Charts, reports and data

Take the evidence
with you.

Self-contained reports for offline review, shareable figures, and full-precision measured data. Downloads preserve the scope of each analysis.

Offline reports

ZIP / CURRENT 2,000-QUERY SAMPLE

The full scatter report

All 838,411 points, source transcripts, filters, dotted references, and 16 selected audio previews. Unzip and open the single HTML file.

Download ↓15.1 MB
HTML / CURRENT SAMPLE

Single-file explorer

The same complete interactive report in one standalone HTML file. The whole dataset is embedded for offline use.

Download ↓20.2 MB
ZIP / EARLIER 200-QUERY SAMPLE

The audio notebook

The earlier listening notebook: 50 curated comparisons and 35 audio clips. A different, smaller diagnostic sample.

Download ↓6.1 MB

Publication figures

PDF / ALL MEASURED POINTS

Speaker × transcript scatter

Four comparison panels with mean, P10, median and P90 reference lines, plus short findings.

Download ↓0.4 MB
PDF / DENSITY VIEW

A view through overplotting

Full-data binned density makes the crowded regions easier to read; all observations contribute.

Download ↓0.1 MB
PDF / FULL BACKFILL

Normalized WER distribution

Full-corpus transcript agreement: exact matches, percentiles and the long error tail.

Download ↓0.0 MB

Full-precision data

GZIP CSV / PART 1

Measured pairs, first half

419,206 rows with cosine, fuzzy and edit scores, source-vector bindings, ranks and exact/audio flags.

Download ↓14.2 MB
GZIP CSV / PART 2

Measured pairs, second half

419,205 rows with the same header. Concatenate the data rows to recover all 838,411 observations.

Download ↓14.2 MB
GZIP CSV / LOOKUP

Clip and speaker identities

Join member_a/member_b to member_index. Includes source IDs, source keys, full ASR text, dataset, language and speaker-slot labels.

Download ↓8.6 MB
The two pair exports are split for convenient hosting. Each has a header. Their member indices resolve against the same clip lookup; source_vector_row fields identify the original frozen speaker index.

Machine-readable summaries

JSON

Diagnostic summary

All comparison counts, correlations, uncertainty intervals, thresholds and sampling configuration.

Download ↓0.0 MB
JSON

Reference statistics

Full-precision means, P10, median and P90 for the original four comparison groups and pooled sample.

Download ↓0.0 MB
JSON

WER summary

Full-backfill raw/normalized WER, normalized CER and the exact normalization policy.

Download ↓0.0 MB