The corpus and the sample
| Scope | Measured coverage | What it means |
|---|---|---|
| Backfill corpus | 10,122,224 clips · 50,612 shards | Includes successful, failed and skipped records. |
| Embedding index | 9,949,618 vectors · 9,541,803 clips | Only clips with exported speaker centroids are searchable. |
| Expanded diagnostic | 2,000 queries · 40,000 returned clips | Uniform sampling over searchable clips, seed 2026102311. |
| Matched background | 10,000 independent clips | Controls match the query dataset and detected language. |
| Measured pairs | 838,411 | Four distinct comparison types, with shared query clusters. |
For each query, choose the local speaker with the greatest embedding support, then search all 9.95 million existing 512-dimensional float32 centroids. Exclude every speaker from the query clip and retain 20 distinct result clips, using the best matching local speaker per result.
The sample contains 1,858 v3 queries, 129 v2 queries and 13 v4 queries. Detected languages include 1,801 English queries plus German, Russian, Korean, French, Japanese, Portuguese, Spanish and Italian. Version-specific conclusions for v4 are limited by its small sample.
The two axes
Speaker cosine on X: the dot product divided by the actual two vector norms, calculated in float64. For two returned results, compare their matched speaker vectors directly. Never reuse either result’s score against the original query as the X value.
Fuzzy transcript similarity on Y: twice the ordered-character longest-common-subsequence length divided by the combined text lengths. Use complete ASR transcripts after Unicode NFKC, casefold, punctuation-to-space and whitespace normalization. Numbers and contractions are not expanded. This is not a percentage of identical words or a semantic similarity model.
Retrieval follows the production raw inner-product ranking on near-unit-norm float32 centroids. The plotted cosine separately divides by actual norms. Every top-20 candidate cutoff lies above the retained search boundary. Original source-array and loaded-manifest SHA256 values were checked.
Interactive coordinates use float32 for compact delivery. Full-precision values and stricter character/word edit metrics are in the downloadable data. The site audits coordinate hashes, and a 1e-7 tolerance keeps displayed threshold filters aligned with the measured results.
Four comparisons
- Query → nearest results: 40,000 query/result pairs.
- Query → matched random: 39,888 query/control pairs.
- Between nearest results: 380,000 pairs within top-20 result lists.
- Between matched random: 378,523 pairs within matched-control lists.
Controls are distinct within each query and use their local speaker closest to that query. Ten rare dataset/language strata have fewer than 20 controls; counts remain explicit. There were no undefined-text pairs to exclude. Between-control comparisons use those selected speakers, matching the structure of the retrieved-pair comparison.
Correlation, repetition and uncertainty
Pearson correlation is 0.169 for query/result pairs, 0.216 between results, and roughly 0.05–0.06 for matched random comparisons. Query-cluster bootstrap 95% intervals are [0.141, 0.199] and [0.193, 0.237] for the two retrieval comparisons. The groups have different selection rules, so a pooled correlation is misleading.
At ≥80% fuzzy similarity, 250/2,000 lists contain at least one non-identical result pair, versus 4/2,000 matched random lists. The largest script family contributes 172 such within-list pairs, and the top ten contribute 371 of 941. Pair counts are therefore not independent incidents.
The association persists in the dominant English-v3 stratum and in single-speaker queries. These checks do not isolate whether wording, voice descriptions, accents, emotion, acoustic context or generation instructions cause the relationship. No global speaker identity labels were established.
Dotted means and percentiles
The explorer recomputes pair-weighted means and P10, median (P50) and P90 for the current filtered observations. Vertical lines summarize X; horizontal lines summarize Y. When several comparison types are active, they are pooled and explicitly labeled. Select one comparison type for a group-specific view.
P90 means 90% of observed values lie at or below that threshold on that axis. These are marginal distributions, not fitted trends, prediction bounds or confidence intervals. Every line is individually switchable.
Verification and provenance
All 52,635 selected vector/member mappings and all retrieval self-exclusions were checked. Every one of the 838,411 cosines was independently recomputed. Independent dynamic-programming text checks covered 48 pairs; three spread query rankings exactly matched the live search service. The maximum coordinate differences were at floating-point roundoff scale.
The standalone report passed offline desktop/mobile checks with no external requests. The hosted explorer preserves every coordinate and transcript, while serving its audio as on-demand same-origin assets. Source corpus objects, models, pipeline code and the production search service were unchanged.