Dataset Viewer

The dataset viewer should be available soon. Please retry later.

Hebrew TTS Listening Annotations

Daniel Rosehill's own listening annotations of Hebrew text-to-speech evaluations (part of Hebrish), made with the TTS Pronunciation Annotation Starter. Read them with their voice notes in the reader Space. Evaluations are listed in Hebrish/Hebrew-TTS-Evaluations.

Method: screen, then rank. Each clip is first passed or failed as a whole (a fail carries a reason category; word-level detail such as the syllable stressed vs the one that should be, or IPA heard vs expected, is optional). Clips that pass then meet head to head, blind, ranked with Bradley-Terry. From 2026-11-18 the Arena is split by voice gender: a voice only meets voices of its own gender (a first listen across genders mostly judges the voice, not the model); earlier votes across genders stay in pairs.jsonl but do not count.

Path What
annotations/<eval_id>/screen.jsonl Pass/fail per clip (tts-annotation/2); the latest line per (run, sentence) wins
annotations/<eval_id>/pairs.jsonl Pairwise comparisons among passing clips (tts-pairwise/1)
annotations/<eval_id>/voice.jsonl + voice/<id>.webm Spoken annotations as recorded (written by the app)
annotations/<eval_id>/voice-applied.jsonl For each spoken note: the verbatim transcript, a lightly tidied one, and the structured reading a model applied
annotations/<eval_id>/screen-voice.jsonl Verdict rows derived from spoken notes (source: "voice:<id>"), merged with screen.jsonl on read
annotations/<eval_id>/decisions.jsonl Every method and scope decision, vote set aside and Arena exit, dated, with its source (tts-decision/1: id, date, kind, decision, runs, source, decided_by). Append-only
annotations/<eval_id>.jsonl One early annotation in the first (1-5 scale) format
Downloads last month
-

Collections including danielrosehill/Hebrew-TTS-Listening-Notes