YouLafWhisper
Timestamped transcription with a vocabulary hint for product names.
Overview
Almost no foreign-language tutorials, lectures or product demos exist in Azerbaijani. Simple machine dubbing has two classic faults: the audio drifts away from the picture, and terms like Xamarin or GitHub get read with Azerbaijani letters and sound wrong.
YouLaf generates every sentence separately from Whisper timestamps and places it at its original moment. Terms keep their spelling on screen but are pronounced correctly — Xamarin is spoken as Zamarin.
30.00a 30 s clip dubbed to the second
≤50 sto process a one-minute video
1.6×maximum speech speed-up
2Azerbaijani voices: Banu and Babək
How it works
Whisper → Gemini → TTS → FFmpeg
Gemini
Sentence-by-sentence contextual translation, protected terms, phonetic version.
Edge TTS
Banu and Babək neural voices, retries and a Gemini voice fallback.
FFmpeg
Each sentence at its time, speed-up capped at 1.6×, final MP4.
Challenges & fixes
What broke, how it was fixed
01
Problem
The first prototype produced one audio track for the whole text — no sync
Fix
Sentence-level timing. A 30-second clip comes back at exactly 30.00 s.
02
Problem
Google Translate rate-limited us, a local 3B model mistranslated terms
Fix
Context-aware Gemini translation with protected terms.
03
Problem
"YouLaf" was heard as "You Laugh"
Fix
A vocabulary hint for Whisper.
04
Problem
The free voice API returned HTTP 500 from the cloud
Fix
Retries and an automatic Gemini voice fallback.
05
Problem
YouTube blocked the server's IP
Fix
An upload path, a clear message and a "try with the sample" button.
Stack
Next project — Social network for developers