Benjamin Netanyahu vs. Yair Lapid

19 statements · 17,808 words · palette: 12 emotions · model: google/gemini-2.5-flash

Speech rate w/s

1.702CI 1.624–1.7881.981CI 1.72–2.085

Pitch variability Hz

21.225CI 14.65–22.9517.7CI 15.0–24.0

Certainty /100

89.809CI 67.78–95.76192.955CI 79.832–100.0

Valence

15.101CI 0.736–51.70340.0CI -1.379–47.034

Vocal dominance

0.616CI 0.574–0.6270.551CI 0.492–0.583

Explore every statement

Each dot is one statement. Pick a metric, hover for a preview, click for the full profile and video link.

Click a dot to see the statement profile.

Audio: yt-dlp (optional per-URL time spans). Transcription: Whisper large-v3-turbo (mlx-whisper, local), auto language detection with per-video verification.
Text annotation: google/gemini-2.5-flash reading the source language — per segment: emotion (12-label palette), certainty, hedging, valence, speech role (prepared/question/answer); per statement: confidence, coherence, topics, tone.
Vocal dimensions: audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim (arousal/dominance/valence). Pitch: librosa YIN per segment.
Semantics: paraphrase-multilingual-mpnet-base-v2 embeddings — clustering (KMeans + silhouette), coherence, message discipline, formulaicity.
Statistics: bootstrap 95% CIs on medians; paired Wilcoxon for the script-vs-Q&A contrast; role labels validated acoustically against the speaker's pitch band.

Corpus sizes: Benjamin Netanyahu n=10, Yair Lapid n=9. Findings below ~10 statements per speaker are unstable.
Same-genre and same-language comparison assumed (v1); differences may still reflect venue and period, not persona.
Vocal dimensions come from an English-trained model — interpret contrasts between the speakers, never absolute values.
Text ratings come from a single LLM; the Q&A role labels are the one annotation validated independently (acoustically).
Whisper transcription noise is not human-quantified; hallucination loops are deduplicated and degenerate transcripts excluded.
Script-vs-Q&A section skipped: too few statements with a detectable Q&A.