Hagari vs. Defrin

Every spokesperson update of the war — transcribed, annotated for emotion, certainty and scripted-vs-spontaneous speech, scored acoustically, and mapped over 33 months.

307 videos · 298 Hebrew statements analyzed · 37.6 h audio · 245k words · 33,072 segments · pipeline: yt-dlp → mlx-whisper → Gemini 2.5 Flash (text) + wav2vec2 A/D/V (voice) + F0 + multilingual embeddings

Statements

298

214 Hagari · 84 Defrin · 7 English aside

With press Q&A

223

the scripted-vs-spontaneous corpus

Hedging: script → Q&A

×4.8

0.5% → 2.2% of segments

Template across handover

held

delivery jumped; the writers stayed

Narrated valence

+7 → +17

Hagari era → Defrin era

Explore all 298 Hebrew statements

Each dot is one televised statement. Pick a metric, hover for a preview, click a dot for its full profile and video link. Shaded bands mark war phases; the dashed line is the Hagari→Defrin handover (March 2025).

Click a dot to see the statement's profile.

The scriptwriter test

Collaborator hypothesis: the briefing template (opening → arena survey → closing) survived the change of spokesperson because the statement writers stayed. Test: compare the semantic similarity of statement openings and closings across the handover boundary against the similarity within each speaker's own era. Result: cross-boundary similarity (0.275 openings / 0.373 closings) is indistinguishable from within-era similarity — the template passed through the handover intact — while delivery jumped at the same boundary (speech rate +21%, pitch variability −61%). The script stayed; the voice changed.

Template similarity across the handover vs delivery metric jumps — Left: opening/closing similarity — within Hagari (blue), across the boundary (grey), within Defrin (red); ±12 statements around the handover. Right: relative change in delivery metrics across the same boundary.

The gem: same man, script vs. spontaneous

221 statements contain a press Q&A — the one part of the podium no scriptwriter controls. Comparing each statement's scripted part against its own spontaneous answers (paired design: same man, same day, same events — the period confound cancels exactly):

Hedging ×4.8 (0.5% → 2.2% of segments; Wilcoxon p<10⁻⁶). "אנחנו בודקים" lives almost entirely off-script.
Certainty drops (99.5 → 97.6, p<10⁻¹⁴) — and the drop is much steeper for Hagari than for Defrin, who holds his assertiveness under questioning.
Speech accelerates (1.94 → 2.33 words/s, p<10⁻³⁶): scripted delivery is deliberately paced; spontaneous speech runs free.
Vocal dominance and arousal both ease off (p<10⁻⁸) — the projected command of the scripted register relaxes when answering.
Median pitch doesn't move (p=0.42) — same voice, sanity check passed. But the speakers diverge: Hagari's pitch settles in Q&A, Defrin's rises.

Paired slopegraphs: scripted vs Q&A for six metrics, each line one statement — Each thin line is one statement (scripted mean → Q&A-answers mean); thick lines are speaker medians.

Hedging share in scripted text vs Q&A answers per speaker — Hedging lives in the answers.

Pitch z-score distributions pooled and per speaker, plus two-sample QQ plot — Validation: journalist questions sit a median 2.9 SD off the spokesperson's scripted pitch band (64% beyond 2 SD) — the role labels are acoustically real. The QQ panel is the formal version: both speakers' question quantiles leave the identity line immediately. The Defrin-era gap is larger (median 7.7 vs 2.2 SD) — likely a different pressroom/mic setup, and off-mic audio makes pitch estimates noisier.

Emotional arcs (Reagan et al.'s machinery, modern instruments)

Per-statement valence arcs on a normalized 0→1 axis, decomposed with SVD and clustered by shape — the "emotional arcs of stories" approach, instrumented with an LLM reading the Hebrew and a prosody model listening to it. Two findings:

Briefings are not novels. The top arc mode explains only 16% of variance (novels: a third or more) — the genre is event-driven, not story-shaped. Still, two canonical shapes emerge: man-in-a-hole (composed open → grim middle → rally close; 162 statements, 60% of Hagari's) and recovery (open with the bad news, climb, stay up; 121 statements, half of Defrin's). The end-of-statement rally is exactly where the scripted closing lives.

The voice barely traces the script's arc. Within-statement correlation between the text's valence arc and the voice's is r≈0.10 (certainty-vs-dominance r≈0.15) — broadcast delivery is emotionally flat relative to content, consistent with the Q&A findings.

SVD modes of standardized valence arcs — Basic arc modes.

Two canonical arc shapes with speaker shares — The two canonical briefing arcs.

Mean statement valence over 33 months with handover marked — The macro arc: the war as narrated from the podium. The Defrin era runs ~10 valence points sunnier than the Hagari era.

Thirty-three months in four signals

Timelines of valence, certainty, speech rate, pitch variability across the war — Grief and certainty dips align with events (the Bibas week of Feb 2025 is the deepest valence dive, −43); the speech-rate step at the dashed handover line is the cleanest speaker signature in the corpus.

Message novelty over time — Message novelty (distance from the previous five statements). Spikes: the first hostage deal (Dec 2023), Rafah (May 2024), the Lebanon escalation (Oct 2024).

Small multiples: share of segments per emotion by quarter — One panel per emotion, zero-based scales. Calm erodes as the war grinds on; pride climbs into the Defrin era; grief double-peaks (Bibas week, then the Oct-2025 hostage deal); anger spikes exactly once — the Bibas week.

What survived the 6.6× scale-up — and what didn't

The first pass of this study compared 44 statements; this corpus is 292. Honest scoreboard:

Finding (44-statement pass)	At full scale
Defrin speaks faster	Holds — 2.00 vs 1.74 w/s median, stable in every period
Hagari's voice more expressive (pitch movement, dominance)	Holds, smaller margin
Defrin narrates a sunnier war	Holds and grows — valence +17 vs +7
Identical within-statement structure	Holds — 0.406 vs 0.408
Defrin's language more certain	Gone — 97.7 vs 97.6; the v1 gap was Hagari's fog-of-war October
Defrin's messaging more repetitive	Gone — identical novelty medians; the v1 gap was his June-2025 Iran streak

Methods

Collection: Playwright (headless Chromium) to scrape idf.il past its bot protection; yt-dlp for YouTube audio extraction.
Transcription: OpenAI Whisper large-v3-turbo run locally via mlx-whisper (Apple Silicon), automatic language detection, segment-level timestamps.
Text annotation: google/gemini-2.5-flash (via OpenRouter) reading the Hebrew directly — per segment: dominant emotion (10-label palette), certainty 0–100, hedging flag, valence −100..100, and speech role (prepared / question / answer); per statement: confidence, coherence, topics, tone.
Vocal dimensions: audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim (trained on MSP-Podcast, natural speech) — continuous arousal / dominance / valence per segment.
Acoustics: librosa YIN pitch tracking (per-segment median F0, voiced ratio); speech rate and pause structure from transcript timing.
Semantics: sentence-transformers/paraphrase-multilingual-mpnet-base-v2 embeddings — semantic clustering (KMeans, k by silhouette), within/between-statement coherence, message novelty, and the opening/closing template similarity behind the scriptwriter test.
Emotional arcs: the Reagan et al. (2016, EPJ Data Science) machinery — normalized narrative time, SVD mode decomposition, KMeans shape clustering — instrumented with the LLM and vocal signals above.
Statistics: paired within-statement design for script-vs-Q&A (Wilcoxon signed-rank); two-sample QQ comparison; acoustic (pitch-band) validation of the LLM's role labels.

Tried and failed

UMUTeam/w2v-bert-emotion-en (categorical speech-emotion, trained on English acted corpora): did not transfer to Hebrew broadcast audio — near-uniform outputs (mean top-class probability 0.172 vs a 7-class uniform of 0.143, zero segments above 0.3, 64% "surprise" bias). Replaced by the dimensional audeering model.
Whisper on noisy stretches: hallucination loops (repeated "תודה רבה") on applause/ambient audio produced two degenerate transcripts (excluded); one duplicate upload was misdetected as Portuguese from pre-briefing room chatter and transcribed as multilingual garbage (dropped as a duplicate).
LLM speech-level scores (confidence/coherence 0–100): ceiling effects — median 90/95 for both speakers, no discrimination. The segment-level certainty measure replaced them as the working instrument.
labMT lexicon (Reagan et al.'s original sentiment instrument): no Hebrew edition exists — per-segment LLM valence used instead.
Plain HTTP scraping (curl / simple fetchers) of idf.il: blocked by Incapsula bot protection; only a real browser with JavaScript execution got through. YouTube downloads likewise failed on an older yt-dlp (403 / SABR streaming) until upgraded.

Caveats

Corpus: the IDF spokesperson-updates YouTube playlist (295 videos; 2 deleted, 6 duplicate uploads removed), deduplicated against the earlier tag-page scrape, then reconciled against the site's own tag pages: 8 statements missing from the playlist (mostly 2026) were recovered from article-page embeds. Speaker attribution by date against the March-2025 role transition; 5 English-language items kept separately as a parallel corpus.
Whisper Hebrew output carries occasional mis-recognitions ("עם כלבי" for מבצע "עם כלביא"); the error rate is not yet human-quantified.
Text ratings come from a single LLM reading the Hebrew directly — no second-model or human cross-check yet; the Q&A role labels are the one annotation validated independently (acoustically).
Vocal dimensions (arousal/dominance/valence): English-trained model — only contrasts, never absolute values, are interpreted. Pitch variability here = IQR of per-segment median F0 (a different, more conservative measure than the first pass's frame-level IQR).
Period ≠ speaker: eras differ in events. The paired script-vs-Q&A design and same-period controls address this where possible; emotion/topic contrasts remain partly event-driven.
English parallel check (3 date-matched pairs): confidence and dominant emotions track across languages; topics differ by audience (international: hospitals narrative; domestic: fronts and hostages).