Three hours of you. Ten seconds Eleven clips worth posting.

Most of a podcast is throat-clearing.

Veldora scores every second of the transcript — hooks, punchlines, story beats, the bit where someone finally says the thing. Below is a real waveform. The lit segments are what survived.

Spoken in Hindi. Subtitled in English. Whisper handles the speech, NLLB handles the translation, and neither one phones home.

You press Start. This is what happens.

Veldora keeps a Python AI engine running beside the app, holding every model in VRAM. Load the models once, not ten times. Here's the log from the run above.

Why a persistent engine? Whisper, NLLB and pyannote take 10–30 s each to load into VRAM. A tool that shells out per step pays that toll ten times. Veldora pays it once and streams progress back over a socket bound to localhost — a socket that has no route to the internet.

The cheapest way to keep your footage private is to never send it anywhere.

Every cloud clipper starts the same way: upload 4 GB, wait behind strangers, pay monthly, hope the terms of service still say what they said last year. Here is the same job, two ways.

Unplug the ethernet cable mid-render. Nothing happens. That's the feature.

What it eats. What it gives back.

Takes in

A YouTube URL, or MP4 / MKV / MOV / AVI / WEBM off your disk

Gives back

MP4 · 9:16 shorts · 1:1 posts · 16:9 clips — subtitles burned in or left off

Speech

Hindi in, English subtitles out — plus whatever else Whisper already knows

Runs on

Windows 10 / 11 · NVIDIA GPU with CUDA — 8 GB VRAM comfortable, more is faster

First launch

Reads your GPU, VRAM, RAM, cores and free disk, then recommends the model sizes that will actually fit

Ships with

FFmpeg and the runtime, bundled — AI models download separately, once

Also does

Background music beds (podcast · cinematic · motivational · vlog) · project history with re-render

Does not do

Accounts. Telemetry. Watermarks. Monthly invoices. Uploads.