diff --git a/CHANGELOG.md b/CHANGELOG.md
index 556de53..f829405 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -21,6 +21,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   edge; `performer_count` reports how many perform in a span as a high percentile of the
   per-frame counts (the wide shots), with median and maximum beside it. Exact for a soloist,
   a duo and a five-piece band from an operated concert camera; a choir is under-counted.
+- `performer_count(..., stat="typical")` takes the median over framings instead of the widest, for
+  lecture halls where the widest framings are the hall and the slides; `detect_people` and
+  `make_proxy` / `camera_motion` take `ffmpeg_input_args` (e.g. `["-hwaccel", "cuda"]`).
 - **Camera cuts and PTZ** — new `_camera` module. `camera_motion` labels each sample of a video
   `still`, `moving` or `cut` from ORB matches and a partial-affine RANSAC fit between consecutive
   frames of a small 2 fps proxy (`make_proxy`), and returns cuts, shots and the share of time in
diff --git a/docs/user-guide/people-and-camera.md b/docs/user-guide/people-and-camera.md
index c9d1888..0c9e95b 100644
--- a/docs/user-guide/people-and-camera.md
+++ b/docs/user-guide/people-and-camera.md
@@ -45,5 +45,18 @@ Two uses follow:
   where the plain percentile counted the audience in the band's wide shot; a choir stays
   under-counted because singers occlude each other.
 
+## Lecture halls and long files
+
+A lecture or defence recording differs in two ways. There is no raised stage, so pass
+`head_below=1.0, cut_head_below=1.0` to switch the audience filter off; and the widest framings
+show the hall and the projected slides, whose figures are not the speakers, so count the *typical*
+framing: `performer_count(det, a, b, camera=cam, stat="typical")` takes the median over framings
+instead of the maximum. On a defence this gave 1, 1, 2, 2 for the lecture, the introduction and
+the two opponent discussions.
+
+A 36 GB 1080p50 file decodes on the GPU by passing ffmpeg input options through:
+`detect_people(video, ffmpeg_input_args=["-hwaccel", "cuda"])` and
+`camera_motion(video, ffmpeg_input_args=["-hwaccel", "cuda"])`.
+
 Both analyses cache well: keep the proxy and the detections next to the recording and the
 counts for any span are instant.
diff --git a/musicalgestures/_performers.py b/musicalgestures/_performers.py
index 8fa936a..61d9e8b 100644
--- a/musicalgestures/_performers.py
+++ b/musicalgestures/_performers.py
@@ -108,7 +108,7 @@ def people_track(detections: dict, **filter_kw) -> tuple[np.ndarray, np.ndarray]
 
 def performer_count(detections: dict, start_s: float = 0.0, end_s: float | None = None,
                     percentile: float = 90.0, camera: dict | None = None, min_framing_s: float = 10.0,
-                    **filter_kw) -> dict:
+                    stat: str = "widest", **filter_kw) -> dict:
     """How many performers a span shows.
 
     Without `camera`: ``estimate`` is the `percentile` of the per-frame counts (the wide shots),
@@ -133,7 +133,8 @@ def performer_count(detections: dict, start_s: float = 0.0, end_s: float | None
             if m.sum() >= 5:
                 vals.append(int(np.percentile(c[m], 75)))
         if vals:
-            return {"estimate": max(vals), "low": min(vals), "median": int(np.median(c)), "max": int(c.max()),
-                    "frames": int(c.size), "framings": len(vals), "method": "framings"}
+            est = max(vals) if stat == "widest" else int(np.median(vals))
+            return {"estimate": est, "low": min(vals), "high": max(vals), "median": int(np.median(c)), "max": int(c.max()),
+                    "frames": int(c.size), "framings": len(vals), "method": f"framings-{stat}"}
     return {"estimate": int(round(float(np.percentile(c, percentile)))), "low": int(np.median(c)),
             "median": int(np.median(c)), "max": int(c.max()), "frames": int(c.size), "method": "percentile"}
diff --git a/tests/test_camera.py b/tests/test_camera.py
index 5defeda..f70acf4 100644
--- a/tests/test_camera.py
+++ b/tests/test_camera.py
@@ -53,4 +53,4 @@ def test_performer_count_per_framing_ignores_the_moving_camera():
         n = 1 if t < 15 else (4 if t < 20 else 3)     # close shot, then the camera swings past a crowd, then wide
         frames.append({"t": float(t), "boxes": [[0.1 * k, 0.2, 0.1 * k + 0.08, 0.7, 0.9] for k in range(n)]})
     c = performer_count({"fps": 1.0, "frames": frames}, 0, 40, camera=cam)
-    assert c["method"] == "framings" and c["estimate"] == 3 and c["low"] == 1
+    assert c["method"] == "framings-widest" and c["estimate"] == 3 and c["low"] == 1