MEASURED REVIEW

YouTube Auto-Dubbing Review: What We Measured (2026)

FAMILIAR LABS RESEARCHUPDATED ALL ARTICLES
READ THIS INEnglishEspañolPortuguês中文日本語한국어FrançaisDeutschالعربيةBahasa IndonesiaРусскийItalianoTiếng ViệtไทยNederlandsΕλληνικάУкраїнськаHrvatskiLietuviųSlovenčinaSvenskaБългарскиDanskLatviešu

ABSTRACT

YouTube auto-dubbing generates translated audio tracks for eligible videos, free, inside the platform. We measured its delivered dubs against the source and against Familiar Labs on a frozen 50-pair cohort (Arabic, Spanish, French, Hindi, Korean; August 8, 2026). Three findings: the dub speaks in a stranger's voice, more than a quarter of important ideas change or disappear, and the background scene flattens. Full data: the full YouTube benchmark.

What it is

  • Alternate audio tracks in supported languages for videos on eligible channels; viewers pick the track.
  • Costs nothing, requires no workflow.
  • Availability is the platform's call: in our catalog, 25 of 43 source clips had a usable auto-dub.

What we measured

DUB PAIRS

50

Frozen August 8, 2026

LANGUAGES

5

Arabic, Spanish, French, Hindi, Korean

ALIGNED PAIRS

31

Speaker resemblance

SOURCE IDEAS

528

Meaning review

VOICE
HOW IT IS SCOREDThe voice is not the speaker's. Speaker resemblance 0.141 vs the source; Familiar Labs 0.399 -- 182.9% higher, on the 31 exactly aligned pairs.
MEANING
HOW IT IS SCOREDMeaning drifts. 145 of 528 important source ideas changed or lost (27.5%); Familiar Labs 43 (8.1%). 12 dubs had a critical meaning failure; Familiar Labs 0.
SCENE
HOW IT IS SCOREDThe scene flattens. Background spectral error 11.21 dB vs Familiar Labs' 3.37 dB; laughter and reactions kept their shape at 0.236 correlation vs 0.998.
FIG. 01 · IMPORTANT SOURCE IDEAS CHANGED OR LOST · OF 528 · 50 REVIEWED DUBS
YOUTUBE145
FAMILIAR LABS43
0.141

speaker resemblance to the real speaker -- a stranger's voice

232.6%

more background-sound error than Familiar Labs

25/43

source clips with a usable auto-dub -- Familiar Labs dubbed all 43

When it's the right call

Testing whether translated audiences exist at all: free and zero-effort is the right price.

When the channel is the person (voice, delivery, reactions), a voice + lip-sync dub closes the gap, delivered up to 4K with a true 5.1 or stereo mix.

Limits

Cohort-scoped. A frozen black-box study of the captured cohort, not every YouTube dub, language, or future version.
Predicted naturalness (UTMOS 2.12 vs 1.85) is a model diagnostic with broad uncertainty. CER/WER excluded; the meaning review accepts valid paraphrases.

REFERENCES

  1. [1]Familiar Labs vs YouTube Auto-Dub: full results, method, and listening examples
  2. [2]Familiar Labs vs YouTube Auto-Dubbing: A Frozen-Cohort Study
  3. [3]YouTube Help: Use auto dubbing
  4. [4]All Familiar Labs benchmarks

Multimodal Translation

Familiar Labs builds generationally better dubbing with a Multimodal Localization Foundation Model (MLFM), the first foundation model for multimodal translation: one model decides what to localize across text, audio and visual, and shows you the choices. Your voice and lips, carried into new languages. They're a team of 7 with PhDs, backed by Samsung, RedBird, Gold House. Our mission

See the full benchmark →Talk with the team