Do closer speaker embeddings retrieve closer wording? Explore an expanded exact-search sample, with matched random controls and the actual speaker vectors behind every pair.
X = actual speaker cosine. The cosine is recalculated from the two specific normalized speaker vectors named by the selected pair. For two returned results, this is their cosine with each other—not either result's similarity to the original query.
Y = full-transcript fuzzy similarity. Ordered-character Indel similarity: twice the longest common subsequence length divided by total text length. Text uses Unicode NFKC, casefold, punctuation-to-space and whitespace normalization. No number or contraction expansion. Exact text matches are included and can be filtered separately.
2,000 fresh uniform clips, top 20 each. Query the best-supported local speaker. Exhaustively search the fixed full index, exclude the entire query clip, and keep unique result clips. Original near-unit-norm inner-product ranking matches production search. Scatter X is separately normalized cosine.
Controls match dataset and detected language. Choose up to 20 distinct clips from a separate 10,000-clip pool, and use each control's speaker closest to the query. Rare strata can have fewer controls; counts are explicit. Between-result/control pairs use the speakers selected for that query.
All observations are plotted. Dense points overlap. Every filtered row is available in the paginated data table. Hover or click a point to inspect the exact pair. The listening set is deliberately selected and does not cover every point; choose “With audio examples” to find it.
Association is not identity or causation. Repeated scripts, voices, accents, genres and generation instructions co-vary. Retrieval conditions on similarity, and pairs share queries; uncertainty is measured by query clusters. Audio previews are compressed full clips, not inputs to these calculations.
Speaker cosine on X. Fuzzy transcript similarity on Y.
Dotted vertical lines summarize X; dotted horizontal lines summarize Y. P10 / median (P50) / P90 mean 10% / 50% / 90% of pairs fall at or below that value on that axis. These are marginal summaries, not fitted trends or confidence intervals.
Audio is available for selected listening examples. All points have full transcripts.
Paginated for readability. The scatter includes the entire filtered set.
| PAIR | COMPARISON | QUERY | SPEAKER COSINE | FUZZY TEXT | CLIP A | CLIP B | AUDIO |
|---|
The nearest-neighbor clouds are selected by retrieval. The matched random clouds are a different sampling distribution. A pooled correlation can therefore be misleading.
The numbers below the scatter describe each comparison type separately. Exact repeats, non-identical near-repeats and between-result pairs can be filtered independently.
At high text similarity, inspect individual script families as well as aggregate counts. A large number of related pairs can come from a single result list.