DRAMA-Eval - Speaker Consistency (Chinese): leaderboard
Speech generators voice whole multi-speaker scenes from structured Chinese scripts built from real podcasts, films and TV series (a set of 537 cases, each turn with delivery directions; systems were scored on up to that many); the score is how similar each character's voice stays across its own turns (speaker-embedding cosine similarity), not similarity to a reference voice as in Seed-TTS-Eval.
Metric: Speaker consistency (-1 to 1): speaker-embedding cosine similarity between the turns of each character in the generated scene, Chinese set (537 cases). Source: aslp-lab.github.io. Saturation forecast: Rough model projection: around 2026. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MOSS-TTS-2.0 Preview | 0.8 | |
| 2 | StepAudio 3 Gen | 0.76 | |
| 3 | Ming-3.0-tts-tiny | 0.74 | |
| 4 | SeedAudio 1.0 | 0.74 | |
| 5 | FireRedTTS2 | 0.73 | |
| 6 | Higgs Audio V2 | 0.7 | |
| 7 | Qwen-Audio-3.1-TTS-Next | 0.7 | |
| 8 | VoxCPM2 | 0.67 | |
| 9 | FireRedTTS3 | 0.62 | |
| 10 | Qwen3-TTS-12Hz-1.7B-VoiceDesign | 0.59 |
Interactive version: theaggregate.ai/benchmark?slug=drama-eval-speaker-consistency-chinese · How It Works · Data refreshed daily, snapshot 2026-11-13.