgram merges distinct voices under one ID, replace direct ID hints with a complete speaker_roster list and describe role or dialogue transitions in additional_context; Deep Transcribe will correct turn-level boundaries with the selected careful model profile.

Output

Each run reports:

• the workspace containing cached media and intermediate results
• the transcript source
• browser-ready HTML

Use --json when another tool or agent needs stable artifact paths. You can also open the workspace with kash to inspect cached and intermediate items.

MCP Server

Run deep-transcribe mcp --help for the available transcription tools and transports. Run deep-transcribe logs --help for server log handling.

Project Docs

For environment setup, see installation.md.

For development workflows, see development.md.

For the manual, agent-reviewed release test, see e2e-test.runbook.md.

For publishing, see publishing.md.

---

This project was built from simple-modern-uv.