Perso Linnet4
Positive judgments across voice, naturalness and emotion in two scenes.
generated: 3/3 · held judgments: 0The same films. Four dubbing services. Explore real outputs and the evidence behind translation, voice similarity, naturalness and emotion.

Sintel and Cosmos received positive AI judgments for voice similarity, naturalness and emotion preservation. Explore the differences across scenes.
Positive judgments across voice, naturalness and emotion in two scenes.
generated: 3/3 · held judgments: 0Strong findings on Sintel; emotion preservation stood out on Tears of Steel.
generated: 3/3 · held judgments: 2Strong findings on Sintel; delivery and emotion differences observed on Tears of Steel.
generated: 3/3 · held judgments: 2Delivery and emotion differences observed in two completed scenes. One generation failed.
generated: 2/3 · held judgments: 1Compare the original with each first output. Playing one sample pauses the others. Open a film to explore its samples and measurements.

| Scene-level automatic measurements | Role 1 / role 2 cosine | UTMOSv2 predicted mean |
|---|---|---|
| Perso Linnet 4 | 0.408 / 0.302 | 2.61 |
| ElevenLabs Dubbing v2 | 0.470 / 0.270 | 2.23 |
| HeyGen Speed · Audio Only | 0.495 / 0.201 | 2.76 |
| Synthesia AI Dubbing | 0.322 / 0.248 | 2.50 |

| Scene-level automatic measurements | Role 1 / role 2 cosine | UTMOSv2 predicted mean |
|---|---|---|
| Perso Linnet 4 | 0.432 / 0.425 | 2.14 |
| ElevenLabs Dubbing v2 | 0.492 / 0.365 | 2.18 |
| HeyGen Speed · Audio Only | 0.279 / 0.364 † | 2.80 |
| Synthesia AI Dubbing | N/A | N/A |

| Scene-level automatic measurements | Role 1 / role 2 cosine | UTMOSv2 predicted mean |
|---|---|---|
| Perso Linnet 4 | 0.589 / 0.255 † | 2.54 |
| ElevenLabs Dubbing v2 | 0.563 / 0.449 | 2.34 |
| HeyGen Speed · Audio Only | 0.385 / 0.187 | 3.06 |
| Synthesia AI Dubbing | held judgments | 2.92 |
Cosine is not a probability; UTMOSv2 is not human MOS. Background audio and short segments affect these diagnostics. They are not combined into ranking scores. Speaker boundaries may be uncertain.
We distinguish recognized evaluation frameworks, reproducible automatic measurements and AI interpretation. These results are not an independent certification.
Mistranslation, omission and addition categories guided transcript review. AI observations of preserved meaning concern cited examples, not 100% translation accuracy or formal human MQM scores.
Methodology source ↗ECAPA embeddings compare selected corresponding speech segments. UTMOSv2 predicts quality of the full audio mix across five fixed-seed runs. Neither establishes overall dubbing superiority.
Methodology source ↗This international recommendation addresses subjective speech-quality evaluation. Our AI pilot is not a P.808 human study or certification. Bilingual listening review remains a follow-up.
Methodology source ↗Perso · Cosmos: role 2 cosine is 0.255 against the original role and 0.244 against the other role. Interpretation is withheld; this cannot establish voice preservation or a speaker swap.
HeyGen · Tears of Steel: role 1 ends at 13.3 s; two transcriptions place the next speaker at 13.0 s and 13.68 s. The segment may mix speakers.
Perso · Sintel: the AI explanation refers ambiguously to speakers. Its positive voice-similarity label is retained with this warning, not treated as verified speaker identity.
We used the same three English originals, first outputs, no manual translation edits or glossary, and no lip-sync. Perso, ElevenLabs and HeyGen used Mexican Spanish; Synthesia used Spanish without a region choice. This is not a comparison of every provider’s highest-tier model.
Gemini received original and dubbed audio without brand names, video or transcripts. Identical-audio, silence and mismatch controls checked input handling, not general judging validity. Perso’s role, five held judgments and one generation failure are disclosed.
The experiment did not calculate an overall ranking. The editorial order was added on November 18, 2026 for this comparison page. All completed outputs had example-level observations of preserved translation; that does not distinguish a winner. More source films and blinded bilingual review are needed.
Listen to the samples, then explore dubbing with Perso.
Evaluated November 18, 2026 · Perso-authored pilot
Tears of Steel / Sintel: Blender Foundation, CC BY 3.0. Cosmos Laundromat: Blender Foundation, CC BY 4.0.
Sintel · Tears of Steel · Cosmos Laundromat