计算机使用智能体分类English
R06.07.01.01 · 任务小类

对白剪辑与音频修复

Dialogue editing & audio repair

Clean, edit and repair speech recordings and podcasts.

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
Audition · iZotope RX · Pro Tools
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。

R06.07.01.01-T01 Remove mouth clicks and coughs without touching clean speech

原始指令

fixtures/vo_take12.wav (synthetic voice, 48 kHz/24-bit, 3 min) has 6 mouth clicks and 2 coughs between words. Remove them in iZotope RX (spectral editor, De-click, Mouth De-click allowed) so the voice sounds untouched. Export output/vo_take12_clean.wav at the same format and length.

平台
mac · macOS win · Windows
应用
iZotope RX
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
path · 连续路径 vis · 视觉判断 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
3 / 5

建议验证方式

Null test vs input outside GT event regions (+/- 50 ms): residual <= -60 dBFS. Inside the 8 events: transient energy reduced >= 20 dB and spectral flux below GT threshold. Same length/sample rate/bit depth. The VM has no terminal, Python or sox/ffmpeg and RX's Batch Processor is disabled; the .rxdoc history must show spectral or module edits in-app, and batch processing of the full file is flagged.

建议部分得分

40% events removed; 40% rest unchanged (null test); 20% format and history.

R06.07.01.01-T02 Reduce HVAC noise while keeping intelligibility

原始指令

fixtures/interview_hvac.wav (synthetic speech plus recorded HVAC noise at -42 dBFS) needs cleanup in Adobe Audition. Learn a noise print from the room tone at 0:00-0:03 and reduce the noise floor by >= 12 dB with minimal speech artifacts. Export output/interview_denoised.wav (48 kHz/24-bit).

平台
mac · macOS win · Windows
应用
Adobe Audition · iZotope RX
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 vis · 视觉判断 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
3 / 5

建议验证方式

Noise floor in GT non-speech regions reduced >= 12 dB; ESTOI vs clean reference drops <= 0.02 relative to input; no musical noise (spectral kurtosis in noise regions below GT threshold); length and format unchanged. Control pair is T16.

建议部分得分

40% noise reduction; 40% intelligibility kept; 20% artifact check and format.

R06.07.01.01-T03 Cut filler words and long pauses from a podcast

原始指令

Edit fixtures/podcast_ep7.sesx (2 speakers, 22 min, synthetic voices): remove the 41 filler words listed in notes/fillers.csv, shorten every pause > 1.5 s to 0.5 s, keep breaths natural, and level both speakers to -16 LUFS within 1 LU. Bounce output/ep7_edit.wav (stereo, 48 kHz).

平台
mac · macOS win · Windows
应用
Adobe Audition · Pro Tools · Descript
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
av · 音视频时序 drag · 拖放 click · 指向与点击
工作产物
WP05 · 时序媒体
条件
licensed · 授权专业软件 long · 长流程
提案难度
4 / 5

建议验证方式

Forced alignment of the bounce vs transcript: >= 38 of 41 fillers absent; 0 non-filler words removed (WER on non-filler words <= 1%); max silence 0.55 s; edits at zero crossings or with crossfades (no click > -50 dBFS at cut points); each speaker -16 +/- 1 LUFS (diarized).

建议部分得分

30% fillers removed; 20% no content lost; 20% pauses; 30% clean cuts (15 pts) and speaker loudness (15 pts).

R06.07.01.01-T04 Spectrally remove a phone ring under dialogue

原始指令

A phone rings for 4 s under dialogue in fixtures/scene14_boom.wav (synthetic voices). Remove the ring with spectral repair or spectral selection attenuation so the dialogue under it stays intact. Export output/scene14_boom_fixed.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition
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
path · 连续路径 vis · 视觉判断 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

Fixture is a synthetic mix with known stems: in the ring region SDR vs clean dialogue improves >= 12 dB; ring partial energy reduced >= 25 dB; outside region null residual <= -60 dBFS; ESTOI in region >= 0.9 of clean. In-app spectral edit recorded in history. The VM has no terminal, Python or sox/ffmpeg; RX's Batch Processor is disabled and WAVs not written by RX or Audition are logged and fail.

建议部分得分

40% ring removed; 35% dialogue intact under the ring; 25% nothing else changed.

R06.07.01.01-T05 Repair clipped dialogue peaks

原始指令

fixtures/lav_clipped.wav has 23 clipped passages (flat-topped at -0.1 dBFS). Repair them with RX De-clip (or Audition DeClipper), reduce gain so restored peaks do not exceed -1 dBFS, and export output/lav_declipped.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition
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 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
2 / 5

建议验证方式

Flat-topped run detector: 0 runs >= 4 identical samples near full scale; restored waveform vs GT unclipped source in clipped regions SNR >= 15 dB; true peak <= -1 dBTP; outside clipped regions the gain-compensated null residual <= -55 dBFS.

建议部分得分

40% clipping removed; 35% waveform restoration; 25% headroom and untouched regions.

R06.07.01.01-T06 Fix plosive pops in a voice-over

原始指令

fixtures/vo_plosives.wav has 8 plosive pops (GT times in notes/plosives.txt). Reduce each pop's low-frequency burst (< 150 Hz) by >= 12 dB with De-plosive or local high-pass edits, without filtering the rest of the file. Export output/vo_fixed.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition · Audacity
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 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
2 / 5

建议验证方式

At each GT pop: energy below 150 Hz in a 60 ms window reduced >= 12 dB; consonant energy above 1 kHz within 1 dB; outside pop windows null residual <= -60 dBFS (no global high-pass).

建议部分得分

50% pops reduced; 25% consonants intact; 25% rest unchanged.

R06.07.01.01-T07 Assemble a scene from selected dialogue takes

原始指令

In fixtures/sc22_dialogue.ptx, 3 takes per line exist on playlists. Using the director's selects in notes/sc22_selects.csv (take per line), cut a single clean dialogue track aligned to picture guide track 'DX_Guide', with crossfades at every edit and room tone under gaps. Bounce output/sc22_dx.wav.

平台
mac · macOS win · Windows
应用
Pro Tools · Nuendo
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 drag · 拖放 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
4 / 5

建议验证方式

Each line's audio in the bounce matches the selected take (fingerprint correlation >= 0.95 vs that take, < 0.6 vs others); line start sync to guide track within 1 frame; no click > -50 dBFS at edits; no digital silence in gaps. Pro Tools' PTSL scripting port and Nuendo's MIDI Remote scripting are disabled and the VM has no terminal or audio CLI tools; session edit list checked.

建议部分得分

40% correct takes; 25% sync to guide; 20% clean edits; 15% room tone fill.

R06.07.01.01-T08 Fill gaps with matching room tone

原始指令

fixtures/dx_gaps.wav has 6 digital-silence gaps (0.4-2.5 s) left by cuts. Fill each with room tone taken from the same file (no generated noise) so no digital silence remains and the background is continuous. Export output/dx_filled.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition · Pro Tools
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
drag · 拖放 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
2 / 5

建议验证方式

No run of >= 100 samples with |x| < 1e-5; each filled gap's spectrum within 2 dB per 1/3-octave band of the surrounding background; fills sourced from the file (cross-correlation finds the source segment); dialogue regions unchanged (null <= -60 dBFS).

建议部分得分

40% gaps filled; 30% tonal match; 15% sourced from file; 15% dialogue untouched.

R06.07.01.01-T09 Reduce excessive room reverb on dialogue

原始指令

fixtures/hall_interview.wav was recorded in a reverberant hall (RT60 about 0.9 s). Use RX Dialogue De-reverb (or Audition DeReverb) and EQ so estimated RT60 drops to <= 0.4 s while intelligibility improves. Export output/hall_interview_dry.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition
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 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
3 / 5

建议验证方式

Blind RT60 estimate (decay-rate method on speech offsets) <= 0.4 s; ESTOI vs dry reference >= input + 0.05; no pumping (envelope modulation in pauses below threshold); level change <= 2 LU.

建议部分得分

40% RT60 reduced; 35% intelligibility gain; 25% artifact and level checks.

R06.07.01.01-T10 Remove mains hum and harmonics from a location recording

原始指令

fixtures/hum_location.wav has 50 Hz mains hum with harmonics to 600 Hz that drifts by +/- 0.3 Hz. Remove it with RX De-hum (adaptive) or notch filters so hum partials drop >= 30 dB while voice fundamentals in the same bands are kept. Export output/hum_removed.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition · Audacity
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 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

Fixture is synthetic hum added to clean speech: energy at 50 Hz and 11 harmonics (tracked frequency) reduced >= 30 dB; speech SDR vs clean >= 25 dB; 100-300 Hz speech band within 1.5 dB of clean in speech frames.

建议部分得分

50% hum reduction across harmonics; 30% speech preserved; 20% drift handled.

R06.07.01.01-T11 Clean wind noise from a lavalier recording

原始指令

fixtures/lav_wind.wav (synthetic voice with recorded wind buffeting) has wind gusts in 9 passages. Reduce wind rumble and gusts (RX De-wind, spectral repair, or clip-gain plus high-pass per passage) without thinning the voice elsewhere. Export output/lav_wind_fixed.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition
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
path · 连续路径 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
3 / 5

建议验证方式

In gust passages: energy below 200 Hz reduced >= 15 dB and ESTOI vs clean >= input + 0.08; outside passages spectral balance within 1 dB per band of input (no global filtering). History shows per-passage edits. The VM has no terminal, Python or sox/ffmpeg; RX's Batch Processor is disabled and WAVs not written by RX or Audition are logged and fail.

建议部分得分

40% gust reduction; 30% intelligibility; 30% unaffected regions preserved.

R06.07.01.01-T12 Match a lav mic to the boom with EQ matching

原始指令

Lines 4-9 of fixtures/sc03_lav.wav must cut seamlessly with boom lines in fixtures/sc03_boom.wav. Use EQ Match (RX or Audition Match Loudness/EQ) plus manual EQ so the lav's long-term spectrum matches the boom within 2 dB per 1/3-octave band (100 Hz-10 kHz). Export output/sc03_lav_matched.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition · Pro Tools
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 vis · 视觉判断 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
3 / 5

建议验证方式

Long-term average spectrum of speech frames: |lav - boom| <= 2 dB per 1/3-octave band 100 Hz-10 kHz; loudness within 0.5 LU; no added noise above -70 dBFS; length unchanged.

建议部分得分

60% spectral match; 20% loudness match; 20% artifacts and format.

R06.07.01.01-T13 Sync an ADR line to the production guide track

原始指令

In fixtures/adr_sc9.ptx, align the ADR take on track 'ADR_9' to guide track 'PROD_9' word by word so lip-sync matches picture, using VocALign/Revoice or manual clip edits and time compression (no pitch change). Bounce output/adr_sc9_synced.wav.

平台
mac · macOS win · Windows
应用
Pro Tools · Nuendo
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 drag · 拖放 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

Forced alignment of both tracks: word onset error <= 1 frame (41.7 ms at 24 fps) on >= 90% of words; median pitch unchanged within 10 cents; no artifacts (spectral discontinuity at edits below threshold); length equals guide +/- 1 frame. Pro Tools' PTSL scripting port and Nuendo's MIDI Remote scripting are disabled, the VM has no terminal or audio CLI tools, and the session edit list is checked.

建议部分得分

55% word sync; 15% pitch unchanged; 15% clean edits; 15% length match.

R06.07.01.01-T14 Reduce lav bleed between two interview mics

原始指令

In fixtures/two_lav_interview.sesx each speaker's lav picks up the other speaker. Reduce bleed on each track (expander/gate tuned per speaker plus manual clip-gain) by >= 15 dB when that speaker is silent, without chopping word onsets or breaths. Export both tracks to output/lavA.wav and output/lavB.wav.

平台
mac · macOS win · Windows
应用
Adobe Audition · Pro Tools · REAPER
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
av · 音视频时序 click · 指向与点击 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

Diarized GT: on each track, energy during the other speaker's turns reduced >= 15 dB; own-speech onset energy within 1 dB in the first 50 ms of words (no clipped onsets); breaths retained >= 70%; no gate chatter (state flips < 3 per second).

建议部分得分

45% bleed reduction; 30% onsets intact; 15% breaths; 10% no chatter.

R06.07.01.01-T15 Lower breaths in a narration without deleting them

原始指令

fixtures/narration_ch3.wav (synthetic narrator, 12 min) has 40 breaths that are too loud. Lower each breath by 10-15 dB (Breath Control or clip gain) without removing them or affecting adjacent words. Export output/narration_ch3_breaths.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition · Pro Tools
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
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

At each GT breath: level reduction 10-15 dB (not > 20 dB, not muted); adjacent word envelopes within 0.5 dB of input; >= 37 of 40 breaths treated; outside breath regions null residual <= -60 dBFS.

建议部分得分

50% breaths reduced in range; 25% not removed; 25% words untouched.

R06.07.01.01-T16 Denoise the interview from the terminal (control pair of T02)

原始指令

Control pair of T02: using only a terminal (sox noisered with a profile from 0:00-0:03, or ffmpeg afftdn/arnndn), reduce the HVAC noise floor in fixtures/interview_hvac.wav by >= 12 dB and write output/interview_denoised_cli.wav (48 kHz/24-bit).

平台
mac · macOS win · Windows lin · Linux
应用
sox · ffmpeg
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
3 / 5

建议验证方式

Same thresholds as T02: noise floor -12 dB, ESTOI drop <= 0.02, musical-noise check, format unchanged. Scored side by side with T02 to measure the GUI-vs-CLI delta.

建议部分得分

40% noise reduction; 40% intelligibility kept; 20% artifact check and format.

R06.07.01.01-T17 Phase-align boom and lav recordings

原始指令

fixtures/sc11_boom.wav and fixtures/sc11_lav.wav are the same performance; the boom arrives 6.4 ms late and is polarity-inverted. Align and correct polarity (clip nudge, Auto-Align or RX Phase) so summing them causes no comb filtering, then export the aligned boom as output/sc11_boom_aligned.wav.

平台
mac · macOS win · Windows
应用
Pro Tools · Adobe Audition · iZotope RX
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
drag · 拖放 vis · 视觉判断 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
3 / 5

建议验证方式

Cross-correlation peak between lav and aligned boom at lag 0 +/- 0.1 ms with positive sign; summed spectrum shows no notches deeper than 6 dB in 200 Hz-4 kHz; length equal to lav; no resampling artifacts.

建议部分得分

45% time alignment; 30% polarity; 25% sum without comb notches.

R06.07.01.01-T18 Restore a degraded archival interview

原始指令

fixtures/archive_1962_sim.wav is a synthetic archival interview with hiss, crackle, 3 dropouts (20-80 ms), 60 Hz hum, and bandwidth limited to 5 kHz. Restore it for broadcast: remove crackle, hiss and hum, repair dropouts by interpolation, gentle bandwidth extension allowed, final -23 LUFS. Export output/archive_restored.wav.

平台
mac · macOS win · Windows
应用
iZotope RX · Adobe Audition
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
path · 连续路径 vis · 视觉判断 av · 音视频时序
工作产物
WP05 · 时序媒体
条件
licensed · 授权专业软件 long · 长流程
提案难度
5 / 5

建议验证方式

Known clean source: SI-SDR improvement >= 8 dB; crackle impulse count reduced >= 90%; hum partials -30 dB; dropouts filled (no gaps, local SNR >= 10 dB vs clean); loudness -23 +/- 0.5 LUFS; no over-processing (ESTOI >= input).

建议部分得分

25% crackle/hiss; 15% hum; 25% dropouts; 35% overall SI-SDR (20 pts) and loudness (15 pts).

R06.07.01.01-T19 Conform production audio to a locked cut from AAF and EDL

原始指令

Picture is locked: import fixtures/reel2_lock.aaf into Pro Tools, relink to the production polywavs in fixtures/prod_audio/ (8 iso tracks, timecode stamped), use notes/reel2.edl to restore lav isos with 48-frame handles where the AAF only carries the mix track, and split dialogue per character. Save output/reel2_dx.ptx and bounce stems.

平台
mac · macOS win · Windows
应用
Pro Tools · Nuendo
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 drag · 拖放 av · 音视频时序
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

Session export (AAF/EDL from the result) parsed: every EDL event has the correct iso clip at its source TC (+/- 0 frames) with >= 48-frame handles; character tracks per notes; bounced stems null against the GT conform within -50 dB. Media all online.

建议部分得分

30% AAF imported and relinked; 30% iso events conformed; 20% handles; 20% stems match GT.

R06.07.01.01-T20 Edit and repair a podcast episode per producer notes

原始指令

Producer notes notes/ep12_notes.md (timestamps) list: cut 3 tangents, fix 2 crosstalk overlaps, swap in a pickup line from fixtures/pickup.wav with matched tone, remove 12 noises, de-ess the host, and deliver at -16 LUFS / -1 dBTP stereo. Session fixtures/ep12.sesx; bounce output/ep12_final.wav.

平台
mac · macOS win · Windows
应用
Adobe Audition · iZotope RX · Pro Tools
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
av · 音视频时序 drag · 拖放 click · 指向与点击
工作产物
WP05 · 时序媒体
条件
licensed · 授权专业软件 long · 长流程
提案难度
5 / 5

建议验证方式

Transcript alignment: tangents absent, all other content present (WER <= 1%); pickup inserted at the noted point with EQ match within 2 dB/band; 12 noises reduced >= 20 dB; sibilance 5-9 kHz peak-to-average reduced >= 3 dB on host; -16 +/- 0.5 LUFS, true peak <= -1 dBTP.

建议部分得分

25% content edits; 20% pickup and tone match; 35% noise fixes (20 pts) and de-ess (15 pts); 20% delivery spec.

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。

同工作流族的小类