better-meeting

Talk to your meeting recordings. bm pulls everything out of a video (a timestamped transcript, the text that appeared on screen, and key screenshots) into one self-contained folder. Give that folder to any LLM or agent as context and ask it anything about the meeting.

bm meeting.mp4
# -> runs/meeting/artifacts/   timeline.md, transcript.md, screens/

It works for any recording: a standup that is all talk, a demo full of terminal commands, a workshop, a plain call.

There is no LLM inside. bm never summarizes or interprets. It only extracts, and the deterministic local pipeline caches every stage. Understanding the meeting is the job of whatever model you connect.

What you get

artifacts/timeline.md interleaves speech, silences and new on-screen text:

# Timeline

- duration: 42:10
- speech languages: uk 71%, en 29%
- speech segments: 612
- screen events: 48
- screenshots in screens/: 30

[12:31] SPEECH: [uk] Давайте подивимось, що з подами на стейджингу.
[12:40] SCREEN (screenshot: screens/007_00-12-40.jpg):
    $ kubectl get pods -n staging
    api-7d9f8c6b5-x2lq   0/1   CrashLoopBackOff   5   3m
[12:44] SPEECH: [en] Okay, so the API is crash-looping again.
[12:58] [silence 14s]

Next to it are transcript.md (speech only), screens/*.jpg (timecode-stamped) and screens_index.md.

Install

You need Python 3.12+ and ffmpeg on your PATH.

# macOS (Apple Silicon): Whisper via MLX and OCR via the Vision framework, out of the box
uv tool install better-meeting

# Linux / Windows: Whisper via faster-whisper; install tesseract-ocr from your package manager
uv tool install "better-meeting[faster]"

Add the [parakeet] extra for the optional Parakeet engine (macOS). Upgrade with uv tool upgrade better-meeting. pipx install works the same way.

On first use the ASR models download from Hugging Face: about 1.6 GB for Whisper large-v3-turbo, 2.5 GB for Parakeet TDT v3.

How it works

The pipeline runs in stages. Each stage caches its result in the video's working folder, so a re-run only redoes what changed:

Each stage records the parameters it was built with. If you change --lang, -O, --frame-interval, --ocr-lang and so on, the affected stages and everything downstream re-run automatically. --force redoes everything.

If OCR finds no readable text (a low-res recording, or --ocr none), screenshots are still picked evenly across the moments the screen changed, so a multimodal model can read them itself.

Multilingual meetings

Whisper's built-in language detection only listens to the first 30 seconds. That is useless for meetings that mix languages, for example a Ukrainian standup with English terms and Russian asides.

With --lang auto (the default), bm transcribes the whole recording once per candidate language (--langs, default uk,ru,en). It then merges the passes: each stretch of the timeline keeps the segments from the pass Whisper was most confident about (avg_logprob). This holds up even when the language changes between sentences. Silence hallucinations are filtered with the standard no_speech_prob and avg_logprob heuristic.

The cost is one pass per language, and each pass is cached on its own. If you know the language, pin it with --lang uk for a single pass. If the meeting is in other languages, set --langs de,en.

Mixed-language transcripts tag every line ([uk], [en]), and the timeline header shows each language's share.

Parakeet: one pass for every language

--asr-backend parakeet uses NVIDIA's Parakeet TDT 0.6B v3 via parakeet-mlx. It is a single multilingual model covering 25 European languages that decodes the recording once, with punctuation. It is several times faster than three Whisper passes.

• The model takes no language input. --langs is ignored and -O is rejected. A pinned --lang only labels segments. With --lang auto, a script heuristic labels each segment uk, ru or en, so treat those tags as a hint.
• macOS only for now. Whisper stays the default.

Usage

bm extract VIDEO [options]     # `bm VIDEO` is the same

Common extract options (run bm extract --help for the full list):

Global flags: -q (errors only), -v (debug), --version.

Examples

bm standup.mov --lang uk                                 # one known language, single pass
bm call-360p.mp4 --ocr none                              # OCR is hopeless at 360p; keep screenshots
bm demo.mkv -O initial_prompt="Kubernetes, Helm, Argo"   # bias Whisper toward domain terms
bm demo.mkv --asr-backend parakeet                       # fastest multilingual run (macOS)
for f in recordings/*.mp4; do bm "$f"; done              # batch: one folder per file

One folder per video

Every recording gets its own working folder, runs/<video-name>/ by default. Running the same file again reuses its cache, and a different file lands in a different folder. --out-root changes the root, and --out pins an exact folder.

runs/standup-2026-08-20/
├── audio.wav, transcript.json, keyframes.json, screen.json, *.meta.json   # stage caches
├── pass_uk.json, pass_en.json, ...                                         # per-language ASR passes
├── thumbs/, frames/
└── artifacts/        <- hand this folder to your LLM

Point queries for agents

Point commands let an agent drill into one moment without re-running the pipeline. Data goes to stdout and logs to stderr:

bm frame meeting.mp4 -t 12:40                 # -> runs/meeting/frames_at/00-12-40.jpg
bm ocr meeting.mp4 -t 12:40                   # recognized text
bm ocr runs/meeting/artifacts/screens/007_00-12-40.jpg --backend tesseract   # second opinion
bm transcript meeting.mp4 --from 10:00 --to 12:30 --lang en
bm transcript meeting.mp4 --from 10:00 --to 12:30 -O initial_prompt="Argo, Helm" --json

Timecodes accept SS, MM:SS or HH:MM:SS. bm transcript results are cached by their full parameter set. If extract already ran with the same model, backend and options, the range is sliced from its cache and no ASR runs at all.

Talking to the video

Tell your agent: "better-meeting is installed as bm; run bm agents and follow that guide to work with <video>". The guide covers extraction, what each artifact means, and the point commands. Then ask anything, for example "what did we decide about the release?", "what command did they run at 12:40?" or "summarize each person's status".

Providers

Speech recognition and OCR are pluggable providers behind small interfaces (AsrBackend, OcrBackend):

Every built-in provider runs locally today, and the interfaces are designed for hosted providers as well. A cloud speech-to-text or vision API is a provider that reads its API key from the environment. The pipeline, caching, language merge and artifact format do not change. See CONTRIBUTING.md for how to add one.

Roadmap

• Hosted ASR and OCR providers (API-key based), next to the local ones
• Parakeet on Linux
• Speaker labels (diarization)

Have a use case? Open an issue.

Contributing

See CONTRIBUTING.md. Contributions are under Apache 2.0, with a DCO sign-off.

License

Apache 2.0. Third-party model attributions are in NOTICE.