TripleU613/yiddish-asr

Additional modern American sources: KJ Weekly public issues and Latest Talks, Pinches Glauber and selected Vaybertaytsh podcasts. See source selection and quality notes. Original PDFs and raw audio are retained; these require review and are not gold training pairs.

For modern American Hasidic Yiddish, start with modern_us_language, modern_us_kaveshtiebel and modern_us_speech_candidates. See selection and review guidance. Text is source-focused and exact-deduplicated; speech is raw audio requiring transcript and speaker review. Preserve evaluation splits and source terms.

Yiddish corpus

One collection for Yiddish language and speech-recognition data. Speech-generation candidates are included where available. No training was launched during preparation.

Configuration	Contents
language	Deduplicated text with provisional thread/document splits
recognition_omnilingual	Original Omnilingual Yiddish speech and upstream splits
recognition_aligned_candidates	Existing news alignments, not independently verified
recordings_<source>	Original untranscribed recordings, sorted by source

from datasets import load_dataset

text = load_dataset("TripleU613/yiddish-asr", "language", streaming=True)
speech = load_dataset("TripleU613/yiddish-asr", "recognition_omnilingual", streaming=True)

Most published configurations use Parquet; external Kashes benchmark subsets retain upstream JSONL. Audio is embedded, with no dependency on server paths. Recordings retain their original encoding and sample rates. SHA-256 digests identify audio; exact duplicates are excluded from the recording queue.

Text is normalized and exact-deduplicated; forum author/timestamp fields are omitted and common contact patterns are redacted. Language identification is heuristic. Known news alignment errors remain flagged. Recording containers and tails are checked; language, transcript and speaker audits remain pending. Unknown speakers stay unknown. These provisional partitions do not establish speaker independence.

Recording publication: complete. Counts and preparation details are in metadata/. See metadata/recordings-upload-status.json for upload progress.

Contributions

See CONTRIBUTING.md. New community batches are submitted under incoming/ for review, then converted to published Parquet. Incoming submissions are not automatically included in the published configurations.

Source terms

Original source terms apply; there is no blanket CC0 license for this mixed collection. Omnilingual data is by Meta/Omnilingual ASR Team under CC BY 4.0; its original card and attribution are in docs/SOURCE_OMNILINGUAL.md. Other source identifiers are retained in records, including the named forum/text collections, Yiddish24, Ashreinu, Yiddish Book Center and Podscribe collections. Community submissions under the contribution guide retain their CC0 declaration.

Additional corpus material

Additional text, pronunciation data, legacy annotations and audio are organized under data/. Candidate tables use review splits. Retained source labels and speaker clusters are not independent verification. Different encodings, vocal stems and segments are explicitly labeled as views. A row with payload_status=deduplicated_reference links by audio_sha256 to an existing payload, rather than duplicating audio bytes. Original video tracks are omitted where an audio stream can be copied losslessly.

Additional audio publication: complete. Progress: metadata/expanded-upload-status.json. Preparation coverage and exclusions: metadata/remaining-preparation.json.

Additional language text, Hebrew-script auxiliary book text, pronunciation/pointing records, and legacy speech labels are available in separate review configurations. Book provenance is in book_sources. These additions retain uncertain source labels; they are not newly audited gold or an independent evaluation set. Source terms continue to apply.

Original source-page scans and booklet PDFs are retained separately for OCR, provenance and review; they are not labeled as ready language-model text.

Supplemental text variants and incremental recognition recordings are available in text_variants_supplement and recognition_supplement (review only). Pointed, pseudo-label and previously rejected text retain their source status. Do not treat inherited splits or textual variants as independent evaluation data. Source terms remain unchanged.

Supplemental source lyric candidates and residual research audio are in lyrics_supplement and research_audio_supplement (review only). Candidate texts and audio slices remain unverified. Do not treat inherited splits or textual variants as independent evaluation data. Source terms remain unchanged.

Residual cached text and timing labels are in cached_annotations_supplement (review only). These are unverified machine or candidate labels. Source IDs are retained, but audio matching and alignment correctness require review. Source terms continue to apply.

The 2026-11-07 residual-value sweep is complete for the audited source snapshot. Its scope, added records and exclusions are recorded in metadata/value-sweep-summary.json. Recovered HTML-cache lyric text is retained separately for review; it is not presumed to be Yiddish or verified training material.

External collections: external_reyd, external_jochre, external_kashes_kashes_mt, external_kashes_flores_plus, external_kashes_ud_pos, external_kashes_ud_dep, external_kashes_ud_translit, external_kashes_ud_lemma, external_kashes_ner_ehri, external_kashes_ner_wikiann, external_kashes_ner_newnlp, external_kashes_piqa, external_kashes_paws_wiki, external_kashes_wikiqa. Source documentation and licenses are retained under docs/external-v1/. REYD retains three text orthographies and source reader IDs; ivrit.ai retains source captions and machine alignments. Jochre annotations remain CC BY-NC-SA 4.0 and preserve source splits. Kashes is evaluation-only and retains its upstream terms; it must not be mixed into training. These sources do not change one another's licenses. No training has been launched.

External collections: external_ivrit. Source documentation and licenses are retained under docs/external-v1/. REYD retains three text orthographies and source reader IDs; ivrit.ai retains source captions and machine alignments. Jochre annotations remain CC BY-NC-SA 4.0 and preserve source splits. Kashes is evaluation-only and retains its upstream terms; it must not be mixed into training. These sources do not change one another's licenses. No training has been launched.

Additional Yiddish web candidates: FineWeb2, HPLT 3.0, MADLAD-400 clean. Provenance, source terms, duplicate references, and limitations: metadata/external-web-v2/summary.json. FineWeb source test remains evaluation-only. Web candidates require language, duplication, and quality review before training. No training has been launched.

Additional source snapshots: Yiddish Wikipedia, Wikisource, Tatoeba sentences and YidTakNL historical editions. Source schemas, attribution, original assets and terms are retained under docs/external-public-v3/ and metadata/external-public-v3/summary.json. These are review candidates requiring cross-corpus deduplication. No training has been launched.

Yiddish WhatsApp readings: external_ivrit_whatsapp retains original audio, alternate Opus, source text, unverified machine alignment, and quality flags. Source provenance and license: metadata/external-whatsapp-v3/summary.json and docs/external-whatsapp-v3/. No training launched.

Additional Yiddish resources: external_kaikki_yi_v4, external_unimorph_yid_v4, external_ppchy_v4, external_yitb_v4. Provenance, source terms and measured counts: metadata/linguistics-v4/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.

Additional Yiddish resources: modern_us_language, modern_us_kaveshtiebel, modern_us_speech_candidates. Provenance, source terms and measured counts: metadata/modern-american-v1/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.

Additional Yiddish resources: modern_us_kjweekly, modern_us_podcast_latesttalks, modern_us_podcast_pinches, modern_us_podcast_vaybertaytsh_us_selected. Provenance, source terms and measured counts: metadata/modern-us-followup-v2/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.

Additional Yiddish resources: modern_us_talkbank_yiddish_annotations. Provenance, source terms and measured counts: metadata/talkbank-yiddish-v1/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.

Additional Yiddish resources: modern_us_press_cover_samples, modern_us_der_yid_daily_bulletins. Provenance, source terms and measured counts: metadata/us-press-samples-v1/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.

Additional Yiddish resources: modern_us_der_yid_daily_pdf_inventory, modern_us_der_yid_daily_pdf_text. Provenance, source terms and measured counts: metadata/us-press-daily-pdf-v1/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.

Additional Yiddish resources: modern_us_open_press_inventory_v2, modern_us_open_press_text_v2. Provenance, source terms and measured counts: metadata/us-press-open-v2/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.

Additional Yiddish resources: modern_us_open_press_text_v3, modern_us_open_press_inventory_v3, modern_us_open_press_scans_v3. Provenance, source terms and measured counts: metadata/us-press-open-v3/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.

Additional Yiddish resources: modern_us_nafshi_article_paragraphs, modern_us_nafshi_article_inventory. Provenance, source terms and measured counts: metadata/nafshi-articles-v1/summary.json. Original annotations are retained; evaluation overlaps remain separate. No training launched.