Communitygithub.com

freudianflips/breakout-launch-videos

为发布影片生成配音的工具集。支持多种语音供应商、自由克隆、自动适配和对齐,确保每个关键词与画面精准同步。

breakout-launch-videos 是什么?

该技能为发布影片提供完整的配音工作流,从角色选择、台词录制到强制对齐和最终合成。它默认使用Higgsfield的预设语音库(约100种)进行角色遴选,同时提供免费占位语音(macOS `say`)和本地自由克隆(Chatterbox)选项,适合预算有限或需要快速原型的情况。核心功能包括:按帧台词录制、基于语速和音调变化的角色排名、通过强制对齐生成词级时间戳(支持torch/whisper.cpp)、自动适配行内停顿和超速(上限12%),以及最终拼接。输出可直接用于`launch-motion`和`launch-sound`模块,实现单词级画面驱动。对运行时间过长的台词提供重写或ffmpeg变速修复;对古风纪录片风格支持频带限制和黑胶噪声效果。整体强调“语音先行,画面跟随”的现代发布影片范式。

兼容平台✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/freudianflips/breakout-launch-videos/tree/HEAD/skills/launch-voice

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档

发布配音

位于 launch-film 下的工具手册。从 skills/launch-film/SKILL.md 开始:那里包含核心内容、钩子加正文模板以及经验教训。凡本文档与此页不同之处,以 launch-film 为准。

在现代发布影片中,语音引导画面逐词演绎:当配音说到“标记”时,某物就被标记。这只有在每个词都有时间戳时才能实现。

供应商

  • 默认:使用你自己的 Higgsfield 账号。 text2speech_v2(引擎包括 elevenlabs、minimax、seed_speech 等,依语音而异)用于在约100种预设语音中进行角色遴选,每行约消耗0.3积分。Seed Audio 则提供一种你可以通过提示词和音频参考来操控的叙述者风格。scripts/hf_voice.py 是 CLI 的封装。
  • 免费占位语音: scripts/scratch_voice.py 使用 macOS 的 say(Linux 上为 espeak-ng)生成占位音频。以此为全片构建并计时,每次评审标明“占位语音”,之后替换为正式音频。
  • 免费本地克隆: scripts/clone_local.py(Chatterbox,MIT 协议)可从已批准叙述者的8到10秒台词中克隆其声音。仅作替身,非正式母带。只克隆你拥有版权的语音。
  • 始终在全片上下文中试听语音,而非零散片段。保留被否决语音的列表,避免重复推荐。

遴选

  1. python3 skills/launch-voice/scripts/hf_voice.py cast "<影片开场台词>" film/vo/audition 将用每一种预设语音朗读该行(约110种语音,约34积分),然后使用 whisper 检查每个音频,并测量语速和音调变化。ranking.md 会优先列出精确匹配的行,最接近每秒2.5个词且音调变化约2个半音(对话风格,非播音腔)的行。--only Name,Name 可以仅重新遴选少数语音。
  2. 从男女声中各取前6名左右进入全剧本(hf_voice.py lines),然后进行适配(见下文)。如果某语音需要超过12%的加速才能适配,说明它对该剧本来说太慢:将其剔除或删减台词。
  3. 每个保留的语音构建一个完整的影片,由用户挑选。ElevenLabs 在测试中语速和音调变化最小,因此是默认引擎。

长度与语速

  • 目标语速为每秒2.2至2.7个词。一个优秀的参考发布影片是2.4(63秒内含151个词)。
  • 一部20秒的影片大约容纳45个词;30秒约70个;60秒约140个。在加速语音之前,先删减台词。
  • 每行对应一个分镜帧,6到20个词,措辞应使关键名词或动词落在节拍点上。

影片模板的工作流程

  1. 剧本位于 film/film.json 的 lines 中。每行录制一个文件到 film/vo/body/line-<id>.wav:使用 hf_voice.py lines 录制正式音频(将台词写为 {"lines": [{"line": "02", "text": "..."}]}),或使用 python3 skills/launch-voice/scripts/scratch_voice.py film/film.json 生成占位音频(添加 --hook film/hooks/<name> 可同时为钩子行配音)。
  2. 运行 python3 skills/launch-voice/scripts/prep_lines.py film/film.json(添加 --align force 以强制使用强制对齐;若已安装 uv,首次运行时会下载 torch,约2 GB)。该脚本会生成 film/vo/body/lines.json,其中包含每行的持续时间和单词时间:当 torch 可用时使用强制对齐,否则使用 whisper.cpp,若两者均不可用则使用均匀估计并标记为 estimate。
  3. film/build.py 将各行放置在影片时间线上,每个画面提示都跟随其对应的单词。替换一次音频只需一次预运行加一次重建。

自由形式影片的工作流程

  1. vo/script-lines.json({"lines": [{"line": "01", "text": "..."}]}),然后每行一个文件:hf_voice.py lines vo/script-lines.json vo/takes/<voice> --voice <Name>。文本、语音、引擎和任务ID都会写入 lines.json。
  2. 适配:python3 skills/launch-voice/scripts/fit_lines.py vo/placement.plan.json vo/takes/<voice> vo/versions/<voice>。计划为每行指定其最早开始时间、呼吸间隙以及(如果某帧必须结束)end_by。行内超过0.25秒的停顿将缩短至0.22秒;若连续多行仍超时,则将这些行整体加速,上限为12%。然后拼接:python3 skills/launch-voice/scripts/stitch.py vo/versions/<voice>/placement.json vo/versions/<voice> vo/versions/<voice>/vo.wav。
  3. 每行获取单词时间。对于视觉效果必须落定的任何内容,使用强制对齐:uv run -q --with torch --with torchaudio --with soundfile python skills/launch-voice/scripts/force_align.py line.wav "确切的文字"。快速读取时,align_lines.py 逐行运行 whisper.cpp。锚点按说话顺序解析(每个锚点是其前一个锚点之后的第一个匹配项),因此任何换用语音都会重新计时影片。对整个影片进行单次 whisper 处理会在行末尾处产生漂移;逐行处理则不会。
  4. 将单词时间交给 launch-motion(每个构建锚定到其对应的单词)和 launch-sound(事件跟随单词)。

复古旁白

用于档案片或纪录片开场:python3 skills/launch-voice/scripts/vintage.py IN.wav OUT.wav --crackle crackle.wav 将带宽限制在150 Hz到5.8 kHz,添加磁带饱和、缓慢的晃动和真实的黑胶爆裂声底层(可用 Seed Audio 生成爆裂声)。Seed Audio 也可以直接以复古旁白声音朗读一行台词(“温暖、浑厚的1960年代纪录片旁白在旧黑胶唱片上……确切文字如下:……”);它读得较慢,并可能重新措辞(例如“gotta”变为“got to”),因此需用 whisper 核对。

同步规则

  • 关键词必须在2帧之内与其对应的视觉效果对齐。以单词的第一帧开始构建;如果允许视觉效果领先于单词,则领先不超过4帧,并且全程保持一致。
  • 在场景切换处让语音有呼吸空间:行与行之间留0.25到0.5秒。
  • 最后一行至少应在徽标出现前1秒结束;徽标应有其自身的近乎静音时段(launch-mix 的下陷)。

当语音供应商不可用时的处理

使用占位语音进行构建,所有评审标明为“占位语音”,画面保持锚定于对齐后的单词,这样换回正式语音只需一次预运行加一次渲染。未经用户同意,绝不更换语音供应商或账号。

修复

  • 一行台词超时少于5%:使用 ffmpeg -af atempo=1.04。超过5%:重写该行。
  • 单词发音错误:在行的 say 字段中按语音重拼(针对 hf_voice.py),切勿修改原始剧本。
  • Whisper 误判软结尾:使用强制对齐或缩短锚点前缀;不要重新录制。
  • 数字:在对齐的文本中拼写出来(例如“twenty four”,而不是“24”)。
  • Whisper 基础模型有时会误判产品名称。请使用小型或更大模型(WHISPER_MODEL)检查最终单词。

Individual skills in this repo

This repo contains 9 individual skills — each has its own dedicated page.

freudianflips/breakout-launch-videos

Build the brand kit a launch film uses from a website. Runs tools/brand-from-url.mjs on the reader's site to extract branding (colours, fonts, logo, name, tagline, copy, product images), then reviews the preview sheet with the reader and fixes any wrong role by hand. Use for "brand from website", "brand kit", "extract branding", "get our colours and fonts", "use our logo", or before the first launch film for a new brand.

freudianflips/breakout-launch-videos

Make or change a product launch film in your own brand. A film is a hook (the story before the cut) on a shared body (the name punching in, hero words cut on the voice, your product at close range, the logo lockup in near silence), built in HyperFrames from brand/brand.json and film/film.json. Holds the core, the template workflow (build, render and mix 1 hook or all of them), the generation recipes, the checks and the lessons. Use for "launch video", "launch film", "product video", "promo", "reveal", "announcement video", "new hook", "change the body", "render all launch videos".

freudianflips/breakout-launch-videos

Tool manual under launch-film (mixdown.py for buses, ducking, loudness and the brickwall). Mix and master the audio of a launch film. Music, SFX and voice-over on separate buses, the bed ducked by the voice's real envelope with a presence carve, the music dropped out for the logo, glue compression, a true-peak-safe limiter, -14 LUFS and -1 dBTP, then laid onto the rendered picture. Use after the score, the sound effects and the voice exist, for "mix", "master", "loudness", "LUFS", "ducking", "mux the audio".

freudianflips/breakout-launch-videos

Tool manual under launch-film (the motion language for launch films built in HyperFrames). Ease tokens, camera doctrine, continuity at every cut, type and cursor craft, and brand-neutral signature moves (logo build, world pass, branch and pulse, waterfall fill, identity morph, macro pull-back, rise) that replace HyperFrames' stock look. Says which HyperFrames blueprints and transitions to keep, transform or avoid. Use whenever a launch film's picture is designed or animated, for "animation", "motion design", "transitions", "camera move", "make it feel like Apple", "less template".

freudianflips/breakout-launch-videos

Tool manual under launch-film (measuring a render and studying a reference film). Review a rendered launch film like a demanding creative director, then drive the v1 to v2 loop. Measures rhythm, static runs, cuts against the beat grid, loudness, true peak and voice pace with a script, then applies an eye checklist for continuity at every cut, legibility, brand, real UI, claims and differentiation. Writes a gap list and a scorecard; nothing ships on numbers alone. Use on every render, and on a reference film to learn from it, for "review the video", "check the film", "what is wrong with this cut", "study this reference".

freudianflips/breakout-launch-videos

Tool manual under launch-film (music fitted to the edit with Higgsfield Sonilo, ACE-Step locally, or the free scratch bed). Score a launch film with music whose sections are the film's acts. Each take is measured for tempo and section energy, stretched onto the film's BPM and offset so its drop lands on the cut. Use after the storyboard's beat sheet exists, for "music", "score", "soundtrack", "background track", "the drop", "beat grid".

freudianflips/breakout-launch-videos

Tool manual under launch-film (the sound palette, the event map, the cue sheet and generated sound kits). Design the sound effects of a launch film. Builds an event map from the storyboard, the voice-over word timings and the rendered timeline, voices every event from 1 consistent palette (the free synthesised, key-locked forge, or real glass, felt, wood, marimba, clicks and air generated on Higgsfield and pitched into the film's key), and renders 1 SFX stem in a shared acoustic space. Use once picture timing and the score are fixed, for "sound effects", "SFX", "sound design", "UI sounds", "sonic logo".

freudianflips/breakout-launch-videos

Tool manual under launch-film (the brief, the story spine, the script, the beat sheet and reference studies). Turn a launch brief into a launch film's story in your own brand, with the one message, the narrative spine, the voice-over script at the right length, a beat sheet on the music's bar grid and a STORYBOARD.md where every shot names its continuity in and out, its signature move and its sound. Includes the reference study and the logo-swap differentiation test. Use first, before voice, score or motion, for "launch video script", "storyboard", "brief", "what should the launch film say", "study this reference video".

freudianflips/breakout-launch-videos

Tool manual under launch-film (backgrounds, plates, Blender glass and 3D shots, shader layers). Build the world a launch film lives in, in your own brand, with generated background plates (cost-gated), Blender for real 3D glass objects and camera moves rendered as alpha layers, WebGL shader layers for atmosphere, and the layer order, depth of field and grade that keep UI legible over living backgrounds. Use when a storyboard needs backgrounds, 3D or atmosphere, for "background", "plate", "3D object", "glass", "Blender", "shader", "atmosphere"; motion of the UI itself is launch-motion.

相关技能