ABE101's picture
Add 6 genre-diverse Yiddish clips (~53 min) with Gemini reference transcripts
9340ba3 verified
|
Raw
History Blame Contribute Delete
2.68 kB
metadata
license: other
language:
  - yi
task_categories:
  - automatic-speech-recognition
tags:
  - yiddish
  - asr
  - benchmark
  - multi-genre
pretty_name: Yiddish ASR Benchmark (Multi-Genre)

Yiddish ASR Benchmark (Multi-Genre)

A small, genre-diverse Yiddish ASR benchmark seed set: 6 clips (~53 minutes total), one per genre, each with a Gemini-generated verbatim reference transcript. Intended to complement the existing single-speaker Yiddish24 collections (Kohn-AI/yiddish24-audio, Kohn-AI/yiddish24-wav, ABE101/yiddish24-dual-asr) with broader genre coverage for benchmarking rather than training.

All source audio is from yiddish24.com.

field meaning
file_name audio (mp3) under data/
genre one of torah_shiur, news, interview, monologue_podcast, general_podcast, vlog
title episode/segment title
speaker presenter/host name, where applicable
category_path yiddish24.com category the source came from
track_id yiddish24 post id(s); news clip concatenates 3 bulletins, ids ;-separated
source_url original cloudfront/cdn url(s); ;-separated for the concatenated news clip
clip_offset_sec start offset (seconds) into the original recording where this clip begins
duration_sec clip duration
text_gemini reference transcript (Yiddish, Hebrew script), verbatim, unclear spans marked [?]
transcription_model model used to produce text_gemini

Genres

genre title source
torah_shiur פרשת דברים הרה"ג ר' גמליאל ראבינאוויטש שליט"א
news 3 concatenated news bulletins Kol Mevaser news desk
interview הרב דוד פינטער - עמעזאן דיסטריבורטער אלגעמיינע אינטערוויוס
monologue_podcast קול מדע משה נחום קרויס
general_podcast טשיקאווע ברעקלעך יושע ווייס
vlog עיטש וואק קאמפאני חיים פערלאוויטש שאו

6 clips · ~53 min.

Reference transcript method

text_gemini was produced by google/gemini-3.1-pro-preview, called via the Vercel AI Gateway chat completions endpoint with the audio file attached as a file content part (media_type: audio/mpeg) and a verbatim-transcription prompt. These are a first-pass gold standard — worth a human spot-check before treating as strict ground truth, especially for proper nouns and numbers. Unclear audio is marked [?] in the transcript.

from datasets import load_dataset
ds = load_dataset("Kohn-AI/yiddish-asr-benchmark")