本地、非商業的「影片 → 音效」創作工具
A local, non-commercial video → sound-effects tool.
設計製作 · Isan(13soul) — 全端設計工程師 · 影像及音樂創作人
你有一段沒有聲音(或只有原聲)的影片,想替它加上對得準的音效:whoosh、撞擊、腳步、玻璃碎、環境底聲…
AudioSFX 在你自己的電腦(本地 GPU)跑兩個 AI 模型生成音效,讓你在時間軸上對齊、混音,再用 ffmpeg 輸出成帶音效的影片。
- Sony Woosh — 文字/影片 → SFX(whoosh / punch / impact,超快)
- MMAudio — 影片 → 同步環境底聲(依畫面生成貼合的背景音)
⚠️ 兩個模型權重是 CC BY-NC(僅限非商業),本工具維持非商業定位,不可接進收費產品。
- 🎬 9:16 直式時間軸編輯器 — 影片/音訊/文字多軌,拖移・裁切・分割
- 📚 內建 CC0 音效庫 — 400+ 即用音效(Kenney.nl),免引擎、即點即用:撞擊/腳步/介面/電子/科技/物件
- 🗂 本機音效資料夾 — 指向你自己的音效夾(如 Sonniss 免費包),自動依關鍵字分中文類型(車輛/撞擊/環境/槍械…)瀏覽、試聽、匯入
- 🤖 三種音效生成入口 — 文字→SFX、影片→V2A 同步音、音軌指定位置生成
- 🎚 DAW 風格混音 — 每軌 EQ / 壓縮器 / 限幅器(可拖曳浮動視窗)+ 母帶防爆 limiter
- 🖥 浮動預覽窗 — 右下可拖/可調大小
- ⚡ ffmpeg 渲染 — NVENC GPU 自動加速,匯出 MP4 / MOV / WebM / WAV
- 📦 桌面安裝版 — Electron 打包,內嵌後端 + ffmpeg,動態 port
- 🧠 AI 可控(MCP) — 內建 MCP server,讓 Claude / Codex 直接驅動編輯器(建專案/生成音效/排時間軸)
桌面版:到 Releases 下載 AudioSFX Setup,執行即可。
開發模式:
npm install
npm run dev # 前端 :6300 / 後端 :6301Python 推論引擎(音效生成需要,GPU)— 首次需自行安裝(下載權重等),詳見 ENGINES.md。 裝好後從 UI 右上「引擎」面板啟動,或手動:
pwsh engines/start-woosh.ps1 # :6302 Woosh
pwsh engines/start-mmaudio.ps1 # :6303 MMAudio沒啟動引擎也能剪輯/匯出,只是不能生成音效。
需求:Node.js 20+、Windows、NVIDIA RTX GPU(音效生成)、ffmpeg。
內建 MCP server(server/mcp-server.mts)讓 Claude / Codex 控制正在跑的編輯器(建專案/生成音效/排時間軸)。在 repo 開 Claude Code 會偵測 .mcp.json 並詢問啟用 audiosfx;需後端在跑(npm run dev)。詳見 Wiki。
例:對 Claude 說「用 audiosfx 建 30 秒專案,生成 glass shatter 音效,加到音軌第 2 秒,存檔」。
- 建專案 + 匯入 — 專案 ▾ → 新建;把影片/音檔拖進視窗。
- 時間軸剪輯 — 拖移・裁切・分割;軌道可改名/鎖定/靜音。選中軌靛紫、靜音軌琥珀。
- 加音效:頂部〰「音效」浮窗,三個分頁——
- 音效庫(最快,免引擎):選分類/搜尋 → ▶ 試聽 → 「+軌」加到時間軸,或「素材」只匯入。
- 本機:指向你自己的音效資料夾(在欄位填路徑,如
D:\sfx_sample),自動分中文類型瀏覽。免費 Sonniss 包可用scripts/download-sonniss.mjs --dest <資料夾>抓(見 Wiki)。 - AI 生成 — 文字 → SFX:英文 prompt(如
glass shatter)→ 生成(需 Woosh 引擎)。 - 影片 → V2A:右鍵影片 clip → 生成同步底聲(MMAudio)。
- 指定位置:右鍵音軌空白 →「建立音效」(可命名)。
- 混音 — 左欄「效果」分頁 → 選軌 → fader / M;「+ 加效果」→ EQ / 壓縮 / 限幅(點名字開浮窗);Main Out 的 LIMIT 防爆母帶。
- 預覽 — 頂部 🖥 開右下浮動預覽窗。
- 匯出 — ⬆ 匯出 →(自動存檔)→ 關於作者 → 下一步 → 設定 → 匯出。成品右下角有「Isan 13soul」浮水印。
小技巧:
Ctrl+S存檔;改過前端Ctrl+Shift+R硬重整。
| Layer | Tech |
|---|---|
| Frontend | React 18 · Vite 5 · TailwindCSS 3 · Zustand 4 |
| Backend | Node.js 20+ · Express 4 · TypeScript 5 |
| 音訊預覽 | Web Audio API |
| 渲染 | ffmpeg(fluent-ffmpeg),NVENC fallback |
| 推論 | MMAudio(py3.10)· Woosh(py3.13),各自 FastAPI |
| Native | Rust + napi-rs(波形/縮圖/probe) |
| Desktop | Electron + electron-builder |
PolyForm Noncommercial License 1.0.0 — 僅限非商業用途。詳見 LICENSE。 依賴的 MMAudio / Sony Woosh 權重亦為 CC BY-NC。 本工具打包並以獨立執行檔呼叫 FFmpeg(GPLv3);第三方授權與來源見 NOTICE.md。
Designed & built by Isan (13soul) — full-stack design engineer · video & music creator
Got a silent (or raw-audio) video clip and want well-timed sound effects — whooshes, impacts, footsteps, glass shatter, ambience?
AudioSFX runs two AI models on your own machine (local GPU) to generate SFX, lets you align and mix them on a timeline, then muxes everything back into the video with ffmpeg.
- Sony Woosh — text / video → SFX (whoosh / punch / impact, very fast)
- MMAudio — video → synced ambient bed (background audio that fits the footage)
⚠️ Both model weights are CC BY-NC (non-commercial only). This tool stays non-commercial and must not ship inside any paid product.
- 🎬 9:16 vertical timeline editor — video / audio / text tracks; move, trim, split
- 📚 Built-in CC0 SFX library — 400+ ready-to-use sounds (Kenney.nl), no engine needed, instant: impacts / footsteps / UI / digital / sci-fi / foley
- 🗂 Local SFX folder — point at your own sound folder (e.g. free Sonniss packs); auto-tagged into categories by keyword (vehicle / impact / ambience / gunfire…) to browse, preview, import
- 🤖 Three ways to generate SFX — text→SFX, video→V2A synced audio, generate at a track position
- 🎚 DAW-style mixing — per-track EQ / compressor / limiter (draggable floating editors) + master brickwall limiter
- 🖥 Floating preview — bottom-right, draggable & resizable
- ⚡ ffmpeg rendering — automatic NVENC GPU acceleration; export MP4 / MOV / WebM / WAV
- 📦 Desktop build — Electron-packaged with an embedded backend + ffmpeg, dynamic port
- 🧠 AI-controllable (MCP) — a built-in MCP server lets Claude / Codex drive the editor (create projects, generate SFX, arrange the timeline)
Desktop: download AudioSFX Setup from Releases and run it.
Dev:
npm install
npm run dev # frontend :6300 / backend :6301Python inference engines (needed for SFX generation, GPU) — first-time setup (download weights, etc.) is on you: see ENGINES.md. Once installed, start them from the "引擎" panel (top-right of the UI), or manually:
pwsh engines/start-woosh.ps1 # :6302 Woosh
pwsh engines/start-mmaudio.ps1 # :6303 MMAudioEditing/exporting works without the engines — you just can't generate SFX.
Requirements: Node.js 20+, Windows, NVIDIA RTX GPU (for SFX generation), ffmpeg.
A built-in MCP server (server/mcp-server.mts) lets Claude / Codex drive the running editor (create projects, generate SFX, arrange the timeline). Opening the repo in Claude Code auto-detects .mcp.json; the backend must be running (npm run dev). See the Wiki.
- New project + import — Project ▾ → New; drag a video/audio file into the window.
- Timeline editing — move / trim / split; rename, lock, mute tracks. Selected track is indigo, muted track is amber.
- Add SFX: top 〰 "音效" floating window has three tabs —
- Library (fastest, no engine): pick a category / search → ▶ preview → "+軌" to drop on the timeline, or "素材" to import only.
- Local (本機): point at your own sound folder (enter a path, e.g.
D:\sfx_sample); browsed by auto Chinese categories. Grab free Sonniss packs withscripts/download-sonniss.mjs --dest <folder>(see Wiki). - AI generate — Text → SFX: English prompt (e.g.
glass shatter) → generate (needs the Woosh engine). - Video → V2A: right-click a video clip → generate synced ambience (MMAudio).
- At a position: right-click empty audio track → "建立音效" (Create SFX, nameable).
- Mix — left "效果" (Effects) tab → pick a track → fader / mute; "+ add effect" → EQ / compressor / limiter (click the name to open a floating editor); Main Out LIMIT for a master brickwall limiter.
- Preview — top 🖥 toggles the floating preview window.
- Export — ⬆ Export → (auto-save) → About → Next → settings → Export. The output carries an "Isan 13soul" watermark, bottom-right.
Tips:
Ctrl+Ssaves; hard-refresh withCtrl+Shift+Rafter frontend changes.
| Layer | Tech |
|---|---|
| Frontend | React 18 · Vite 5 · TailwindCSS 3 · Zustand 4 |
| Backend | Node.js 20+ · Express 4 · TypeScript 5 |
| Audio preview | Web Audio API |
| Rendering | ffmpeg (fluent-ffmpeg), NVENC fallback |
| Inference | MMAudio (py3.10) · Woosh (py3.13), each a FastAPI service |
| Native | Rust + napi-rs (waveform / thumbnail / probe) |
| Desktop | Electron + electron-builder |
PolyForm Noncommercial License 1.0.0 — non-commercial use only. See LICENSE. The MMAudio / Sony Woosh weights it relies on are likewise CC BY-NC. This tool bundles and invokes FFmpeg (GPLv3) as a separate executable; third-party licenses and sources are listed in NOTICE.md.
Isan(13soul) — 全端設計工程師 · 影像及音樂創作人 / full-stack design engineer · video & music creator
- 官方網站 / Website · https://www.poofone.com.tw/
- Threads · https://www.threads.com/@isan1314558
- YouTube · https://www.youtube.com/@13Neosoul
- Instagram · https://www.instagram.com/isan1314558/
- Facebook 社團 / Group · https://www.facebook.com/groups/26340062805675868
