~/projects/byu-speeches/laughter
What makes a hall laugh
A BYU devotional is a room of several thousand students, and the recording keeps what the room did. This study listens for laughter and applause in a stratified sample of the audio with a general-purpose sound tagger, maps each laugh back to the transcript by its position in the talk, and asks when speakers get laughs, where in a speech they come, and whether the words alone can say which paragraphs earned them.
What stands out
- 83% of the sampled speeches got at least one laugh, 402 events in all, 161 of them loud enough to be unmistakable through a lectern microphone.
- The 1970s were the funniest decade in the sample at 2.5 laughs per ten minutes; the 1990s the least at 0.61. The older recordings are also the ones most likely to have been edited.
- BYU faculty get the most laughs (3.17 per ten minutes); Quorum of the Twelve the fewest (0.94). The people the hall came to hear joke least.
- 46% of laughs come in the first three tenths of a speech. The opening is where speakers earn the room. Applause at the close was detected in only 3% of the sample: a devotional ends in a prayer, not an ovation.
- Mary Ellen Edmunds's “If Thou Art Merry, Praise the Lord” (2004) got the most laughs in the sample, 9.13 per ten minutes across 29.6 minutes.
- The loudest laugh in the sample lands on a transcript paragraph that ends They are more willing to be impressed than they are eager to impress.” (Kate L. Kirkham, 1987), which is what a story punchline looks like on the page.
- The words alone are a weak predictor: AUC 0.633. Timing and delivery are not in the transcript, and the paragraph model knows it; the Try-it box lets you test a line against it.
Loading the study…
Model card
- Data
- 80 BYU speeches with audio, ten per decade from the 1950s, drawn with a fixed seed from the speeches of twelve to seventy minutes (sample_audio.py); the official mp3 files are read once and only per-window scores are kept. Speeches without a full transcript (older audio-only recordings) count toward the rates but not toward the text model.
- Detection
- PANNs Cnn14, a convolutional network trained on AudioSet, scores four-second windows every two seconds for 527 sound classes; laughter is the maximum of Laughter, Giggle, Chuckle and Belly laugh, applause of Applause and Clapping. Audience sound is faint behind the speaker's microphone, so each speech is thresholded against its own median plus six median absolute deviations with a floor of 0.04; adjacent windows merge into one event; loud means a peak of 0.10 or more.
- Alignment and model
- The transcript has no timestamps. Opening and closing music are found from the tagger's Music and Singing scores, and each event's position within the speaking span maps to the transcript by cumulative word count. The paragraph model is a logistic regression over TF-IDF words and bigrams (minimum three documents), balanced classes, five folds with speeches held out; the browser demo recomputes the same features from the exported weights.
- Known failure modes
- A general sound tagger was not trained on audience laughter under speech, so quiet chuckles are missed and the scores are small; a vigorous speaker can trigger Laughter on their own voice. Proportional alignment slips by a paragraph when a speaker ad-libs, skips text, or the recording includes an introduction. Edited recordings, common before the 1990s, may have had audience reaction trimmed. Thresholds were chosen by inspection, not against hand labels.
Methods and data for the corpus as a whole.