My live dubber waits a sentence before it talks. That's not a bug, it's the honest one.

Description and use

LiveDub embraces the natural sequential delay in real‑time dubbing, creating a breathing space that improves comprehension and authenticity.

Problem

Real‑time dubbing demos often cut silence, creating unnatural pauses that mislead users about latency and quality.

Outcome

LiveDub delivers a realistic dubbing experience by embracing the natural delay of sequential processing, resulting in a breathing space that enhances comprehension and authenticity.

Workflow

• 1. Capture audio from the tab using Whisper for transcription.
• 2. Feed the full transcribed sentence to a local LLM in LM Studio for translation.
• 3. Pass the translated text to Kokoro TTS to synthesize the dub voice.
• 4. Mix the original audio and the synthesized voice, ducking the original to create a breathing space.
• 5. Play the combined audio in real time on a PC running Python 3.10 (no GPU required).

Evidence

My local dubber waits until the speaker finishes before it translates. Which means you hear the original line in the language you don't understand, a short gap, then the dub.

Source

The whole thing runs on my PC on Python 3.10, no GPU required (it helps, but it's not the gate).

Source