DeepVideo 2.0 Beta: A Complete Rebuild of Video Translation & Dubbing

🎓 Expert Video Translation Pro Workflow

A full-pipeline translation workflow for professional users: a 10-step pipeline with advanced translation options — set translation style, target industry, and glossary; native multi-speaker mode and emotion recognition, plus pause-after-translation review so you can review before synthesis continues.

⬇️ Video Download Free to UseBuilt-in yt-dlp

Built-in yt-dlp tool for one-click downloads from major platforms, with the option to create a translation project right after downloading — download and translation closed-loop in one app, no third-party tools needed. This feature is free to use — no subscription required.

🖼️ Image Watermark Removal Free to UseLaMa Inpainting

AI automatically detects watermark locations and removes them with the LaMa inpainting algorithm, with batch processing support — import many images at once, export in one go. This feature is free to use — no subscription required.

🎬 Video Watermark Removal Free to UseFrame-by-Frame AI Erasing

Smart detection of hardcoded subtitles and watermarks with frame-by-frame AI erasing: the new GrowPolicy reduces accidental damage by 61–87%, validated on a corpus of 10 videos / 118 frames. This feature is free to use — no subscription required.

🎙️ SenseVoice Speech Recognition Recognition Engine

The brand-new SenseVoice engine is now live. For casual Chinese speech, noisy environments, phone recordings and more, its recognition accuracy far exceeds Whisper:

• Completely fixes Whisper’s homophone errors — cases like mis-hearing that no bigger model could ever fix are gone for good
• Cantonese support — where Whisper forces Cantonese into Mandarin, SenseVoice outputs authentic Cantonese text
• Smart routing: Chinese, Cantonese and Japanese automatically go to SenseVoice; other languages go to Whisper
• Word-level timestamps for more precise subtitle alignment

🗣️ Chatterbox Voice Cloning

The new Chatterbox voice cloning engine is fully integrated into the production pipeline:

🎶 RoFormer Vocal Separation New Default Engine

RoFormer replaces Demucs as the default vocal separation engine, validated by A/B blind testing on 19 audio clips: RoFormer 14 wins / 4 losses vs Demucs 11 wins / 7 losses.

• Fixes female vocals being cut off and voiceovers being swallowed
• Significantly less residual vocals in background music
• Streaming processing cuts memory for 45-minute audio from 3.4GB to 70MB (48×)

👥 358 System Voice Characters Voice Library

A brand-new system voice library covering 10 languages (Chinese, English, Japanese, Korean, German, French, Spanish, Italian, Portuguese, Russian) and 8 categories (characters, narration, business, education, entertainment, news, stories, animals), with a built-in smart assignment algorithm that matches the most suitable voice based on speaker gender.

⚡ One-Click Translation Minimal Workflow

The brand-new one-click translation mode compresses the entire pipeline into a single step:

1. Select a video, click translate
2. Automatically: audio extraction → vocal separation → speech recognition → translation → speech synthesis → muxing
3. After translation, go straight to the preview page with original and translated videos side by side
4. Need fine-tuning? Jump into the subtitle editor with one click

✏️ Editor Upgrades Editor

• Per-segment TTS preview — no need to regenerate the whole audio track; click and listen per segment
• Smart edit journal — only writes back changed settings, never overwrites unmodified ones
• SRT import/export — supports SRT files with speaker info
• LLM source-subtitle correction — automatically fixes speech recognition errors (original_corrected.srt)
• Font size control — adjust subtitle styling in real time

🧽 Hard-Subtitle Removal Improvements Erasing Algorithm

Major improvements to the hard-subtitle erasing algorithm: the new GrowPolicy reduces accidental damage by 61–87%; the mask retention cap leaves no residue after subtitles leave the erasing zone; exit frames no longer freeze the background. Validated on a corpus of 10 videos / 118 frames plus 363-frame end-to-end tests.

👥 Speaker Diarization Improvements Diarization

• Dual-metric merging — weak clusters use group cosine ≥ 0.35, strong clusters use pairwise centroid cosine ≥ 0.45, fixing different speakers being fused together
• Iterative splitting at every speaker boundary, not just the first
• Ignores sub-1-second bursts, eliminating ghost speakers

🍎 Mac App Store in Preparation App Store

DeepVideo 2.0 is being prepared for Mac App Store release: fully sandbox-compatible builds, security-scoped bookmarks persisting file access across restarts, legacy data migration shrunk from ~32GB to ~4MB, and the privacy manifest (PrivacyInfo.xcprivacy) already included.