MEET LINNET 4
A consistent voice. Clear speech. Controlled timing. Linnet 4 brings these strengths together across English, Spanish, Korean and Portuguese in the new benchmark.
0.604
Korean voice identity
Highest mean; no significant difference from Finch.
Voice consistency
Across four targets. Consecutive-sentence similarity.
0.0–1.6%
Timing overrun
Across four targets. Sentences ending over 0.3s late.
Measured on read speech, October 2026. Strengths vary by language and metric; explore the full comparison below.
PERSO DUBBING · LISTENING ROOM
Hear the delivery.
Choose a scene and switch between the original and the dub.
Nightingale is available on PRO and above.
These are model showcases using different scenes and language pairs. They are separate from the read-speech benchmark below.
Explore the model benchmarks →Each engine brings a different balance of voice, delivery and timing. These profiles summarize the October 2026 read-speech benchmark; compare the full results below.
Premium · Expressive storytelling
Nightingale
A premium engine built for expressive delivery. The new benchmark shows strong translation quality and expression tracking across the four tested target languages.
Highest translation mean in EN, ES and KO. Expression tracking has no significant difference from the best mean in all four targets.
Balanced multilingual delivery
Linnet 4
Consistent voice, low recognition error and controlled timing across four targets. Explore this generation when those qualities matter to your content.
Korean voice identity: 0.604. Timing overrun: 0.0–1.6% across four targets.
Natural delivery · Consistent voice
Oriole
The highest automated English naturalness mean, alongside consistent voice delivery across sentences. Compare it with the other engines in the same statistical group.
English naturalness: 4.11 / 5. Voice consistency: 0.857–0.868 across four targets.
Voice identity · Expression tracking
Finch
Strong similarity to the original speaker, with the highest expression-tracking mean in all four tested target languages.
Highest voice-identity mean in EN, ES and PT. Korean identity is statistically grouped with Linnet 4.
English naturalness · Clear speech
Wren
Strong automated English naturalness and low recognition error. A useful point of comparison when assessing clear, steady delivery.
English naturalness: 4.08 / 5. Recognition error: 1.2–2.6%; Korean uses character error rate.
Previous-generation reference
Linnet 3
Included alongside Linnet 4 so you can compare the two generations under the same updated evaluation. Check each dimension instead of assuming every score moves together.
All six engines were remeasured for the October 2026 benchmark.
Profiles describe measured strengths, not an overall ranking. Model availability can vary; check the options shown in Perso Dubbing.
Benchmark update · 23 October 2026
Compare what matters to your content.
Explore six engines across four target languages. Choose one quality dimension at a time to see the measured strengths and trade-offs.
| Model | 01 | Mean |
|---|---|---|
| Wren | 0.384 | |
| Linnet 3 | 0.382 | |
| Oriole | 0.384 | |
| Linnet 4 | 0.510 | |
| Finch | 0.651 | |
| Nightingale | 0.545 |
● marks the best observed mean. ◇ reproduces the source’s group with no statistically significant difference from the best mean (paired tests, 5% level, multiple-comparison correction). This does not establish equivalence or an overall winner.
Finch has the highest mean in English, Spanish and Portuguese. Linnet 4 has the highest Korean mean, with no significant difference detected from Finch.
Perso engine-team benchmark · 23 October 2026
All six engines were remeasured using the same read-speech evaluation design. The results below replace the earlier benchmark; changes should not be treated as gains against the previous test.
16
Readers in the source dataset
FLEURS recordings in English, Korean, Portuguese and Chinese. Each recording is dubbed into the target languages other than its source language.
51
Dubs per engine
Six engines. Four targets: English, Spanish, Korean and Portuguese. Naturalness is evaluated only for English.
End-to-end evaluation, with context.
Each engine performs speech recognition, translation and voice generation. Outputs use the same evaluation processing and scoring. These results compare the full dubbing pipeline, not isolated text-to-speech models.
Paired comparisons use the same recordings, a 5% significance level and correction for multiple comparisons. A result with no detected significant difference is not proof that two engines are equivalent.
Source: Perso engine-team benchmark, 23 October 2026. Data: Google FLEURS (CC BY 4.0). Evaluators: Microsoft WavLM (CC BY-SA 3.0), UTMOS ’22, Unbabel COMET ’22 and Whisper large-v3. Read-speech results do not cover every production scene; no aggregate ranking is calculated.
What makes Nightingale different?
Nightingale is our premium engine for expressive storytelling. In this benchmark it has the highest translation mean in English, Spanish and Korean, and its expression tracking has no significant difference from the best mean in all four targets. Premium positioning does not mean it leads every metric.
What stands out about Linnet 4?
Linnet 4 combines consistent voice delivery, low recognition error and low timing overrun across the four tested targets. It has the highest Korean voice-identity mean, with no significant difference from Finch. Its English naturalness also belongs to the source’s group with no significant difference from the best mean.
Can the scores predict every dubbing result?
No. This is a read-speech benchmark across four target languages. Naturalness is an automated English-only estimate, and expression measures vocal arousal rather than the full emotional range. Listen to samples and evaluate your own content before choosing a model.
Bring your own video and explore the models available in Perso Dubbing.
Explore Perso Dubbing →