Wes Roth didn't build another Premiere. He wanted to open a talking-head recording, remove retakes and long pauses by editing its transcript, and export a cut he could publish. In his November 21 video, he builds that narrower tool, calls it Splice and Dice, and moves it to a server so it can keep working after he closes his laptop.
That's the part of the SaaS argument worth taking seriously: you can replace a recurring workflow without recreating the entire product. You may drop an editor subscription, but you still own the storage, fixes, backups and occasional 11 p.m. export failure.
Watch Wes Roth Build the Editor
Source: Wes Roth's November 21, 2026 video and its transcript. Hostinger sponsored the VPS segment. The app and timings are from Roth's demo; the workflow and prompts below are our adaptations. Product documentation was checked on November 21.
From Raw Footage to a Reviewable Cut
| Step | What happens |
|---|---|
| Transcribe | OpenAI's Whisper turns speech into timestamped text. Roth first runs it on his own GPU and later shows it running in the hosted workflow. Running the open-source model on your hardware is different from calling a transcription API. |
| Find cuts | The editor can mark long silences using a duration threshold. For repeated lines and false starts, Roth shows Claude Opus 5.5 reviewing the transcript and identifying likely retakes. That sends text to a model provider even if transcription is self-hosted. |
| Review | Click transcript text to seek, mark or restore a cut, and use the waveform for adjustments the transcript doesn't capture. Compare the edited playback with the original before applying changes. |
| Clean and export | Roth shows basic audio cleanup alongside optional external tools such as Descript Studio Sound and ElevenLabs audio isolation. He exports an edited MP4 and can also export transcripts or data for other editing tools. |
In the footage test, Roth reports a roughly 35-second render for a five-minute source clip at 720p. That's one export, not a measured upload-to-export time or a promise for 4K. For your own test, time upload, transcription, review and rendering separately. A quick encode doesn't help much if you're waiting on a large upload or fixing cut-off words.
Build One Editing Loop Before Adding AI Features
Roth starts with upload and playback in Codex CLI. Then he adds transcript navigation and undo, silence and retake suggestions, audio cleanup, a waveform and multi-clip export. The useful order is playback, precise cuts, recovery and export. Add model-driven suggestions after the basic editing loop works.
Keep original media unchanged and save edits as a list of time ranges, separate from the source. That way you can revise or undo a cut without repeatedly encoding the media. For a new prototype, use a media tool such as FFmpeg for rendering rather than writing an encoder. Then compare export against preview on footage you actually use.
Build the smallest transcript editor that works
What Changes When You Put It on a VPS
Roth uses a Hostinger KVM 2 VPS. The provider lists 2 vCPUs, 8 GB RAM and 100 GB disk for that plan. The demo uses Docker and Traefik for HTTPS, plus a separate login for the editor. Treat those specs as context for Roth's test. Raw recordings, temporary audio, exports and concurrent jobs all change the sizing.
In the demo, Codex connects to the hosting account, then Hostinger's browser terminal agent assists with deployment. The Hostinger MCP docs show why that permission deserves a separate review: the integration can expose account actions, not just read project files. Before connecting anything, confirm what the integration can change. Develop in a local workspace with appropriate permissions and approve infrastructure changes separately.
- Control access: Keep a solo editor private where possible. HTTPS encrypts the connection; authentication and authorization decide who can view, upload or download footage. Check that a signed-out browser cannot reach the app or its file routes.
- Limit uploads and jobs: Keep uploads out of the public web directory, limit file sizes, and cap simultaneous renders. Apply the OWASP upload guidance and keep the host and media dependencies patched.
- Know what leaves the server: Keep API credentials out of chats, source control and build archives. If an external model reviews the transcript or cleans audio, share only what it needs and what you're authorized to send.
- Test restores: Hostinger distinguishes regular backups from short-lived snapshots. A Docker image doesn't contain edit decisions or media stored in mounted volumes. Keep an independent copy and test restoring one real project, including files, cut lists and a playable export.
Test the Failures That Cost an Editor Hours
Roth asks several models to review the same code without making edits. The brief names three failure modes that matter more than aesthetic polish: lost edits, clips exported in the wrong order, and drifting audio. You don't need three models to get started. Use two short clips with obvious visual markers and distinct sounds, restart the editor halfway through, then compare the final MP4 with the saved cut list.
Review for lost edits, clip order and sync
Use branches or test copies for experimental features, and keep a known-good version for the next video. A reviewer can flag a risk; an actual export tells you whether it happened.
When to Replace the Subscription
Build when the work is frequent, the scope is narrow and you can recover without losing a day. Keep paying when your team depends on shared review, broad format support, vendor support or uptime you don't want to manage. A useful middle ground is to self-host the rough cut and hand the result to your existing editor for finishing.
Run both workflows on three real projects before cancelling anything. Track cash costs (hosting, storage, backups and API usage), then compare how much hands-on time each takes, including rework and maintenance. The point isn't to own another small app. It's to cut the time between a recording and a good edit.