diff --git a/performance/README.md b/performance/README.md
index b58257f..1bdfc89 100644
--- a/performance/README.md
+++ b/performance/README.md
@@ -30,3 +30,4 @@
 - [memo](memo/README.md): memo's own lists — paging past 500 rows, what a brief answer saves, and `people list` with kinds kept in the store.
 - [funnel](funnel/README.md): wirecat-funnel, the tsnet helper for Tailscale Funnel: whether Tailscale's
   log upload stays off, and Funnel end to end.
+- [beeline](beeline/README.md): Beeline Cloud PBX call recordings, starting with local transcription speed.
diff --git a/performance/beeline/README.md b/performance/beeline/README.md
new file mode 100644
index 0000000..1f696d4
--- /dev/null
+++ b/performance/beeline/README.md
@@ -0,0 +1,6 @@
+# Beeline
+
+Measurements of beeline, the CLI for Beeline Cloud PBX call recordings.
+
+- [transcription](transcription/README.md): how fast local CPU recognition transcribes stereo calls, one channel at a
+  time (2026-11-13).
diff --git a/performance/beeline/transcription/README.md b/performance/beeline/transcription/README.md
new file mode 100644
index 0000000..e2597a7
--- /dev/null
+++ b/performance/beeline/transcription/README.md
@@ -0,0 +1,56 @@
+# Local call transcription speed in beeline
+
+Measured 2026-11-13.
+
+## Why it was run
+
+The owner wanted to show a potential client a working demo of beeline, and asked how fast it transcribes calls.
+beeline recognises speech on the user's own machine, on the CPU, with no GPU and no cloud service.
+
+The question: is CPU-only local recognition fast enough to keep up with a call center's daily volume?
+
+The decision it fed: ship local transcription in beeline 0.1.0, without a GPU or a cloud service.
+
+## Setup
+
+- **Engine.** `sherpa-onnx` 2.0.0 compiled to WebAssembly, on the CPU, 1 thread.
+- **Model.** `gigaam-v3`, Russian with punctuation, about 232 MB.
+- **Audio.** Real stereo calls from the owner's account. Channel 0 is the manager and channel 1 is the client. Each
+  channel is recognised separately.
+- **beeline.** A build of beeline-cli 0.1.0 (cli-core 0.647.0) from a local checkout.
+- **Machine.** The owner's laptop: AMD Ryzen AI 9 HX 470 (24 threads), 30 GB RAM, Linux 7.0 (Ubuntu), Node
+  v24.19.0.
+- **Run.**
+
+  ```sh
+  performance/beeline/transcription/run.sh <profile> <beeline-cli checkout> [calls]
+  ```
+
+  [run.sh](run.sh) reads the profile's token from the system keyring. It then runs beeline in a throwaway `HOME`
+  with D-Bus turned off, so the run does not touch the owner's own store or keyring entry. The model folder is
+  linked in from the real one, so the model is not downloaded again. The script syncs the calls of the last two
+  days, then times `beeline <profile> meetings transcribe --max-meetings <calls> --json`. The time is wall time
+  for the whole command. The audio length comes from the command's own `audio_seconds` total.
+
+## Results
+
+| Audio | Wall time | Speed |
+|---|---|---|
+| 1644 s (27.4 min) | 222 s (3.7 min) | 7.4× faster than real time |
+
+The run used the owner's real calls, so this page gives totals only. It has no call ids, names, phone numbers or
+text.
+
+## What it showed
+
+- One CPU thread transcribes about 7 hours of calls in one hour of work.
+- The owner took that as fast enough for a call center's daily volume, so beeline 0.1.0 ships local transcription
+  with no GPU and no cloud service.
+
+## Not measured
+
+- Several threads.
+- Accuracy. There was no word error rate check.
+- macOS and Windows.
+- Calls longer than 60 minutes.
+- Memory use.
diff --git a/performance/beeline/transcription/run.sh b/performance/beeline/transcription/run.sh
new file mode 100755
index 0000000..53730f3
--- /dev/null
+++ b/performance/beeline/transcription/run.sh
@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+# Usage: run.sh <profile> <beeline-cli checkout> [calls]
+# Reads the profile's token from the keyring, then runs beeline in a throwaway HOME with no D-Bus,
+# so the run never changes the owner's own state or keyring entry.
+set -euo pipefail
+[ $# -ge 2 ] || { echo "usage: $0 <profile> <beeline-cli checkout> [calls]" >&2; exit 2; }
+profile=$1 root=$2 n=${3:-3}
+BEELINE_TOKEN=$(secret-tool lookup service beeline-cli username "token:$profile" 2>/dev/null) || true
+[ -n "$BEELINE_TOKEN" ] || { echo "no token for profile $profile" >&2; exit 1; }
+export BEELINE_TOKEN
+sandbox=$(mktemp -d); trap 'rm -rf "$sandbox"' EXIT
+mkdir -p "$sandbox/data/WireCat/models"
+ln -s "$HOME/.local/share/WireCat/models/audio" "$sandbox/data/WireCat/models/audio"
+b() { HOME="$sandbox" XDG_CONFIG_HOME="$sandbox/config" XDG_DATA_HOME="$sandbox/data" XDG_STATE_HOME="$sandbox/state" \
+  XDG_CACHE_HOME="$sandbox/cache" MESSAGING_STORE="$sandbox/wirecat.db" DBUS_SESSION_BUS_ADDRESS=disabled: \
+  BEELINE_NO_UPDATE_CHECK=1 node "$root/dist/bin/beeline.js" "$profile" "$@"; }
+since=$(node -p 'new Date(Date.now() - 2 * 864e5).toISOString().slice(0, 10) + "T00:00:00Z"')
+b meetings sync --since-time "$since" --json >/dev/null
+start=$(date +%s.%N)
+b meetings transcribe --max-meetings "$n" --json 2>/dev/null | node -e '
+let s="";process.stdin.on("data",d=>s+=d).on("end",()=>{const r=JSON.parse(s).item
+const wall=(Date.now()/1e3)-Number(process.argv[1]);
+console.log(`calls ${r.transcribed}, audio ${Math.round(r.audio_seconds)} s, wall ${Math.round(wall)} s, ` +
+`speed ${(r.audio_seconds/wall).toFixed(1)}x real time, failed ${r.failed.length}`)})' "$start"