How Vocelle is built

Measured on real jobs

• About 1 min

  A 25 second clip into two languages

  4 jobs, 58 to 64 seconds

• About 2½ min

  A one minute clip into four languages

  3 jobs, 134 to 152 seconds

• 49 s

  Processing per minute of video, per language

  Median of 14 jobs

• 30 min

  Longest video one job takes

  Files up to 2 GB

From the records of 14 dubs run through the live pipeline in September 2026, on clips 15 to 60 seconds long. Times run from submitting the job to its files being ready.

AI is the engine

How the AI decides

• What the AI hears

  The original soundtrack, word by word. The speech model gives every word a timestamp, and the words are grouped into subtitle length lines at sentence ends and pauses.

• What it produces

  A timed transcript, a translation of every line in every target language, a neural voice track per voiced language, and SRT and VTT subtitles for all of them.

• How it decides

  The translation model keeps each line aligned to its source line and close to its length. The timing engine places each voiced line at its original start and speeds it up by at most 1.45 times to fit.

• Why it gets better

  Every job records, line by line, what needed a speed up or reads too fast, and which engine made each language. That record shows where a language needs a better engine, and the engine layer switches a language with a setting.

The pipeline, service by service

From upload to download

1. 01

   Upload or link

   Cloud Storage

   Files go straight from the browser to storage in a resumable upload. A pasted link is fetched by the worker, which first checks that it points to a public address.

2. 02

   API

   Cloud Run service

   Checks your sign in, opens the upload, writes the job record and starts a worker for that video.

3. 03

   Worker

   Cloud Run job, 2 vCPU, 4 GB

   One worker per video. It writes progress to the job record as it goes, so your workspace updates live.

4. 04

   Speech recognition

   Chirp 2, Speech-to-Text v2, Europe

   Transcribes the original audio with a timestamp on every word, then cuts it into subtitle length lines at sentence ends and pauses.

5. 05

   Translation

   Gemini 2.5 Flash on Vertex AI

   Translates 40 lines per request, JSON in and out, retried until every line comes back. Glossary terms stay exactly as written.

6. 06

   Voices

   Cloud Text-to-Speech, Chirp 3 HD

   The newest voice family available for each language, four lines at a time. Languages without a voice are delivered as subtitles.

7. 07

   Timing and render

   Timing engine and FFmpeg

   Each line starts where the original did and is sped up by at most 1.45 times. The new track is laid on your original picture without re-encoding when the codec allows.

8. 08

   Quality check and delivery

   Cloud Storage, EU

   Lines that run long or read too fast are flagged on the job. Files are downloaded through links that expire after 6 hours.

Engines

Built so the engines can change

Translation and voices sit behind one interface, so a language can move to a better engine with a setting. If an engine fails, the job falls back to the other one instead of stopping.

Formats

What goes in, what comes out

• Video files

  MP4 or MOV, up to 2 GB and 30 minutes

  Live

• Links

  A link to a video you own

  Live

• Dubbed video

  MP4 per voiced language, original picture kept

  Live

• Captioned video

  MP4 with burned in subtitles per subtitled language

  Live

• Subtitle files

  SRT and VTT for every language, plus the source transcript

  Live

• API

  Create jobs, follow progress and download results from your own systems

  Planned