Hugging Face's logo

Datasets:
Fritter-ASR
/
Training-Sources

Modalities:
Tabular
Text
Formats:
json
Size:
< 1K
Tags:
index
speech
asr
tts
fritter
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Training-Sources / README.md
danielrosehill's picture
danielrosehill
Update index
d8b014d verified
preview code
|
Raw History Blame Contribute Delete
6.13 kB
metadata 
license: cc-by-4.0
pretty_name: Fritter Training Sources
tags:
  - index
  - speech
  - asr
  - tts
  - fritter
configs:
  - config_name: sources
    data_files: sources.jsonl
    default: true
  - config_name: releases
    data_files: releases.jsonl
Fritter icon

Fritter Training Sources

An index of the datasets the Fritter speech models are trained on. It holds no audio; each row links to the dataset itself.

  • sources: every dataset involved and how Fritter uses it (fritter_asr: used, partial, not used, indirect). All training audio comes from one primary dataset, danielrosehill/Read-Aloud-Voice-Sentences; the topic datasets listed with it are subsets of it, there for browsing by topic.
  • releases: each published model version, its training run and the exact revision of the primary dataset it was trained on.

How Fritter (ASR) is trained

23 of 29 topic subsets are used. The training mix is the speaker's English takes that are not code-switching, with these tags left out: israeli-places, israeli-streets, israeli-tv, israeli-names, tanach, biblical-verses, talmud, rabbinic-works, rabbis, halacha, name-cipher, paragraphs, weird-voices. About 10% of each mix is LibriSpeech replay, so general English does not degrade. From v1.2 each release continues from the previous one on the new takes plus a sample of earlier ones.

Dataset Fritter ASR Takes Hours
danielrosehill/Read-Aloud-Voice-Sentences used 28,570 43.75
openslr/librispeech_asr used
Hebrish/Acronym-Sentences used 99 0.12
Hebrish/Background-Speech used 12 0.02
Hebrish/Brand-Name-Sentences used 378 0.70
Hebrish/Currency-Sentences used 160 0.24
Hebrish/Disfluent-Read-Sentences used 55 0.22
Hebrish/Expanded-Acronym-Sentences used 50 0.07
Hebrish/Hard-ASR-Sentences used 68 0.07
Hebrish/Irish-English-Sentences used 258 0.38
Hebrish/Noisy-Condition-Sentences used 7,202 10.10
Hebrish/Pronunciation-Variants used 170 0.23
Hebrish/Repeated-Word-Sentences used 19 0.06
Hebrish/Self-Correction-Sentences used 27 0.04
Hebrish/Singing-Sentences used 16 0.02
Hebrish/Spoken-Aside-Sentences used 5 0.01
Hebrish/Whispered-Sentences used 312 0.49
danielrosehill/AI-Interactions-Sentences used 583 0.74
danielrosehill/General-English-Sentences used 1,910 2.89
danielrosehill/Homophone-Context-Sentences used 158 0.25
danielrosehill/Linux-Admin-Sentences used 530 0.86
danielrosehill/Personal-Vocabulary-Sentences used 191 0.25
danielrosehill/Smart-Home-Sentences used 185 0.29
danielrosehill/Tech-Sentences used 7,495 11.63
danielrosehill/Voice-Command-Sentences used 1,266 0.69
Hebrish/Israeli-Name-Sentences not used 117 0.17
Hebrish/Jewish-Religious-Terminology-Sentences not used 2,146 3.86
Hebrish/Judaica-Sentences not used 772 1.46
Hebrish/Weird-Voice-Sentences not used 6 0.01
Hebrish/Yiddish-Sentences partial 348 0.55
danielrosehill/English-Hebrew-Codeswitching-Sentences not used 8,233 14.05

Sizes measured 2026-11-08.

Releases

Model Version Run Primary dataset revision
Fritter (ASR) v1.0 r11-general 520e7506e379
Fritter (ASR) v1.1 r25-general 5ce184058554
Fritter (ASR) v1.2 r30-general 306a7450433d
Fritter (ASR) v1.3 r35-general 39f21d714103
Fritter TTS v1 fritter-tts-v1 not recorded

Generated by scripts/fritter_training_sources.py in the training repository. Regenerated after each release.