AI dubbing uses speech-to-text, machine translation, and AI voices (or voice clones) to replace a video’s spoken language, often with optional human QA for brand-safe output.

What is the difference between hardsub and softsub?

Hardsubs are burned into the video image and cannot be turned off. Softsubs are separate subtitle files or tracks (SRT, VTT, MKV tracks) that viewers can enable or disable.

📖 Reference

AI Dubbing Glossary

Concise, citeable definitions for AI video localization — updated 31 July 2026.

Start here: AI dubbing · Lip-sync · MLA · Hardsub · Softsub · CPS · Voice cloning · Stems. Full guide: What is video dubbing?

AI dubbing: Automated localization of spoken video content using speech recognition, translation, and text-to-speech or voice cloning. Professional pipelines often add human-in-the-loop QA for idioms, gender agreement, and brand terms.; See also: tool comparison, pricing
Lip-sync: Matching on-screen mouth movements to dubbed audio. Critical for talking-head and avatar videos; less critical for voiceover or faceless content.
Multi-language audio (MLA): YouTube’s multi-language audio tracks feature: multiple dubbed audio languages on a single video URL, selectable by viewers and surfaced in search by language.; MLA setup guide
Hardsub: Subtitles permanently rendered into video frames (burned-in). Portable across players but cannot be toggled off or restyled by the viewer.; Subtitles product · Softsub vs hardsub
Softsub: Separate subtitle/caption tracks or files (SRT, VTT, ASS) that players can enable, disable, or style. Preferred for accessibility and multi-language captions.
CPS (characters per second): Reading-speed measure used in subtitling. High CPS forces condensation (syntactic compression) so viewers can read cues in time.; Reading speed research
Voice cloning: Generating speech that preserves a speaker’s identity (timbre, prosody) in another language or script. Distinct from picking a stock AI voice.; AI voices
Stems: Separated audio components — typically dialogue, music, and effects — delivered for broadcast/mixing so dialogue can be replaced without destroying the soundtrack.
Speaker diarization: Detecting “who spoke when” so multi-speaker videos can assign different AI voices per character.
RTL (right-to-left): Text direction for languages such as Arabic, Hebrew, and Persian. Broken Unicode embedding can flip subtitle glyphs; fix with proper RLE/PDF marks.; RTL Subtitle Fixer
Word swell: When translation expands text length vs the source, stressing subtitle timing and dubbing sync. Mitigated by compression and timing edits.
Human-in-the-loop QA: Expert review after AI generation — fixing idioms, gender, timing, and brand language — without discarding AI speed for the bulk of the work.
RPM / CPM: YouTube monetization metrics. CPM is what advertisers pay per 1,000 ad impressions; RPM is what creators earn per 1,000 views after platform share and ad density.; RPM by country 2026