Sources & snapshots

Get the text-to-speech benchmark data

Every ranking this site has ever shown is an immutable snapshot: the raw board scores, the alias mappings that were applied, and the rules the composite was computed with. Download any of them as JSON — re-running the published formula over a snapshot reproduces the exact board it rendered.

The boards

BoardSignalStatus

artificialanalysis.comSpeech Arena Elolive

voicearena.comarena Elo, US Englishlive

humannessindex.vapi.aihumanness score 0–100live

Latency sources

SourceMetricRead

Coval · primaryTime to first audio, p50, us-east-1 · 20 modelsOct 11, 2026

SpekoTime to first audio, p50, US East · 22 modelsOct 11, 2026

Humanness IndexTime to first audio, median · 19 modelsOct 11, 2026

Snapshot archive

SnapshotModelsDownload

2026-10-11T21-20-12Z · current110 models · captured Oct 11, 2026JSON

2026-10-11T21-08-50Z110 models · captured Oct 11, 2026JSON

2026-10-06T10-45-43Z108 models · captured Oct 6, 2026JSON

2026-10-04T05-26-43Z107 models · captured Oct 4, 2026JSON

2026-09-27T04-07-09Z104 models · captured Sep 27, 2026JSON

2026-09-25T18-15-02Z104 models · captured Sep 25, 2026JSON

2026-09-24T14-46-08Z105 models · captured Sep 24, 2026JSON

2026-09-06T04-00-13Z107 models · captured Sep 6, 2026JSON

2026-08-30T04-01-20Z110 models · captured Aug 30, 2026JSON

2026-08-26T07-11-50Z110 models · captured Aug 26, 2026JSON

2026-08-19T14-05-26Z108 models · captured Aug 19, 2026JSON

2026-08-19T04-01-20Z103 models · captured Aug 19, 2026JSON

2026-08-17T04-00-44Z103 models · captured Aug 17, 2026JSON

2026-08-15T18-49-58Z103 models · captured Aug 15, 2026JSON

2026-08-14T10-43-32Z99 models · captured Aug 14, 2026JSON

Snapshots are append-only. Each file contains boards (source list, status, normalization strategy, board size), models (raw scores per board, aliases applied, metadata, and each latency source's own figure), latency (which latency sources were read and when — display only, never part of the score), and rules (the scoring mode, board authorities, and the dataset constants the score was computed with). Ranks and the BPRA score are recomputed from these values — nothing pre-baked, nothing editable after the fact.