Two Saudi preachers, one orthodoxy, opposite instruments — the classical division of pulpit
labor between warning and reassurance (الترهيب والترغيب), measured across 24 sermons and lectures.
The one-paragraph version. The hypothesis behind this pair was that two scholars of the
same country, generation and creed would differ almost purely in delivery — and the data agrees,
then sharpens it. Al-Arefe preaches the classical rhetoric of warning: admonition
carries a third of his emotional weight, his pitch swings half again as wide as his counterpart's, and his
sermons run on colloquial mini-dramas — staged everyday dialogues that pull the listener into the story.
Al-Maghamsi teaches: calm instruction and reverence dominate, delivered in a
narrower, steadier melodic band anchored in Quranic citation and liturgical formula. And beneath the opposite
instruments, one identical engine: their linguistic certainty is statistically indistinguishable —
same creed, different nervous system.
Explore every sermon
Each dot is one recorded sermon or lecture. Pick a metric, hover for a preview, click for the full
profile and video link.
Click a dot to see the sermon profile.
The division of pulpit labor
The gap-sorted comparison reads like a definition of الترهيب والترغيب. Admonition — hellfire, reckoning,
consequence — belongs to al-Arefe (−18 pp). Calm instruction (+15 pp) and reverence
(+7 pp) belong to al-Maghamsi. Urgency and indignation follow al-Arefe; grief, hope
and exaltation lean al-Maghamsi. Per sentence, admonition dominates 18% of al-Arefe's talk against 4% of
al-Maghamsi's.
Homiletic registers, gap-sorted dumbbells: dot = share, connector = the difference. Left: statement-level weighted emotions. Right: dominant register per sentence.
فإذا فنيت حسناته قبل أن يُعطى ما عليه أُخذ من خطاياهم فطُرحت عليه ثم طُرح في النار عياذًا بالله
al-Arefe — the admonition register: the bankrupt soul on the Day of Reckoning
ولقد وصينا الذين أوتوا الكتاب من قبلكم وإياكم أن اتقوا الله
al-Maghamsi — the instruction register: Quranic citation as the sermon's spine
A stylistic signature the palette almost missed: al-Arefe's "urgency" segments are largely
storytelling — everyday dialogues in colloquial Gulf Arabic ("how much does he owe you? — ten
riyals, wallahi ya rabbi"), staged as miniature moral dramas. Al-Maghamsi's corresponding segments cite
scripture. Fear-drama versus scripture-calm, down to the sentence level.
The voice: half again as much melody
The vocal channel is where this pair was predicted to split, and it does: al-Arefe's median sermon moves
through a 60 Hz pitch interquartile range against al-Maghamsi's 41 — and both dwarf every
politician we have measured with the same instruments (17–25 Hz for Biden, Trump, Netanyahu, Lapid).
Vocal arousal and dominance tell the same story twice: al-Arefe 0.79/0.75 vs al-Maghamsi 0.76/0.72 — high
and higher. The pulpit out-projects the podium: the mosque register beats every political
speaker on prosodic command (politicians: 0.48–0.62).
Pitch variability per sermon; bar = median.Speech rate — nearly identical (1.82 vs 1.75 w/s): the difference is melody, not tempo.
Vocal arousal, dominance, valence per sermon, and vocal dominance vs. text certainty per sentence.
Same canon, same map
Semantically the two preach from one world: the clusters — God and His servants, the Prophet, human
conditions and divine will — are shared territory, and message consistency across sermons is essentially
identical (0.978 vs 0.980; sermon corpora are the most internally consistent genre we have measured, above
even IDF war briefings). One structural difference: al-Arefe's openings are near-verbatim formula
(similarity 0.90 vs 0.73) — the fixed khutba preamble — while al-Maghamsi varies his entries and instead
standardizes his closings (0.79 vs 0.75).
Each dot is one transcript sentence, placed by meaning; shared PCA axes.Share of each preacher's sentences per semantic cluster, gap-sorted.
Every sermon rises at the end
Emotional arcs show a genre law rather than a personal signature: both canonical shapes end with a sharp
climb in the final tenth — the closing duʿāʾ, where warning resolves into supplication and glad tidings.
The two preachers distribute nearly evenly across both shapes: the arc belongs to the liturgy; the voice
inside it belongs to the man.
The two canonical sermon arcs (standardized valence): both resolve upward at the close.
Methods
Audio via yt-dlp; transcription: Whisper large-v3-turbo run locally (mlx-whisper), Arabic, segment timestamps.
Text annotation: google/gemini-2.5-flash reading the Arabic — per segment: dominant register from a
12-emotion homiletic palette (reverence, admonition, glad tidings, grief, compassion, exaltation, urgency,
calm instruction, indignation, humility, polemic, hope), certainty 0–100, hedging, valence −100..100.
Vocal dimensions: audeering wav2vec2 (MSP-Podcast) — arousal/dominance/valence per segment;
pitch via librosa YIN.
Statistics: bootstrap 95% CIs on per-sermon medians; hygiene: Whisper hallucination loops deduplicated
(one sermon required forced-Arabic retranscription after a subtitle-credit hallucination).
Caveats
12 sermons per speaker: medians are directionally solid (bootstrap CIs shown on the scoreboard where they
matter) but fine-grained contrasts should not be over-read.
Corpus curation leaned on official channels for al-Maghamsi and on re-uploads for al-Arefe (his originals
are scattered); recordings span years and mosques, so room acoustics vary within each corpus.
The vocal model is English-trained: only the contrast between the two speakers is interpreted, never
absolute values. Mosque PA reverb and audience sounds are part of both corpora symmetrically.
Text ratings come from a single LLM reading the Arabic; the homiletic palette was purpose-built for this
genre. Whisper Arabic is strong on the fuṣḥā register both preachers use, noisier on colloquial asides —
which al-Arefe uses more, so his transcripts likely carry slightly more recognition noise.