Dev Blog: How a Turkish AI Video Pipeline Wrested Audio Control from the Model

A developer building AI-generated character videos in Turkish discovered that the video service's built-in speech synthesis was unreliable, mangling words, repeating phrases, and mispronouncing foreign terms. To fix this, the team removed audio generation from the video model entirely and replaced it with Microsoft's edge-tts, a free, quota-free Turkish TTS engine, tuning speech rate to -8% to pass a Whisper-based quality gate requiring word confidence above 0.80. Lip-sync was handled by Wav2Lip, but its low-resolution mouth output looked blurry on high-resolution faces, so GFPGAN face restoration was applied selectively only to the mouth region using a frame-difference mask to prevent flickering elsewhere. When the audio clip ran longer than the base video, the team used ffmpeg to extend footage with a slow-down and boomerang loop rather than letting Wav2Lip create a visible repeat cut. The resulting pipeline — edge-tts, Whisper gate, Wav2Lip, GFPGAN, and ffmpeg — produced publishable lip-synced Turkish character videos without burning commercial TTS character quotas.