Dataset Viewer
Auto-converted to Parquet Duplicate
eval_id
string
title
string
task
string
date
timestamp[s]
purpose
string
scope
string
sentence_set
string
sentence_set_revision
string
results_dataset
string
listening_space
string
leaderboard
string
data_json
string
repo
string
annotation_axes
list
models
list
n_models
int64
n_runs
int64
n_clips
int64
repo_commit
string
hebrew-tts-bakeoff-1126
Hebrew TTS Bake-off 11/26
text-to-speech
2026-11-18T00:00:00
Pick the reference voice the Hebrish recorder plays before a learner reads a Hebrew sentence aloud from nikud; measure each model's nikud inference (plain text vs checked nikud).
Open models (weights on Hugging Face; Fish S2 Pro through its publisher's API), plus four commercial APIs added the same day (Gemini 3.1 Flash TTS, OpenAI gpt-4o-mini-tts, ElevenLabs v3 via fal, Google Cloud Chirp 3 HD)
Hebrish/Hebrew-TTS-Eval-Set
667f8dac2eaab969cb0c275ffe976d1aab9896fc
Hebrish/Hebrew-TTS-Bakeoff-1126
Hebrish/Hebrew-TTS-Bakeoff-1126
https://huggingface.co/spaces/danielrosehill/Hebrew-TTS-Leaderboard
https://huggingface.co/spaces/Hebrish/Hebrew-TTS-Bakeoff-1126/resolve/main/data.json
https://github.com/danielrosehill/Hebrew-TTS-Bakeoff-1126
[ "naturalness (1-5)", "clarity (1-5)", "hebrew (1-5)", "nikud (yes/no)", "use (yes/maybe/no)", "word_issues (stress, mispronunciation, minor)", "note" ]
[ "bluev3", "chatterbox-multilingual", "elevenlabs-v3", "elevenlabs-v4", "f5-hebrew-v2", "fish-s2-pro", "gemini-2.5-flash-tts", "gemini-2.5-native-audio", "gemini-2.5-pro-tts", "gemini-3.1-flash-live", "gemini-3.8-flash-lite-tts", "gemini-3.8-flash-tts", "gemini-3.8-live", "gemini-flash-tts"...
26
61
536
b2e00d62b08067b306fe2cae82a8e7e5dd1110d9

Hebrew TTS Evaluations

The register of Hebrew text-to-speech evaluations run for Hebrish: when each ran, which models it included, the sentence set and the raw run parameters. One source of truth, added to as evaluations are run.

  • evals: one row per evaluation. Date, purpose, scope, sentence set and its revision, models, counts, and links to the results dataset, the listening Space, the clip index (data_json) and the code at the commit that made it.
  • runs: one row per model run inside an evaluation. The model and the revision it ran, the input (plain Hebrew or nikud), voice or reference clip, hardware, sample rate, load and per-sentence time, licence, publisher, release date, training origin, the script, and notes from the run.

Rows are written by scripts/register_eval.py in each evaluation's repo; re-registering an evaluation replaces its rows. Listening notes are kept privately and are not part of this dataset.

Evaluation Date Models Runs
Hebrew TTS Bake-off 11/26 2026-11-18 10 17

Index of all Hebrew evaluations: Hebrew-AI-Evaluations.

Downloads last month
206