Listen to the overlap.
Almost the same words. Sometimes an entire page of them. A listening notebook about the script families that surface in speaker-embedding search.
result pair with ≥80% fuzzy similarity.
The rest is nearly word for word.
Plus one exact transcript match.
What we measured, what the scores mean, and how to listen
200 uniformly sampled queries. Full-corpus search. Each query searched all 9,949,618 existing speaker vectors and returned 20 distinct clips, excluding the query clip. These 50 selected comparisons illustrate the strongest examples; they are not a representative sample.
The large score measures text, not voice. Fuzzy similarity compares complete normalized transcripts while preserving character order: 2 × longest common subsequence length ÷ combined character lengths. Exact normalized matches are excluded from these examples. Character-edit and word-edit scores offer stricter checks. Highlights show unmatched words, rather than the score calculation itself.
Repeated language is concentrated. At ≥80% fuzzy similarity, 29/200 result lists contained a similar pair; matched random lists had none. With the stricter character-edit score, 8/200 lists qualified. Shared scripts, voice descriptions, accents and scenarios can co-vary. These observations do not establish speaker identity or a corpus-wide duplicate rate.
| Fuzzy threshold | Result lists | Matched random |
|---|---|---|
| ≥70% | 83 / 200 | 1 / 200 |
| ≥80% | 29 / 200 | 0 / 200 |
| ≥90% | 3 / 200 | 0 / 200 |
Listen to A, then B. The previews include the whole clip, which can contain multiple speakers. They are MP3 at 96 kbps, mono, without trimming or gain normalization. Source audio hashes were verified before conversion. Similarity scores come from the original analysis and are unaffected by these previews.
Unicode, case and punctuation are normalized; numbers and contractions are not expanded. Scores compare whole-clip ASR text, even when the search query uses one local speaker.
The listening room.
Hear the evidence. Words that differ are highlighted.
35 UNIQUE AUDIO CLIPS