Crisp

Crisp is a local transcript-based audio and video editor prototype. It lets a user import media, transcribe it, select words in the transcript, preview edits, and render a cleaned audio/video copy.

The current prototype focuses on filler-word cleanup and timestamp-based editing.

Features

• Import audio or video.
• Extract audio from video for transcription.
• Transcribe media with Groq Whisper.
• Display word-level transcript timestamps.
• Highlight filler candidates such as um, uh, ah, erm, and hmm.
• Select transcript words.
• Preview selected edits.
• Render selected edits as:
  • Silence: mute the selected interval while preserving duration.
  • Remove: cut the selected interval and close the gap.
• Preserve the original upload until final download cleanup.

Tech Stack

• Frontend: React + Vite + TypeScript.
• Backend: Node.js + Express + TypeScript.
• Transcription: Groq Audio Transcriptions API.
• Media processing: FFmpeg and FFprobe.

Requirements

• Node.js
• npm
• FFmpeg available on PATH
• FFprobe available on PATH
• A Groq API key (create one at console.groq.com) — transcription fails without it. The landing page checks for it on load and tells you if it is missing.

Environment

Create a .env file in the project root:

GROQ_API_KEY=your_groq_api_key_here
PORT=8787

The app reads .env through dotenv.

Install

npm install

Run

npm run dev

This starts:

• API: http://localhost:8787
• Web app: http://localhost:5173

Build

npm run build

Project Structure

DESIGN.md
  Crisp visual identity used by the app and demo assets.

server/index.ts
  Express API, upload handling, Groq transcription, FFmpeg rendering.

src/App.tsx
  Main React editor UI.

src/BoundaryEditor.tsx
  Waveform panel for adjusting each cut by hand.

src/languages.ts
  Languages supported by Whisper large-v3, listed on the landing page.

src/styles.css
  Application styling.

Example Audio/
  Optional sample audio file.

demo/
  Public demo poster, demo video, and HyperFrames source composition.

data/uploads/
  Uploaded originals. Ignored by git.

data/proxies/
  Temporary transcription audio proxies and chunks. Ignored by git.

data/exports/
  Preview and final rendered files. Ignored by git.

Current Workflow

1. Import an audio or video file.
2. Crisp stores the original upload.
3. For video, Crisp extracts/normalizes an audio proxy for transcription.
4. Crisp transcribes the media.
5. The transcript appears with word timestamps.
6. Filler candidates are highlighted.
7. User selects words.
8. User adjusts each cut's boundaries on the waveform if needed.
9. User chooses Silence or Remove (audio supports Remove only).
10. User previews the selected edit.
11. User prepares a final download.

Rendering Model

Crisp uses non-destructive edit operations:

{
  "id": "operation_001",
  "type": "ripple_delete",
  "source_start": 7.54,
  "source_end": 8.021,
  "word_ids": ["word_0024"],
  "boundary_source": "user_adjusted"
}

The original transcript timestamps are kept as source-media timestamps. Rendering is done from the original media file into a generated copy.

Current render paths share the same backend function:

/api/projects/:id/preview
/api/projects/:id/render
/api/projects/:id/finalize

Important Limitation

Word timestamps are not always safe cut boundaries.

ASR can mark a filler word in a way that overlaps nearby real speech. If Crisp removes that interval exactly, it can remove part of a real word. This is why every cut is adjustable by hand before rendering.

Example:

participating, um, in

The provided timestamp may mark um too early, causing a cut to remove part of participating.

This is the main known product limitation. Internal timestamp/debug notes are kept out of the public repository.

Transcription

Whisper cleans up speech by default and omits filler words entirely, folding them into the neighbouring word's timing. Crisp counters this by passing a prompt of deliberately disfluent speech, which acts as a style prior rather than an instruction — demonstrating the style works where describing it ("keep every um") does not.

Measured on the bundled samples: zero fillers without the prompt, all three with it, identical across repeated runs. It also improves the surrounding transcript, recovering the repeated access access and Ellen (unprompted runs produced Alan).

Set WHISPER_PROMPT in .env to tune it for a different speaker or language.

Whisper large-v3 transcribes 100 languages (the full list is on the landing page), though accuracy varies considerably between them. The default prompt above is English, so for another language set WHISPER_PROMPT to a sample of disfluent speech in that language or fillers will go undetected.

Boundary Adjustment

Because word timestamps are only ever an approximation, every selected word can be adjusted by hand before rendering. Selecting a word opens an editor showing the real waveform around it:

• Drag either edge of the cut, or nudge with ← / → (10ms, or 50ms with Shift).
• Dashed guides mark where the neighbouring words start and end.
• Play selection plays exactly what sits between the handles, so you hear what is about to be cut — if part of a real word is audible, drag that edge in.
• Reset returns to the transcript's own range.

Hand-placed boundaries are stored separately from the transcript and are tagged boundary_source: "user_adjusted". The renderer applies those intervals exactly as given and skips its automatic boundary refinement, so an adjusted cut is never silently second-guessed.

Recommended Next Work

• Consider dropping the heuristic acoustic detector now that prompted transcription recovers fillers directly (see Transcription above).
• Add safer defaults for video, where Silence is usually less destructive than Remove.

License

Crisp is released under the MIT License.

Copyright (c) 2026 Alpha AI.

The license notice must be preserved in copies or substantial portions of the software.

Contributing

Crisp is maintained by Alpha AI. To contribute, extend the editor, or discuss collaboration, contact:

corporate@alphaai.biz

Notes

• The product should never automatically delete words without user confirmation.
• Preview files are generated copies under data/exports.
• Originals remain under data/uploads until final cleanup.
• The current project state is in memory; restarting the API loses active project objects, though files remain on disk.