meetscribe

Meeting audio in → speaker-attributed transcript out. On your own GPU.

Platform Review — weekly
104.0 min · 1074 turns · 16 speakers
==================================================================

[0:00:00] Dana Whitfield
  Meeting open, welcome everybody.
  So we'll start with the apologies and I guess we have Priya, everyone
  else seems to be here, I think. So could I have a mover please?

[0:00:12] Marcus Elle
  Dana, just before you do, I have to leave at 1.30 to attend the
  roadmap session. So sorry about that.

Name a voice once and it is recognised in every meeting after that — including ones you already transcribed.

---

Contents

Quickstart · Speed · Remote GPU · How it works · Requirements · Layout · Licence

---

Quickstart

git clone https://github.com/suzuenhasa/meetscribe-cli.git
cd meetscribe-cli && ./setup.sh        # 2-10 min, mostly 2 GB of weights

cp ~/recordings/*.mp3 inbox/
./transcribe

Each recording becomes a directory in library/ — transcript, audio, and a few seconds of each voice. The inbox empties as they finish.

Everyone starts as Speaker 1. Name them once and they are recognised in every recording afterwards, and in the ones you already have:

./speakers link --apply                   # group each voice across meetings
./speakers review                         # the groups waiting for a name
./speakers name 12 "Dana Whitfield"       # names them in every meeting at once
./speakers apply --apply                  # backfill the transcripts you have

review prints a line of what each voice said and a clip to play, which is usually enough to tell who it is: ./speakers play <meeting> G02.

Supported formats: wav mp3 m4a flac ogg opus aac m4b aiff wma mp4 webm mkv — any sample rate, auto-converted to 16 kHz mono.

Unfamiliar names? Proper nouns the model has never heard come out as the nearest familiar English. Give it a vocabulary — a file per subject, or a list:

./transcribe --glossary crypto.txt
./transcribe --glossary "Dana Whitfield,Northwind"

A glossary.txt beside where you run is picked up automatically.

Keep the engine warm

The engine costs ~60-70 s to load and that is paid per run. On a 3-minute clip it is nearly the whole wall clock: 99 s cold against 6.6 s with it already up, measured on a 3090.

./engine start      # once, then it stays
./engine stop       # hand the card back

If the engine is running, commands use it. Otherwise they load their own.

Every command and flag is in RUNBOOK.md.

Day to day, the loop is in WORKFLOW.md: transcribe, group the voices, name whoever is new, backfill.

---

Speed

One batch, 6 meetings, 6.27 hours of 16 kHz mono WAV, default flags. Three rented boxes, each a fresh git clone and ./setup.sh.

GPU	VRAM	Engine load	6.27 h of audio
RTX 5090	32 GB	56 s	29.5 s — 766×
RTX 3090	24 GB	66 s	70.5 s — 320×
RTX 2060	6 GB	127 s	580 s — 39×

Decode is 85–97% of that time, and every run prints the breakdown. Engine load is once, not per meeting, and the first load on a new machine is ~3× slower while torch.compile builds its cache — 219 s against 66 s on the same 3090.

Feeding it MP3 or any other format adds a conversion pass first, which runs across files in parallel and splits a long recording across cores: 6.27 h of MP3 converts in about 22 s on a 48-core box, once.

---

Remote GPU — optional

Everything above runs on your own card. This is for when you have not got one, or want a faster one for a big batch. Skip the whole section if neither applies.

--host runs the GPU work on another machine. Your library is the thing that stays local — transcripts, clips and speakers.db live here and are the authoritative copy.

The audio itself does go over. It is uploaded, transcribed, and sits in the box's own library until you destroy the instance. Nothing goes to a third-party service, but it is on that machine while it works, so rent from someone you are willing to put the recording on.

One-time setup on the remote box:

ssh mybox 'git clone https://github.com/suzuenhasa/meetscribe-cli.git /opt/meetscribe \
           && cd /opt/meetscribe && ./setup.sh'

Then from your laptop:

./transcribe --host mybox

speakers.db syncs both ways, so the box recognises your people and anything it learns comes home. Destroy the instance — lose nothing.

For rented boxes, vast/provision.sh handles unattended install. See vast/README.md.

---

How it works

audio ──► MOSS on vLLM        →  text + per-window speaker labels
             │
      WeSpeaker ResNet34-LM  →  voice embedding per segment
             │
      constrained clustering  →  Speaker 1, Speaker 2, …
             │
      profile store (sqlite)  →  Dana Whitfield, Marcus Elle

MOSS-Transcribe-Diarize labels speakers within each window, but window 4's "S01" is not necessarily window 5's. WeSpeaker embeddings link segments across a recording, and the profile store puts names on the result — across recordings and across months.

---

Requirements

GPU	NVIDIA, 6 GB+ VRAM, RTX 20-series or newer. Pascal will not run vLLM. The GTX 16-series meets the compute requirement but has no tensor cores and is untested.
System	FHS-style Linux. On NixOS/non-FHS distros, bring your own Python and set library paths for NVIDIA driver, GCC runtime, zlib, plus TRITON_LIBCUDA_PATH.
Tools	ffmpeg

A 6 GB card transcribes and embeds in two passes rather than simultaneously — slower, but complete.

./setup.sh --check verifies your environment without installing anything.

---

Layout

inbox/          drop recordings here; empties as it finishes
library/        one directory per meeting
speakers.db     who people are — the one thing not rebuildable from audio

Back up speakers.db. Everything else can be rebuilt from the audio.

---

RUNBOOK.md — full command reference, flags, and troubleshooting.

---

Licence

Apache License 2.0 — see LICENSE. Use it, change it, ship it, sell it; keep the notices. Contributions are taken under the same terms.

MOSS, WeSpeaker and vLLM are Apache-2.0 as well and are downloaded at install time rather than bundled — NOTICE lists them and setup.sh pins the exact revisions.