meetdown 0.1.1

Generate Markdown meeting transcripts from speech-to-text JSON.

meetdown

meetdown is a small Python CLI for turning meeting audio into a Markdown transcript.

The CLI sends a local audio or video file to a speech provider, normalizes the response, and writes a .md file with the full text and speaker-by-speaker transcript when the provider returns speaker information. When --provider is omitted, meetdown auto-detects the provider from configured provider-specific credentials.

Install and run

Run it directly with uvx; cloning this repository is not required:

uvx meetdown meeting.m4a -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key"

For local development:

uv run meetdown meeting.m4a -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key"

Set credentials with environment variables:

$env:CLOVA_SPEECH_INVOKE_URL="https://your.invoke.url"
$env:CLOVA_SPEECH_SECRET_KEY="your-secret-key"

You can also pass them directly:

uvx meetdown meeting.m4a -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key"

Run meetdown without arguments, or use --help, to see the current defaults and provider auto-detection status without transcribing anything. In interactive terminals, the defaults are color-coded by area; set NO_COLOR=1 to keep plain text output.

uvx meetdown
uvx meetdown --help

Recognition language defaults to auto, so meetdown does not force a single language when audio contains a mix such as English and Korean. Pass --language ko-KR, --language en-US, or another provider-supported language when you want to bias recognition toward one language.

Python API

Use the top-level package API when you want to call meetdown from Python instead of shelling out to the CLI:

from meetdown import render_markdown, transcribe_file, write_markdown

response = transcribe_file("meeting.m4a", provider="openai")
markdown = render_markdown(
    response,
    title="Weekly sync",
    source_path="meeting.m4a",
)
write_markdown("meeting.md", markdown)

transcribe_file uses the same provider resolution rules as the CLI. For more control, import resolve_provider_config and transcribe_with_provider from meetdown and keep the provider config in your own application code.

Providers

If --provider is omitted, meetdown picks a provider only when exactly one provider-specific API key is configured. Pass --provider explicitly when more than one provider key is present, or when you pass only a generic --api-key or MEETDOWN_API_KEY.

CLOVA Speech can be selected automatically when you provide a CLOVA Invoke URL and key:

uvx meetdown meeting.m4a -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key"

For CLOVA, --api-url is the CLOVA Speech Invoke URL from NAVER Cloud Platform. You may pass either the base Invoke URL or a URL that already ends with /recognizer/upload.

OpenAI and Gemini can be selected automatically from their provider-specific environment variables. Use --api-url when you need to override the provider endpoint or base URL:

$env:OPENAI_API_KEY="your-openai-key"
uvx meetdown meeting.m4a -o meeting.md --chunk-duration 10m
uvx meetdown meeting.m4a -o meeting.md --provider openai --api-url "https://api.openai.com/v1"

$env:GEMINI_API_KEY="your-gemini-key"
uvx meetdown meeting.m4a -o meeting.md --chunk-duration 10m

Use --model to override the provider default model. CLOVA does not support --model; the CLOVA Speech model is determined by the Invoke URL/domain. OpenAI defaults to a diarization-capable transcription model when diarization is enabled, and to a smaller transcription model when --no-diarization is used. Gemini uses inline audio requests, so use --chunk-duration for larger files.

Chunk long recordings

Use --chunk-duration when a recording is too long for one sync upload. The CLI extracts audio chunks with ffmpeg, sends each chunk to the selected provider, shifts segment timestamps back to the original file timeline, and writes one Markdown file. If your PC does not have system ffmpeg, meetdown falls back to the bundled imageio-ffmpeg executable installed with the package. If system ffprobe is missing, meetdown reads duration metadata through ffmpeg instead.

uvx meetdown meeting.mp4 -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key" --chunk-duration 10m

Accepted duration forms include 600, 600s, 10m, and 1h. Uploads default to --compress smallest, which extracts audio and uploads a 64 kbps MP3 copy. Use --compress lossless for FLAC or --compress none to avoid compression. When chunking or range extraction is used, --compress none creates WAV chunks:

uvx meetdown meeting.mp4 -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key" --chunk-duration 10m --compress lossless
uvx meetdown meeting.mp4 -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key" --chunk-duration 10m --compress none

You can still force the exact temporary upload format with --chunk-format mp3, --chunk-format flac, or --chunk-format wav.

You can transcribe only part of the source file. Timestamps in the Markdown stay on the original media timeline:

uvx meetdown meeting.mp4 -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key" --start 10m --end 45m
uvx meetdown meeting.mp4 -o meeting.md --api-url "https://your.invoke.url" --api-key "your-secret-key" --start 00:10:00 --end 00:45:00 --chunk-duration 10m

Offline conversion

If you already have a CLOVA Speech JSON response, convert it without calling the API:

uvx meetdown --from-json examples/clova_response.sample.json -o sample.md --title "Sample Meeting"

Generated Markdown includes a Processing options section with the provider, model, language, compression, chunking, range, timeout settings, and a replay command for the run. Raw API keys are never written. They are stored only in masked form, such as sk-p********7890. Replay commands use <...> placeholders for secret values; replace them before running the command.

When a provider returns unnamed speakers, meetdown writes speaker placeholders like [[SPEAKER_1]] and [[SPEAKER_2]]. These are intentionally easy to find with Replace All and change to real names after you identify the speakers.

Initial scope

This project intentionally starts small. It supports one local file, provider-backed speech recognition, full text output, speaker-segment output, and Markdown file generation. Summaries, decisions, action items, database storage, web UI, async callbacks, and batch processing are later extensions.