Descript

Converts uploaded video or audio files into editable text transcripts using multi-language speech recognition. The system detects and labels up to 8+ distinct speakers automatically, supporting 25 languages. Transcription output is synchronized with video timeline, enabling text-based editing that maps back to media segments. Processing occurs server-side in the cloud with latency described as 'in moments' (specific SLA unknown).

Core editing engine that maps text transcript edits back to video/audio output. When a user deletes, modifies, or reorders text in the transcript, the system automatically re-renders the corresponding video segments, removing or adjusting audio/video timing to match. This requires frame-accurate synchronization between transcript tokens and media segments, likely using alignment metadata generated during transcription. Regeneration consumes AI credits and processes asynchronously (latency unknown).

One-click automation that applies professional formatting, scene composition, and layout to existing video. System analyzes video content, automatically inserts B-roll, applies transitions, adjusts pacing, and applies consistent styling (fonts, colors, animations). Quick Design generates multiple formatted variations that users can choose from. Processing consumes AI credits and generates new video variants without modifying original.

Agentic AI system that interprets natural language editing instructions and applies corresponding video edits automatically. Users describe desired edits in plain English (e.g., 'remove the pause after the first sentence', 'make the intro 5 seconds shorter', 'add B-roll to the second paragraph'), and Underlord parses instructions, identifies relevant video segments, and applies edits. Underlord has limited access on Free tier and full access on Creator tier+. Operates asynchronously and consumes AI credits.

System tracks media consumption (video/audio duration uploaded and processed) against monthly per-user quotas. Free tier: 1 hour/month; Hobbyist: 10 hours/month; Creator: 30 hours/month; Business: 40 hours/month. Quotas reset monthly. When quota is exceeded, users must upgrade tier or purchase top-up minutes (pricing unknown). Consumption is tracked per user and per project. Dashboard displays remaining quota and usage breakdown.

Consumption-based credit system where different AI features (voice cloning, B-roll generation, eye contact correction, etc.) consume different amounts of credits. Monthly credit allowances: Free: 100 credits; Hobbyist: 400 credits; Creator: 800 credits; Business: 1500 credits. Credits reset monthly. When credits are depleted, users must upgrade tier or purchase top-up credits (pricing unknown). Consumption rates per operation are not documented, creating unpredictable usage patterns.

Enables multiple users to work on the same project simultaneously. Users can share projects, assign roles (editor, viewer, commenter unknown), and see real-time changes. Collaboration is limited by tier: Creator tier supports 3 users; Business tier supports 5 users; Enterprise supports unlimited users. Shared projects have shared media hour and AI credit quotas (quota sharing model unknown). Real-time synchronization and conflict resolution mechanisms unknown.

Built-in screen recording tool that captures screen, audio, and optional webcam video. Recordings are automatically transcribed and imported into Descript project for editing. Users can record tutorials, presentations, or demos without external recording software. Recordings are stored in project and consume media hour quota. Screen recording quality and resolution unknown.

Detects and removes common filler words ('um', 'uh', 'like', 'you know', 'basically', etc.) from video/audio by identifying them in the transcript and triggering automatic regeneration. The system likely uses a predefined filler word dictionary and removes matching tokens from the transcript, then re-renders video to remove the corresponding audio segments. No user control over which fillers to remove; fully automated with no preview.

Analyzes video frames to detect faces and eye gaze direction, then synthesizes corrected eye contact by adjusting gaze to face the camera. Uses computer vision for face detection and likely generative AI (unknown model) to synthesize eye movement and pupil position. Operates on video segments where faces are detected; fails silently on frames without detectable faces. Correction is applied during video regeneration and consumes AI credits.

Applies AI-powered audio processing to reduce background noise, enhance voice clarity, and improve overall audio quality without requiring professional microphones or soundproofing. Uses 'regenerative AI' (specific model unknown) to analyze audio spectrograms, identify noise patterns, and synthesize clean voice audio. Processing is non-destructive and applied during video regeneration. Consumes AI credits and operates on entire audio tracks (no selective application).

Records a sample of a user's voice, creates a digital voice clone, and regenerates video/audio with the cloned voice speaking new text. The system uses speaker embedding and voice conversion techniques to match the original voice characteristics, then synthesizes mouth movements to match the new speech using video generation (model unknown). Cloned voices are stored in Descript and cannot be exported. Regeneration consumes AI credits and processes asynchronously.

Generates video clips (B-roll) that match content context or user-provided prompts using generative video models (specific model unknown; claims 'latest AI models'). Users can select from predefined styles or provide custom prompts describing desired B-roll. Generated clips are inserted into the timeline to supplement talking-head footage. Generation consumes AI credits and processes asynchronously. Generated clips are stored in project and can be customized (trim, speed, effects unknown).

Automatically generates captions/subtitles from transcript and applies dynamic styling, animations, and branding. Captions are synchronized to video timeline and can be customized with fonts, colors, animations, and positioning. System supports multiple caption styles (burned-in, overlay, separate track) and export formats (SRT, VTT unknown). Captions are accessibility-focused and can be toggled on/off in exported video.

Translates video content into 30+ languages and generates dubbed audio in target language while maintaining speaker voice characteristics. System uses machine translation (model unknown) to translate transcript, then synthesizes speech in target language with mouth movement regeneration. Translations are marked as 'proofread' (implies human review, but process unknown). Dubbing consumes AI credits and processes asynchronously. Available on Business tier+ only.

Generates talking-head videos with AI avatars speaking provided text. Users can select from a gallery of predefined avatars (Creator tier+) or create custom avatars from their own photos (Business tier+). System synthesizes speech in avatar's voice and generates lip-sync mouth movements. Generated videos can be customized with backgrounds, clothing, and gestures (customization depth unknown). Avatar videos are stored in project and can be edited like regular video.