R06.04.02.03 · 任务小类

字幕与版本

Captions, subtitles & versions

Create timed captions and produce cut-downs, reframes and language versions.

20 条本项目编写的提案 · 分类 v1.1 · 新增基准运行 0 次。下方验证方式与部分得分规则是设计，不是执行结果。

典型应用: Premiere Pro · DaVinci Resolve · Aegisub
小类 GUI 需求标签: HY · 混合

数字员工关联

本项目的小类归属，不代表员工能力实证或已导入基准覆盖。验收与价值指标均为建议，未测量 ROI。

关联岗位

创意制作与修订 · 按改变的作品关联：矢量路径、照片调整、图层合成、剪辑、字幕、模型、修复对白及响度合规混音。

对照全部数字员工工作流设计

任务提案 · 20

展开查看完整原文与建议检查方式；两种界面均保留提案英文原文。

R06.04.02.03-T01 Create English captions to reading-speed rules in Premiere Pro

原始指令

Premiere Pro sequence 'Explainer_3min' has a single speaker. Create an English caption track (Speech to Text allowed) and correct it so every caption is ≤ 42 characters per line, ≤ 2 lines, ≤ 20 characters per second, and ≥ 5/6 s long. Export out/explainer_en.srt.

平台: mac · macOS win · Windows
应用: Premiere Pro
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: type · 文字与表单 av · 音视频时序 click · 指向与点击
工作产物: WP05 · 时序媒体
条件: licensed · 授权专业软件
提案难度: 3 / 5

建议验证方式

SRT parse: WER vs GT transcript ≤ 5%; 100% captions meet chars/line, lines, cps and min duration; caption spans overlap GT speech ≥ 95% and never start > 0.5 s before speech onset. Caption track must also exist in the .prproj.

建议部分得分

40% text accuracy; 35% formatting rules; 15% timing to speech; 10% track in project.

R06.04.02.03-T02 Burn in captions that avoid the lower thirds

原始指令

Premiere Pro sequence 'Panel_Talk' has lower-third graphics at 6 moments. Burn the existing caption track into a render: bottom-center, white sans 48 px with a 60% black box, and move captions up whenever a lower third is on screen so they never overlap it. Export out/panel_burned.mp4.

平台: mac · macOS win · Windows
应用: Premiere Pro
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: click · 指向与点击 geom · 精确几何 vis · 视觉判断
工作产物: WP05 · 时序媒体
条件: licensed · 授权专业软件
提案难度: 3 / 5

建议验证方式

Per frame: OCR caption bboxes vs GT lower-third bboxes overlap = 0; outside lower-third windows captions are bottom-centered (center x within 2% of frame center, bottom margin 6–10%); box opacity ~60% (luma sampling); text matches caption track (CER ≤ 2%).

建议部分得分

40% no overlaps; 25% default position; 20% styling; 15% text fidelity.

R06.04.02.03-T03 Re-segment a Spanish translation onto existing caption timings

原始指令

Resolve project 'Doc_Short' has an English subtitle track (84 events). input/es_translation.docx provides the Spanish as paragraphs, not split. Create a Spanish subtitle track with the identical 84 timings, splitting the Spanish text to match each event's meaning, ≤ 42 chars/line, ≤ 2 lines. Export out/doc_es.srt.

平台: mac · macOS win · Windows lin · Linux
应用: DaVinci Resolve
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: type · 文字与表单 click · 指向与点击
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 2 / 5

建议验证方式

SRT: 84 events with in/out identical to English (±0 frames); concatenated Spanish equals the provided text (normalized whitespace); per-event alignment vs GT split (≥ 85% events identical); line rules met.

建议部分得分

30% timings identical; 30% text complete and unaltered; 25% alignment to events; 15% line rules.

R06.04.02.03-T04 Cut 15-second and 6-second versions from a 30-second spot

原始指令

Premiere Pro sequence 'Spot30_Final' (30 s) — create 'Spot15' (15:00) and 'Spot06' (6:00) that both keep the end card (last 2 s) and the 2 brand-name shots (markers BRAND1, BRAND2; Spot06 needs only BRAND2), with the music edited to end cleanly on the end card. Export both to out/.

平台: mac · macOS win · Windows
应用: Premiere Pro
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: drag · 拖放 av · 音视频时序 vis · 视觉判断
工作产物: WP05 · 时序媒体
条件: licensed · 授权专业软件
提案难度: 4 / 5

建议验证方式

Durations exact; end card frames present in last 2 s; brand shots present (fingerprint) per spec; music ends with the final hit within 2 frames of end and edits at bar boundaries (no spectral-flux spike > 1.5× median); no shot < 12 frames.

建议部分得分

20% durations; 20% end cards; 25% brand shots; 35% music edits and no micro shots.

R06.04.02.03-T05 Reframe a 16:9 ad to 9:16 and fix Auto Reframe misses

原始指令

In Premiere Pro, create a 1080×1920 version of sequence 'Ad_Wide' (20 shots) using Auto Reframe, then fix the shots where the subject leaves the frame by keyframing position manually. Export out/ad_vertical.mp4.

平台: mac · macOS win · Windows
应用: Premiere Pro
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: geom · 精确几何 vis · 视觉判断 nav · 平移 / 缩放 / 滚动
工作产物: WP05 · 时序媒体
条件: licensed · 授权专业软件
提案难度: 3 / 5

建议验证方式

Per frame GT subject bbox (mapped through the export's crop) fully inside frame on ≥ 95% frames and ≥ 90% within every shot; crop motion smooth (no jumps > 5% width between frames except at cuts); 1080×1920. ExtendScript/UXP panels are disabled, ffmpeg is absent, and .prproj writes by other processes are detected, so reframing must be keyframed in Premiere.

建议部分得分

60% subject coverage; 25% per-shot minimum; 15% smooth motion and format.

R06.04.02.03-T06 Export three social aspect-ratio variants to spec

原始指令

From Resolve timeline 'Launch_Master' (UHD 16:9), deliver three files: out/launch_9x16.mp4 (1080×1920), out/launch_4x5.mp4 (1080×1350), out/launch_16x9.mp4 (1920×1080), all H.264 High, 30 fps, AAC 48 kHz stereo, with the logo fully inside each frame.

平台: mac · macOS win · Windows lin · Linux
应用: DaVinci Resolve
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: click · 指向与点击 geom · 精确几何
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 2 / 5

建议验证方式

ffprobe: resolution, codec profile, 30 fps, AAC 48 kHz 2 ch for each; GT logo template found fully inside the frame in all three on the logo frames; durations equal master.

建议部分得分

60% three encodes to spec (20 points each); 30% logo inside frame; 10% durations.

R06.04.02.03-T07 Time subtitles to shot changes and speech in Aegisub

原始指令

In Aegisub, open fixtures/drama_ep2.ass (rough timings) with fixtures/drama_ep2.mkv. Retime all 240 lines so each starts within 2 frames of speech onset, snaps to a shot change when within 12 frames of one, ends 12 frames after speech or at the next shot change, with ≥ 2-frame gaps. Save out/drama_ep2_timed.ass.

平台: win · Windows mac · macOS lin · Linux
应用: Aegisub
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: av · 音视频时序 key · 快捷键 nav · 平移 / 缩放 / 滚动
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 3 / 5

建议验证方式

Against GT speech onsets/offsets and keyframe list: ≥ 90% starts within 2 frames (or snapped to shot change per rule); ≥ 90% ends per rule; no overlaps; gaps ≥ 2 frames; text unchanged.

建议部分得分

40% start timing; 30% end timing and snapping; 20% gaps/no overlaps; 10% text intact.

R06.04.02.03-T08 Karaoke-time a song's lyrics syllable by syllable in Aegisub

原始指令

In Aegisub, fixtures/song_op.ass has 24 lyric lines already line-timed to fixtures/song_op.flac. Add karaoke timing (\k tags) so each syllable's start matches the sung onset, using the audio spectrum view. Save out/song_op_karaoke.ass.

平台: win · Windows mac · macOS lin · Linux
应用: Aegisub
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: av · 音视频时序 click · 指向与点击 key · 快捷键
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 4 / 5

建议验证方式

Parse \k durations into syllable onsets: ≥ 85% within ±50 ms of GT annotated onsets; each line's \k sum equals line duration ±1 cs; syllable split matches GT romaji segmentation ≥ 95%. Aegisub's Lua automation is disabled and .ass writes by other processes are detected, so syllable timing must be set in Aegisub's karaoke mode.

建议部分得分

60% syllable onset accuracy; 20% line sums consistent; 20% syllable segmentation.

R06.04.02.03-T09 Typeset a translated street sign that follows camera motion

原始指令

In Aegisub, add a typeset line translating the shop sign in fixtures/scene12.mkv (frames 1440–1560, camera pans and the sign shears slightly). Match the sign's color and angle, position it over the sign, and motion-track it (Aegisub-Motion or manual \pos/\frz/\fax per frame). Save out/scene12_ts.ass.

平台: win · Windows mac · macOS lin · Linux
应用: Aegisub
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G2 · GUI 必要
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: geom · 精确几何 vis · 视觉判断 nav · 平移 / 缩放 / 滚动
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 5 / 5

建议验证方式

Render with libass: per frame, rendered text bbox center within 4 px of GT sign center, rotation within 2° of GT sign angle, visible only in frames 1440–1560; text color ΔE2000 ≤ 6 to sign color; text = translation in input/ts.txt. Aegisub automation is limited to the bundled Aegisub-Motion macro and .ass writes by other processes are detected, so tracking must go through the GUI.

建议部分得分

45% per-frame position tracking; 20% angle; 15% timing window; 20% color and text.

R06.04.02.03-T10 Write SDH captions with speaker IDs and sound cues

原始指令

Resolve project 'Thriller_Ep1' has a 6-min scene with 3 speakers, music and key sound effects. Create an SDH English subtitle track: speaker IDs when off-screen or ambiguous, [door slams]-style cues for the 9 GT sound events, ♪ for music, per input/sdh_style.pdf. Export out/ep1_sdh.srt.

平台: mac · macOS win · Windows lin · Linux
应用: DaVinci Resolve
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: type · 文字与表单 av · 音视频时序 vis · 视觉判断
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 4 / 5

建议验证方式

WER on dialogue ≤ 5%; speaker IDs present on ≥ 90% of GT off-screen lines and absent where speaker is on-screen and unambiguous (≥ 80%); 9 sound events captioned within ±1 s with bracket format; music cues present; style rules met.

建议部分得分

30% dialogue accuracy; 25% speaker IDs; 25% sound cues; 20% music cues and style.

R06.04.02.03-T11 Deliver an MP4 with an embedded selectable subtitle track

原始指令

From Resolve project 'Lecture_Pub', render out/lecture.mp4 (H.264) with the English and French subtitle tracks embedded as separate selectable tracks (not burned in), languages tagged eng/fra, plus sidecar out/lecture_en.vtt and out/lecture_fr.vtt.

平台: mac · macOS win · Windows lin · Linux
应用: DaVinci Resolve
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: click · 指向与点击
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 4 / 5

建议验证方式

ffprobe: two subtitle streams (mov_text) with language tags eng and fra; no burned text in frames (OCR on caption times finds none); VTTs valid and identical in timing/text to the project tracks.

建议部分得分

40% two embedded tracks; 20% language tags; 20% not burned in; 20% VTT sidecars.

R06.04.02.03-T12 Fix captions to respect shot changes and minimum gaps

原始指令

In Subtitle Edit, open fixtures/feature_reel3.srt with fixtures/feature_reel3.mp4. Apply the delivery rules: no caption may cross a shot change unless it spans ≥ 1 s on both sides; ≥ 2-frame gap between captions; min duration 5/6 s; max 7 s. Fix all violations without changing text. Save out/reel3_fixed.srt.

平台: win · Windows lin · Linux
应用: Subtitle Edit
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: av · 音视频时序 click · 指向与点击 nav · 平移 / 缩放 / 滚动
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 3 / 5

建议验证方式

Rule checker against GT shot changes: 0 violations; caption starts/ends moved ≤ 0.5 s from original on average; text unchanged; count of captions unchanged or splits only where required.

建议部分得分

50% zero rule violations (scaled); 25% minimal timing drift; 25% text unchanged.

R06.04.02.03-T13 Create a German version of on-screen text graphics

原始指令

Premiere Pro sequence 'Product_Tour' has 5 Essential Graphics titles in English. Duplicate the sequence as 'Product_Tour_DE' and replace the text with German from input/de_titles.csv, adjusting tracking or size so each title fits its design box without overflow. Export out/tour_de.mp4.

平台: mac · macOS win · Windows
应用: Premiere Pro
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: type · 文字与表单 geom · 精确几何 vis · 视觉判断
工作产物: WP05 · 时序媒体
条件: licensed · 授权专业软件
提案难度: 3 / 5

建议验证方式

OCR of 5 title frames vs German CSV (CER ≤ 2%); rendered text bbox inside GT design box with ≥ 10 px margin; font size ≥ 85% of English; EN sequence unchanged.

建议部分得分

45% German text correct; 35% fits box; 10% size floor; 10% EN untouched.

R06.04.02.03-T14 OCR image-based Blu-ray subtitles into a clean SRT

原始指令

In Subtitle Edit, import fixtures/film_pgs.sup (image-based, 812 events). OCR it to text, fix OCR errors (l/I, rn/m, italics), and save out/film.srt with italics tags preserved and timings unchanged.

平台: win · Windows lin · Linux
应用: Subtitle Edit
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: type · 文字与表单 click · 指向与点击 vis · 视觉判断
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 3 / 5

建议验证方式

SRT: 812 events with timings equal to the .sup (±1 ms); CER vs GT text ≤ 1%; italics tags on ≥ 95% of GT italic events, none elsewhere.

建议部分得分

20% events/timings; 60% CER (scaled from 0.05 CER); 20% italics.

R06.04.02.03-T15 Build a textless version by covering burned titles

原始指令

Resolve timeline 'Brand_Film' has 7 burned-in titles over picture; the textless backgrounds are in bin 'Textless'. Create 'Brand_Film_TXTL' replacing each titled section with its textless background, frame-accurately aligned, and render out/brand_textless.mov.

平台: mac · macOS win · Windows lin · Linux
应用: DaVinci Resolve
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: drag · 拖放 vis · 视觉判断 nav · 平移 / 缩放 / 滚动
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 3 / 5

建议验证方式

OCR finds no text across the render; each replaced section matches its textless background frame-for-frame (SSIM ≥ 0.98 outside the text area vs original); duration and other frames identical to the titled version.

建议部分得分

40% all titles removed; 35% frame-accurate alignment; 25% rest identical.

R06.04.02.03-T16 Cut a 60-second spot to 30 seconds and carry its captions

原始指令

Premiere Pro sequence 'Charity_60' has a caption track. Create 'Charity_30' (exactly 30:00) keeping the 3 key VO lines marked KEEP and the end card, then make the captions follow the cut: remove captions for removed dialogue, retime split ones, keep rules (≤ 42 cpl, ≤ 20 cps). Export out/charity30.mp4 and out/charity30.srt.

平台: mac · macOS win · Windows
应用: Premiere Pro
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: drag · 拖放 type · 文字与表单 av · 音视频时序
工作产物: WP05 · 时序媒体
条件: licensed · 授权专业软件
提案难度: 4 / 5

建议验证方式

Duration 30:00; KEEP lines present (ASR); end card present; SRT text equals ASR of the cut (WER ≤ 3%) with every caption inside its line's speech span; no captions for removed audio; rules met.

建议部分得分

20% duration; 25% KEEP lines/end card; 35% captions follow cut; 20% rules.

R06.04.02.03-T17 Control: produce the T01 caption file with Whisper and a script

原始指令

Control pair of T01 (CLI half). Generate out/explainer_en_cli.srt from fixtures/explainer.wav with whisper (or whisper.cpp) plus your own segmentation script enforcing ≤ 42 cpl, ≤ 2 lines, ≤ 20 cps and ≥ 5/6 s.

平台: lin · Linux mac · macOS win · Windows
应用: whisper.cpp · Python
交互表面: term · 终端
GUI 需求 · 任务设定: CL · CLI 足够
GUI 需求 · 普通权限: CL · CLI 足够
交互原语: code · 代码与命令
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 2 / 5

建议验证方式

Same SRT checks as T01 (WER ≤ 5%, formatting rules, timing to speech). Paired with T01 to compare the in-app caption workflow with a CLI pipeline.

建议部分得分

40% text accuracy; 35% formatting; 25% timing.

R06.04.02.03-T18 Add stacked bilingual captions to a vertical video in CapCut

原始指令

In CapCut desktop, project fixtures/street_food_9x16 (60 s, Mandarin speech). Generate captions, correct them to input/zh.txt, add English lines from input/en.txt under each Chinese line (bilingual captions), Chinese 44 px above English 34 px, both inside the safe area. Export out/street_food_bi.mp4.

平台: mac · macOS win · Windows
应用: CapCut
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: G3 · 仅 GUI
GUI 需求 · 普通权限: G2 · GUI 必要
交互原语: type · 文字与表单 geom · 精确几何 av · 音视频时序
工作产物: WP05 · 时序媒体
条件: 未声明
提案难度: 3 / 5

建议验证方式

OCR: Chinese CER ≤ 3%, English WER ≤ 5%; Chinese line above English on every caption; glyph heights in ratio ~1.3; all text within the central 80% width and above the bottom 15%; timing overlaps speech ≥ 90%. ffmpeg is absent from the VM and draft files written by other processes are detected, so the export must come from CapCut (encoder metadata plus matching app draft).

建议部分得分

30% Chinese accuracy; 25% English accuracy; 20% stacking/sizes; 25% safe area and timing.

R06.04.02.03-T19 Assemble a full versioning delivery package

原始指令

From Premiere Pro project 'Campaign_Q4' (master 60 s), deliver per input/delivery_matrix.xlsx: 60/30/15 s cut-downs, each in 16:9, 9:16, 1:1 and 4:5, plus SRTs in EN/ES/FR for every cut, named <cut>_<ratio>_<lang>.ext in out/delivery/. Burn-in only for 9:16; sidecars for others.

平台: mac · macOS win · Windows
应用: Premiere Pro · Media Encoder
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: HY · 混合
交互原语: click · 指向与点击 drag · 拖放 type · 文字与表单 multi · 跨应用
工作产物: WP05 · 时序媒体
条件: licensed · 授权专业软件 long · 长流程 multiapp · 多应用
提案难度: 5 / 5

建议验证方式

Matrix check: all 12 videos + 9 SRTs present with exact names; durations exact; resolutions per ratio; subject inside frame (GT bbox ≥ 95%); burned captions only in 9:16 (OCR); SRT WER ≤ 5% vs cut ASR/translations; cut-downs keep brand shots.

建议部分得分

20% file matrix complete; 25% cut-down durations/content; 20% reframes; 35% captions per language and burn-in policy.

R06.04.02.03-T20 Author forced narratives and full subtitles in TTML

原始指令

Resolve project 'Border_Film' has dialogue in English with 4 scenes in Arabic and on-screen signs to translate. Create (a) a full English subtitle track and (b) a forced-narrative track only for Arabic dialogue and signs, positioned to avoid the signs; export both as IMSC1 TTML (out/full.ttml, out/forced.ttml) with the forced flag set.

平台: mac · macOS win · Windows lin · Linux
应用: DaVinci Resolve · Subtitle Edit
交互表面: desk · 桌面 GUI
GUI 需求 · 任务设定: HY · 混合
GUI 需求 · 普通权限: HY · 混合
交互原语: type · 文字与表单 av · 音视频时序 geom · 精确几何
工作产物: WP05 · 时序媒体
条件: multiapp · 多应用
提案难度: 5 / 5

建议验证方式

TTML validates against IMSC1 text profile; forced file covers 100% of GT Arabic dialogue and sign events and nothing else (precision ≥ 0.95); sign captions positioned not overlapping GT sign bboxes; full file WER ≤ 5%; forced flag present.

建议部分得分

30% valid IMSC1 and forced flags; 35% forced coverage/precision; 15% positioning; 20% full track accuracy.

已有研究映射

以下沿用研究目录中的基准与小类映射，不是逐题标注、执行记录或产品背书。

AgenticVBench

字幕与版本

数字员工关联

关联岗位

任务提案 · 20

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

原始指令

建议验证方式

建议部分得分

已有研究映射

同工作流族的小类