VOICE & WORDS / Research observatorySNAPSHOT · 23 OCT 2026
Generated speech, examined

Where voices
meet words.

A research view of the H3 speech corpus. Explore speaker embeddings, follow repeated scripts, and listen to the actual clips behind the points.

Explore the full scatter ↗Listen to paired clips

838,411 measured pairs · 2,000 exact-search queries
Built from the frozen 23 October 2026 corpus snapshot.

Voice similarity × text similarityMEASURED DATA
Retrieved pairsBetween retrieved resultsMatched random controls
Preview: 4,000 sampled pointsOpen all 838,411 pairs ↗
10.12M10,122,224 clip records across 50,612 shards.
9.54M9,949,618 clip-local speaker centroids.
2,000Uniform random queries, 20 exact matches each.
6.61%Corpus word-weighted prompt agreement, not human ASR accuracy.
What stands out

Related voices. Familiar wording.

Retrieved clips show more text overlap than matched random clips. The association is modest, and a few script families can dominate individual result pages.

Read the evidence ↗
Near-identical transcript pairs
12.5%

Of the 2,000 result lists contain a pair of non-identical transcripts scoring at least 80% fuzzy similarity. Matched random lists reach that threshold in 0.2% of queries.

Retrieved result lists
12.5%
Matched random lists
0.2%
Percentages describe result lists in this sample. They do not estimate the fraction of the entire corpus with duplicate text.
A modest association

Close embeddings do not require the same script.

Pearson correlation between speaker cosine and fuzzy text similarity is 0.17 for query-to-result pairs, and 0.22 between returned results. Many close speaker embeddings still have different transcripts.

Generated voices, accent, emotion, setting and repeated instructions can co-vary. The scatter is a diagnostic of that relationship, not a test of verified speaker identity.

Hear close vectors with different words ↗
Three ways into the evidence

Inspect the point. Hear the pair.

01 / EXPLORE

The similarity map

Inspect every measured pair. Filter comparison type, language, dataset, rank and text overlap. Dotted means and percentiles follow the current view.

Open the explorer ↗02 / LISTEN

The listening room

Eight deliberate comparison cases with full-clip previews. Read highlighted word differences, compare the voices, and trace source identities.

Listen to the examples ↗03 / QUALITY

Transcript agreement

See the full-corpus normalized WER distribution, its exact-match share, and the distinction between corpus and per-clip averages.

See the quality results ↗