ru_extra / README.md
WTForbes's picture
tonewebinars: 44,246 rows / 311.77 h (ToneWebinars admitted by GigaAM-v3 + Qwen3.8 root adjudication) + README configs tonebooks_adj/tonewebinars
a952ec3 verified
|
Raw History Blame Contribute Delete
11.1 kB
metadata
license: apache-2.0
language:
  - ru
task_categories:
  - automatic-speech-recognition
pretty_name: WTFO ru_extra
size_categories:
  - 10K<n<100K
extra_gated_prompt: >-
  This dataset contains audio redistributed from Vikhrmodels/ToneBooks
  (Apache-2.0, audiobook fragments) and Vikhrmodels/ToneWebinars (Apache-2.0;
  upstream ZeroAgency/shkolkovo-bobr.video-webinars-audio, MIT, school exam
  webinars). Agree to use it for research / ASR training only and to respect the
  rights of the original audiobook publishers and narrators.
configs:
  - config_name: tonebooks
    data_files:
      - split: train
        path: tonebooks/data/train-*.parquet
  - config_name: tonebooks_adj
    data_files:
      - split: train
        path: tonebooks_adj/data/train-*.parquet
  - config_name: tonewebinars
    data_files:
      - split: train
        path: tonewebinars/data/train-*.parquet

WTFO/ru_extra

俄語 ASR 補充資料:俄語音訊 → 俄文轉寫。用來補充基線 YuKuanFu/ru(約 4,465.6 h)。 Russian ASR supplement (Russian audio → Russian text) for the YuKuanFu/ru baseline (~4,465.6 h).

config 來源 / source 列 rows 時數 hours 狀態 status
tonebooks Vikhrmodels/ToneBooks @ 4fd7ba96 (train) 60,695 105.24 已清理 / cleaned
tonebooks_adj Vikhrmodels/ToneBooks @ 4fd7ba96 (train),tonebooks 以外的新收列 / rows not in tonebooks 13,996 35.41 已清理(三 ASR + LLM 裁決)/ cleaned (3-ASR + LLM adjudication)
tonewebinars Vikhrmodels/ToneWebinars @ a9bda19f (train) 44,246 311.77 已清理(三 ASR + LLM 裁決)/ cleaned (3-ASR + LLM adjudication)

tonebooks 與 tonebooks_adj 不重疊(id 不重複),可同時載入。 / tonebooks and tonebooks_adj are disjoint.

來源與授權 / Provenance and license

ToneBooks

  • Hugging Face 卡片宣告的授權是 Apache-2.0。
  • 內容是俄語有聲書片段,共 10 位朗讀者(speaker_id)。卡片致謝 its5Q/bigger-ru-book 協助收集。
  • 上游 its5Q/bigger-ru-book(rev 18a38573)沒有 dataset card,也沒有授權標記。
  • 書籍與錄音的著作權屬原出版者和朗讀者。Apache-2.0 只是 Vikhrmodels 的宣告,不是我們的法律評估。

ToneBooks (English)

  • The Hugging Face card declares Apache-2.0.
  • Content: Russian audiobook fragments read by 10 narrators (speaker_id). The card credits its5Q/bigger-ru-book for collection help.
  • The upstream its5Q/bigger-ru-book (rev 18a38573) has no dataset card and no license tag.
  • Copyright of the books and recordings stays with the original publishers and narrators. Apache-2.0 is Vikhrmodels' declaration, not our legal assessment.

轉寫的獨立性 / Transcript independence

  • ToneBooks 的文字有書籍排版特徵。
    • label 中出現對話破折號 - 的有 25.1%,括號 2.5%,分號 1.5%,?! 0.5%。
    • faster-whisper 在同一批音訊上,這幾項全是 0%。
    • 因此判定 label 是書本原文,與 Whisper 互相獨立。
  • ToneBooks text shows book typography that faster-whisper never produced on the same audio.
    • Dialogue hyphen 25.1%, parentheses 2.5%, semicolons 1.5%, ?! 0.5% in labels; 0% for each in Whisper output.
    • We therefore treat the labels as source book text, independent of Whisper.
  • ToneWebinars 的上游卡片寫明轉寫是「whisper and VAD」產生的,label 與 Whisper 不算獨立,因此加做 GigaAM-v3 重驗與 LLM 裁決(見下方 tonewebinars 一節)。
  • ToneWebinars' upstream card states its transcripts were made "with whisper and VAD", so its labels are not independent of Whisper; GigaAM-v3 verification and LLM adjudication were added (see the tonewebinars section).

ToneWebinars

  • Hugging Face 卡片宣告 Apache-2.0;上游 ZeroAgency/shkolkovo-bobr.video-webinars-audio 宣告 MIT。內容是 Shkolkovo 免費升學考試課程的網路研討會錄音(約 2,573 場)。
  • The card declares Apache-2.0; upstream ZeroAgency/shkolkovo-bobr.video-webinars-audio declares MIT. Content: recordings of free exam-prep webinars by Shkolkovo (~2,573 webinars).
  • 錄音著作權屬原講者與 Shkolkovo;上述授權是上游的宣告,不是我們的法律評估。

清理內容 / What was cleaned (tonebooks)

  • 範圍 / scope:上游 train 共 91,976 列。

    • 先做精確音訊去重:比對 PCM SHA 與原始 FLAC SHA,對象是 baseline train/eval 與來源內部。
    • 再做 faster-whisper-large-v3 盲轉寫 QC:不給參考文字、hotword 或語言提示。
    • 通過以上兩步的 83,577 列 / 164.04 h 是候選。
  • 收錄條件 / admitted:只收 source label 與 Whisper 在正規化後一致的列。

    • 正規化是指不分大小寫、去標點、ё→е,連字號與空白視為相同。
    • 收錄共 60,695 列 / 105.24 h。
    • 書本原文加上一個獨立 ASR 家族,構成兩個獨立來源。
    • Only rows whose source text equals the independent faster-whisper-large-v3 hypothesis after normalization (case, punctuation, ё/е, hyphen/space) are included.
  • 排除 / excluded(見 EXCLUDED.jsonl):

    原因 / reason 列 rows 時數 h
    仍有用字差異、待第二 ASR / lexical difference, pending GigaAM-v3 22,190 58.54
    HOLD:外部評測命中、非俄文、重複等 / eval hit, non-Russian, duplicates 481 0.18
    保守排除:外部評測短句文字命中 / conservative short-text eval hit 197 0.07
    加標點後仍無句讀(只有破折號) / still no sentence punctuation 14 0.01

    上游另有兩類列,不在 83,577 列候選中:

    • 已知外部評測重疊:11 列。
    • 音訊/ASR 待審:8,377 列 / 15.1 h。
  • 外部評測隔離 / eval isolation:

    • 比對方式:精確文字、近似文字(3-gram 加相似度 ≥90 或 partial ≥95)、精確 PCM。
    • 比對集合:Golos crowd/farfield、Common Voice 22 ru test、FLEURS ru test、RuLS test/val、SOVA RuDevices 與 RuAudiobooksDevices test、ToneBooks/ToneWebinars validation、YuKuanFu/ru eval、ToneSpeak validation(部分)。
  • 基線去重 / baseline dedup:

    • 對象:YuKuanFu/ru@6cf44587 train(4,650,979 列)與 eval。
    • 方式:精確音訊,加上正規化文字(長度 ≥20 字元)。
  • 標點 / punctuation:

    • 每個 label 都有標點。
    • 196 列原本完全沒有標點。這些列以 1-800-BAD-CODE/xlm-roberta_punctuation_fullstop_truecase 為參考,只補 .,?!。
    • 文字鎖定、大小寫不改。驗證條件:去掉新增標點後,必須與原文逐字相同。
    • 這些列的 text_state=punct_only_overlay,qc_flags 含 punct_added_text_locked。
  • 數字 / numbers:保留原書寫法,不做 ITN 改寫。含數字的列在 qc_flags 標 label_has_digits。

  • 只收俄文 / Russian only:

    • 範圍內沒有英文翻譯。
    • 拉丁字母(品牌、符號)保留,標 label_latin。
    • 拉丁字母多於西里爾字母的列已排除。

清理內容 / What was cleaned (tonebooks_adj, tonewebinars)

  • 證據 / evidence:每列有三份盲轉寫:faster-whisper-large-v3(W)、GigaAM-v3 RNNT(G)、GigaAM-v3 CTC(C)。G 與 C 同一家族,只算一個來源。

  • LLM 裁決 / LLM adjudication:本機 Qwen3.8-27B-FP8(只看文字,不聽音訊)對 label 與 ASR 不一致的列提出 KEEP / FIX / HOLD / DROP,再由 root 規則 ru_clean912_root_adj_v1_20261025 決定:

    • KEEP:差異只在格式(標點、大小寫、ё/е、數字寫法)、填充詞或已知 ASR 誤聽;或 label 與兩個獨立來源一致(AGREE)。
    • FIX:只接受有 ASR 支持的最小修改;人名、數字、拉丁字串的修改需要額外證據,否則 HOLD。
    • HOLD/DROP 的列不收錄(見各 config 的 EXCLUDED.jsonl)。
  • 收錄 / admitted:

    config 候選 rows / h KEEP(LLM) KEEP(AGREE) FIX 收錄 rows / h HOLD 未收 rows / h DROP rows / h
    tonebooks_adj 83,577 / 164.04(含 tonebooks 已收) 7,364 5,675 957 13,996 / 35.41 8,207 / 23.14 678 / 0.25
    tonewebinars 105,194 / 747.72 38,438 3,405 2,403 44,246 / 311.77 59,929 / 428.82 1,019 / 7.14
  • text_state:root_adj_accept_keep_LLM、root_adj_accept_keep_AGREE、root_adj_accept_fix_LLM;text_original 是上游原文。

  • qc_tier:stage-2 分級(T0/T1 正規化後與 Whisper 一致;T2 有用字差異,經上述裁決)。

  • 評測隔離、基線去重與標點規則同 tonebooks;每個 label 都有標點。

  • 仍在處理 / still in progress:HOLD 列(約 452 h)正以第二輪 LLM 裁決(加上下文與外部查證)處理,通過者日後以新 config 發佈。

尚未完成 / Still pending (honest limits)

  • ToneSpeak validation:只取得 64/700 列 metadata。它是 GPT 合成文本,重疊風險低,但尚未完整比對。
  • GigaAM evaluation 中的 OpenSTT 驗證集:未取得,沒有比對。
  • 沒有做重新編碼或重切段的近似音訊比對,也沒有做 speaker 層級的評測隔離。
  • 專名:ToneBooks 的專名拼寫以原書為準。Whisper 與 label 有差異的列都已排除,等第二 ASR 重驗。
  • text 保留原書的排版,包括對話破折號 - 、引號和省略號。

English summary of the same limits:

  • ToneSpeak validation: only 64 of 700 rows of metadata were obtained.
  • The OpenSTT eval sets inside the GigaAM evaluation were not obtained.
  • No re-encode / re-segment near-audio check and no speaker-level eval isolation were done.

欄位 / Columns

欄位 說明
id 列 ID
audio {bytes, path};FLAC,16 kHz mono(Audio(16000))
text 清理後的俄文 label(有標點)
text_original 上游原文
duration 秒
source_dataset、source_revision、source_file、source_row、source_id 上游 parquet 與列號、上游 mp3 名稱
speaker_id、license 朗讀者、授權
asr_hyp、asr_model、asr_model_revision、asr_wer、asr_cer faster-whisper-large-v3 盲轉寫與距離
qc_tier T0 完全一致;T1 只差空白或連字號
text_state source 或 punct_only_overlay
eval_gate、dedup_gate、qc_flags 評測與去重檢查結果、旗標
audio_sha256、pcm_sha256 音訊雜湊
clean_rule 清理規則版本

使用 / How to load

from datasets import load_dataset
ds = load_dataset("WTFO/ru_extra", "tonebooks", split="train")   # or "tonebooks_adj", "tonewebinars"   # gated: accept terms first, then `huggingface-cli login`
print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])

清理證據(本地,未發佈):tp1_new_data/ru/clean912_20261022/,含 RU_CLEAN_PLAN.md、out/SUMMARY.json、out/stage2_final.sqlite、out/PUNCT_RECEIPT.json。 Cleaning evidence (local, not published): the same directory and files.