deepgram-transcribe

Transcribe audio and video recordings (meetings, calls, interviews) with Deepgram and get a Markdown transcript plus the raw JSON response.

For meeting.m4a you get meeting.json (the full Deepgram response, never your API key) and meeting.md (frontmatter, optional summary and topics, speaker-labelled timestamped transcript).

Install

uvx deepgram-transcribe --help          # run without installing
uv tool install deepgram-transcribe     # installs the `dgt` command

Requires Python 3.11+. Both dgt and deepgram-transcribe are installed.

API key

Get a key from the Deepgram console, then provide it in one of these ways (highest precedence first):

1. --api-key on the command line
2. the DEEPGRAM_API_KEY environment variable
3. api_key in the config file

export DEEPGRAM_API_KEY=your-key

Usage

Transcribe recordings (globs are expanded by the tool; quote them so ** works):

dgt run meeting.m4a
dgt run --output-dir ~/Transcripts "recordings/**/*.mp3"
dgt meeting.m4a                      # shortcut for `run`
dgt run --summarize --topics --language en call.wav
dgt run --keyterm Reaktor --param tier=enhanced meeting.m4a

Re-render Markdown from saved JSON, without calling the API:

dgt render meeting.json
dgt render --force --date 2026-10-07 *.json

Common options: --output-dir DIR, --force (overwrite; otherwise files with existing outputs are skipped), --date mtime|now|YYYY-MM-DD, --config PATH, -q/--quiet, -v/--verbose. Run dgt run --help for the full list of Deepgram flags (--model, --language, --diarize, --smart-format, --paragraphs, --keyterm, --redact, --param KEY=VALUE, ...).

Exit codes: 0 success or skipped, 1 at least one file failed, 2 usage or config error.

Config file

Location: $XDG_CONFIG_HOME/deepgram-transcribe/config.toml, falling back to ~/.config/deepgram-transcribe/config.toml. A missing file is fine.

api_key = "..."                # optional
output_dir = "~/Transcripts"   # optional
date = "mtime"                 # optional
timeout = 600                  # optional, seconds

[deepgram]                     # defaults for any Deepgram query param
model = "nova-3"
language = "multi"
keyterm = ["Reaktor", "Hermes"]

Precedence per option: CLI flag > config [deepgram] > built-in default.

English-only features

--summarize, --topics, --intents and --sentiment only work for English. They are sent only when the effective language starts with en and language detection is off. With the default language = "multi" they are dropped with a warning, so pass --language en to use them.

More

Full behaviour is described in the spec. Licensed under MIT.