18 Voice Candidates, Tested Before Recommendation
We ran 126 automatic synthesis checks across 18 Kokoro-82M voices, then used a fixed listening rubric to decide which voices deserved a featured position. Eight English voices passed. The tested Chinese voice set did not.
Results at a Glance
How to Read the Scores
Each score is from 1 to 5 in this order: naturalness, pronunciation accuracy, long-sentence stability, long-form consistency, audio integrity, and Chinese-English code-switching. English code-switch scores are diagnostic only because these voices are recommended as English voices. Mixed-language speech is a release gate for Chinese candidates, so cross-language averages would be misleading and are not published.
The listening review was completed by one internal reviewer using headphones in an ordinary room. This is a product quality gate, not an independent laboratory study or a statistically representative preference test.
Candidate Results
| Voice | Locale | Six scores | Gate result | Recommended scenes | Live short preview |
|---|---|---|---|---|---|
| Aoede af_aoede |
en-US | 5 / 5 / 5 / 5 / 5 / 1 | Featured Passed every applicable featured gate |
course, podcast, article | |
| Bella af_bella |
en-US | 5 / 5 / 5 / 5 / 5 / 1 | Featured Passed every applicable featured gate |
article, course, podcast | |
| Kore af_kore |
en-US | 5 / 5 / 5 / 5 / 5 / 1 | Featured Passed every applicable featured gate |
article, course, podcast | |
| Nicole af_nicole |
en-US | 5 / 5 / 5 / 5 / 5 / 1 | Featured Passed every applicable featured gate |
article, podcast | |
| Fenrir am_fenrir |
en-US | 5 / 5 / 5 / 5 / 5 / 1 | Featured Passed every applicable featured gate |
podcast, article, course | |
| Michael am_michael |
en-US | 5 / 5 / 5 / 5 / 5 / 1 | Featured Passed every applicable featured gate |
podcast, video | |
| Emma bf_emma |
en-GB | 5 / 5 / 5 / 5 / 5 / 1 | Featured Passed every applicable featured gate |
course, article, podcast | |
| Puck am_puck |
en-US | 5 / 5 / 5 / 5 / 5 / 1 | Did not pass no use-case score reached 4/5 |
None | |
| Heart af_heart |
en-US | 4 / 5 / 5 / 5 / 5 / 1 | Featured Passed every applicable featured gate |
article | |
| Yunjian zm_yunjian |
zh-CN | 4 / 5 / 5 / 5 / 5 / 1 | Did not pass blocking listening issue, Chinese-English code-switch score below 3/5 |
article, course | |
| Xiaobei zf_xiaobei |
zh-CN | 4 / 4 / 4 / 4 / 4 / 1 | Did not pass blocking listening issue, Chinese-English code-switch score below 3/5 |
podcast | |
| Xiaoni zf_xiaoni |
zh-CN | 4 / 4 / 4 / 4 / 4 / 1 | Did not pass blocking listening issue, Chinese-English code-switch score below 3/5, no use-case score reached 4/5 |
None | |
| Sarah af_sarah |
en-US | 3 / 5 / 5 / 5 / 5 / 1 | Did not pass naturalness score below 4/5 |
article, podcast | |
| Xiaoxiao zf_xiaoxiao |
zh-CN | 3 / 5 / 5 / 5 / 5 / 1 | Did not pass blocking listening issue, naturalness score below 4/5, Chinese-English code-switch score below 3/5 |
article, course, podcast | |
| Yunyang zm_yunyang |
zh-CN | 3 / 4 / 4 / 5 / 5 / 1 | Did not pass blocking listening issue, naturalness score below 4/5, Chinese-English code-switch score below 3/5 |
article | |
| Yunxia zm_yunxia |
zh-CN | 3 / 4 / 4 / 5 / 5 / 1 | Did not pass blocking listening issue, naturalness score below 4/5, Chinese-English code-switch score below 3/5, no use-case score reached 4/5 |
None | |
| Xiaoyi zf_xiaoyi |
zh-CN | 3 / 4 / 4 / 5 / 4 / 1 | Did not pass blocking listening issue, naturalness score below 4/5, Chinese-English code-switch score below 3/5 |
course, podcast, video, article | |
| Yunxi zm_yunxi |
zh-CN | 3 / 4 / 3 / 5 / 5 / 1 | Did not pass blocking listening issue, naturalness score below 4/5, Chinese-English code-switch score below 3/5, no use-case score reached 4/5 |
None |
Live previews use the current production preview sentence and are not the original benchmark recordings.
Why the Chinese Launch Was Paused
All eight tested Chinese voices passed automatic audio-integrity checks, but none met the featured-quality gate. Every candidate failed the Chinese-English mixed-speech requirement, and the listening reports also recorded blocking pronunciation or abbreviation issues. Because the launch condition required at least four featured-quality Chinese voices, the Chinese-market validation phase did not start.
This is a provider-quality result, not a demand result. It does not show that Chinese creators do not want document-to-audio software; it shows that the tested voice set was not strong enough to run a fair demand test.
Limitations and Reproducibility
- One reviewer means the scores should be treated as a release decision, not universal preference.
- Automatic checks verified that audio was generated and structurally valid; they did not prove naturalness.
- The candidate set was deliberately narrow: ten English voices and all eight available Chinese voices.
- Provider, model version, evaluation date, criteria, and public result data are disclosed so later runs can be compared honestly.