doanhlt/cabin-interpreted-interviews

Cabin interpreted interviews (pilot)

Real consecutive interpretation recorded on YouTube: interviews, press conferences and Q&A where a speaker (MC / journalist / foreign guest) talks in one language and an interpreter repeats it in another. Built for the Cabin Offline app (two-way on-device interpretation): ASR on interpreter-style speech, spoken language ID at turn changes, turn segmentation and speech-translation pairs. Companion of doanhlt/cabin-interpreter-interview-test (same video ids / start_sec / end_sec / lang / role names mc, guest, interpreter).

Licence: CC BY 3.0. Every video was published on YouTube under Creative Commons Attribution (reuse allowed) — checked on its own watch page with yt-dlp on the date in licence_evidence.jsonl; videos from YODAS2 (CC BY 3.0) were re-checked on YouTube and dropped if the uploader has since switched to another licence (kept if still CC BY or no longer online; see licence). Attribution for every video (channel, title, URL) is in the table below and attribution.jsonl. Text labels are machine transcripts by our own model (no Gemini / OpenAI / Soniox output anywhere).

Configs

<pair> configs have one row per speech segment; <pair>-pairs have one row per source turn + interpreter turn. Splits are video-disjoint (~20 % of videos in test, frozen in the build so later versions keep them).

config	split	videos	segments	hours	hours by language	pairs	pair hours
de-ru	train	1	43	0.07	de 0.05, ru 0.02	6	0.01
en-ar	train	1	193	0.27	ar 0.18, en 0.08	4	0.02
vi-en	test	1	36	0.06	en 0.03, vi 0.03	6	0.03
vi-ko	train	1	36	0.06	ko 0.04, vi 0.02	5	0.03

Raw audio hours of videos where two languages were detected, by pair: ar-de 0.15 h, ar-fr 1.02 h, de-fr 0.10 h, de-it 0.04 h, de-ko 0.05 h, de-ru 0.48 h, en-ar 0.66 h, en-de 1.22 h, en-es 0.14 h, en-fr 0.07 h, en-it 0.13 h, en-ja 3.98 h, en-ko 2.61 h, en-th 0.34 h, en-zh 0.38 h, es-it 0.21 h, es-zh 0.47 h, fr-it 0.09 h, fr-ja 0.33 h, it-ko 0.17 h, ja-ko 0.38 h, ja-th 0.04 h, ja-zh 0.40 h, ko-pt 0.03 h, ko-zh 0.22 h, vi-en 0.94 h, vi-es 0.05 h, vi-ko 0.50 h, vi-zh 1.02 h. Video outcome counts: {'monolingual': 247, 'bilingual_no_interpretation': 78, 'eval_overlap': 2, 'interpreted': 4}.

Segment columns

audio (16 kHz mono FLAC), video_id, url, channel, channel_id, title, upload_date, licence, source (youtube / yodas2), start, end (= start_sec, end_sec, seconds in the original video), duration, turn_id, speaker_cluster (per video), bilingual_speaker (cluster speaks both languages: usually the interpreter), lang, lid_p, text (R13x transcript), caption_text (YouTube caption over the same span, manual track first, else the original-language auto track), caption_kind, agree_wer (interior WER / CER of text vs caption), label_check (agree or teacher_only = no caption to compare), pair_id, role (mc / guest / interpreter, only inside a pair), pair_role (speaker / interpreter), pair_xsim, split.

Pair columns

src_audio, tgt_audio, pair_id, video_id, url, channel, licence, src_lang, tgt_lang, src_text, tgt_text, xsim (multilingual MiniLM cosine of the two transcripts), src_start/end, tgt_start/end, src_speaker_cluster, tgt_speaker_cluster, split.

Method

1. Search (search.py): about 35 interpreter / press-conference queries in vi, en, zh, ja, ko, fr, de, es, each as a plain search and with YouTube's Creative Commons filter, plus the seed videos' channels / events (MSI 2019, Park Hang-seo press conferences, ...). Every candidate's licence string is read from its watch page; only Creative Commons Attribution license (reuse allowed), public, not age-restricted, not live, 2-90 min is kept. Sign-language and children's content are excluded by title / tags.
2. YODAS2 mining (yodas_mine.py): caption text of YODAS2 vi/ko/ja/zh/en/fr/de/es shards scanned for language alternation and interpreter keywords; audio taken from YODAS2.
3. Segmentation (segment.py, CPU): Silero VAD -> segments <= 15 s; VoxLingua107 ECAPA LID (int8, renormalised over 13 interview languages); SpeechBrain ECAPA speaker embeddings + agglomerative clustering (cosine 0.6). Video pair = two dominant languages (2nd >= 8 % of confident speech, else monolingual and dropped); per-segment language re-decided inside the pair. Turns = same speaker + language runs; candidate pairs = adjacent turns in different languages by different speakers, gap <= 4 s, both >= 1.5 s, length ratio 1/3-3.
4. Transcription (GPU): R13x (our Qwen3-ASR-1.7B fine-tune) with the segment language as a hint.
5. Filters (build.py): caption agreement (WER / CER <= 0.30 when a caption exists, else teacher_only), wrong teacher language, repetition loops, eval overlap (video ids of all our eval / dev manifests and Vietnamese eval transcripts are excluded); a pair is kept only if all its segments pass and xsim >= 0.5 (the interpreter turn really translates the source turn); a video is kept with >= 2 such pairs.

Segment verdicts: {'unlabelled': 6025, 'teacher_only': 1995, 'teacher_lang': 296, 'agree': 249, 'empty': 17, 'cap_length': 50, 'disagree': 144}. Pairs checked 61, kept 22.

Example pairs

de-ru (6 pairs)

• wx38yRQYgPo_p005 xsim 0.74 [de] Ja, natürlich selbstverständlich. Genau, und dann bringen Sie mir das Dokument einfach wieder, wenn es ausgefüllt ist. -> [ru] И потом вы принесете просто документ, когда вы все заполнили.
• wx38yRQYgPo_p009 xsim 0.69 [ru] до скорой встречи, я вам желаю еще хорошего дня. -> [de] Dann beende ich das Gespräch jetzt. Vielen Dank.
• wx38yRQYgPo_p000 xsim 0.64 [de] Okay, super. Also Sie können sich verstehen, ja? -> [ru] Мы понимаем друг друга, да? Да, все понятно.

en-ar (4 pairs)

• W0w-u0Ua3xw_p022 xsim 0.58 [en] You stated earlier that while you were waiting for the cops, you saw someone that looked suspicious to you hanging around the place and looking at your ride. Is that what happened? -> [ar] تماماً، حالما أغلقت الهاتف بعد أن تحدثت إلى عامل المقسم، لاحظت بأن رجلاً يرتدي... خيرقا يتسكعون حول سيارتي وينظروا من خلال الشباك.
• W0w-u0Ua3xw_p007 xsim 0.50 [en] And you said that a deputy was arriving at this time. Did he immediately approach you? -> [ar] نعم هذا صحيح. خرج من سيارته. وسأل من اتصل.
• W0w-u0Ua3xw_p023 xsim 0.51 [en] And you said that a deputy was arriving at this time. Did he immediately approach you? -> [ar] نعم هذا صحيح. خرج من سيارته. و سأل من اتصل.

vi-en (6 pairs)

• vcgBfZ5ZN9Y_p006 xsim 0.60 [en] Well, we are just coming to beat everyone game by game. Every game is important, so every game matters, and. We are taking every game as serious as we can, even though. I think we can show a lot stronger performance. -> [vi] tất cả những trận đấu tại giải em bé sai này rất là quan trọng cho nên chúng tôi sẽ đánh bại tất cả các đối thủ trên con đường đi của chúng tôi
• vcgBfZ5ZN9Y_p001 xsim 0.74 [en] Well, I feel good that we are performing kind of well. I think our games can be better. And today I played pretty poorly, so. I want to play better. -> [vi] mặc dù chúng tôi rất là vui vì chúng tôi giành được chiến thắng nhưng mà ngày hôm nay tôi thi đấu không được tốt lắm cho nên tôi sẽ cố gắng hơn trong những chiến đấu về sau
• vcgBfZ5ZN9Y_p002 xsim 0.51 [en] Like just very stupid mistakes, so. Yeah, I think the game was there was not point in game that I felt like we were gonna lose the game. So yeah. -> [vi] thực ra thì chúng tôi không nghĩ là chúng tôi sẽ có nhiều khó khăn như vậy đến flashbook nhưng mà bản thân tôi đã mất khá nhiều sai lầm như là dịch chuyển có học lên hay là tóc biến trước mặt tiêm địch và có những giây p

vi-ko (5 pairs)

• C9yjoixv3-E_p005 xsim 0.58 [ko] 응원해 주셔서 정말 감사하고, 열심히 하겠습니다. -> [vi] rất cảm ơn mọi người, tôi sẽ xin cố gắng, khai thinh.
• C9yjoixv3-E_p002 xsim 0.55 [vi] và trong một vài trận đấu khó khăn vừa rồi mà các bạn đã trải qua thì bạn đã có những vị tướng có lẽ như là bạn rất là ít khi sử dụng như là Sona hay là Miss Fortune thì bạn có nghĩ đó thật sự là vấn đề hay không -> [ko] 어 최근에 몇 개의 경기에 진행하셨는데 SKT에 조금 힘들었던 경기들이 있었어요. 혹시나 저희 그 자주 쓰지 않는 그 체험 변 때문에 그렇습니까? 혹시 예를 들면 소나타인지 그리고 저희 그 미스트 포티튠 포티튠 인지 그것 때문에 자주 쓰지 않으신. 그 챔피언 때문에 조금 휘들었던 경기 있었는지 혹시 그렇게 좀 얘기할 수 있을까요?
• C9yjoixv3-E_p004 xsim 0.90 [vi] và như bạn cũng thấy đó là bạn có rất nhiều người hâm mộ ở Việt Nam vậy cuối cùng bạn có lời nào muốn gửi đến họ không? -> [ko] 도시다시비. 보시다시피 여기는 베트남 팬이 많이 있다고 보시죠. 혹시 우리 여기에 베트남 팬들한테 하고 싶은 말씀 있으세요?

Known issues

• Pilot size; speaker clusters are per video and not a real diarizer (overlap not modelled; MC and journalists may share a cluster). role is a heuristic from pair order + host language.
• text is a machine transcript (R13x); teacher_only rows had no caption to cross-check.
• Some videos are about interpreting rather than interpreted (lectures, practice tapes); xsim + min_pairs remove most, not all. Language-pair coverage follows what CC BY material exists, which is thin for vi-zh / vi-ja.
• Adults only by intent (children's titles excluded) — no age estimation was run.
• YouTube Terms of Service restrict automated download even for CC-licensed videos; we fetched audio at a low rate, without login or cookies. CC BY applies to the uploader's content; if you are the rights holder of a video and want it removed, open a discussion on this repo.
• Non-CC interpreted interviews found during search are listed only (no audio) in eval_candidates_noncc.jsonl.

Attribution (CC BY 3.0)

video	channel	title	source	config	split	licence checked
C9yjoixv3-E	Channel LOL	[MSI 2019] Phỏng vấn SKT Teddy: " cảm ơn các bạn Việt Nam"	youtube	vi-ko	train	2026-10-28
W0w-u0Ua3xw	Texan Translation	English-Arabic Consecutive Practice for the State Court Interpreter Oral Exam	youtube	en-ar	train	2026-10-28
vcgBfZ5ZN9Y	Channel LOL	[MSI 2019] Phỏng vấn Perkz: "chúng tôi sẽ đánh bại mọi đối thủ"	youtube	vi-en	test	2026-10-28
wx38yRQYgPo	ZwischenSprachen	Besonderheiten beim Telefondolmetschen	youtube	de-ru	train	2026-10-28