DRAMA-Eval - Interaction Naturalness (English): leaderboard
Speech generators voice whole multi-speaker scenes from structured English scripts built from real podcasts, films and TV series (a set of 501 cases, each turn with delivery directions; systems were scored on up to that many); a Gemini 3.8 Flash judge compares the turn-taking and interaction between speakers with the real recording of the whole scene, and the score is a tie-adjusted win rate where 50 means as good as the recording. It judges the scene, where EmergentTTS-Eval judges single utterances.
Metric: Tie-adjusted win rate (%) against the real recording on turn-taking and interaction between the speakers, Gemini 3.8 Flash A/B judge (50 = parity), English set (501 cases). Source: aslp-lab.github.io. Saturation forecast: Rough model projection: around 2026. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | SeedAudio 1.0 | 56.8 | |
| 2 | StepAudio 3 Gen | 43.4 | |
| 3 | Qwen-Audio-3.1-TTS-Next | 42.79 | |
| 4 | MiniMax-H3 | 37.8 | |
| 5 | Ming-3.0-tts-0.5B | 37.6 | |
| 6 | Qwen3-TTS-12Hz-1.7B-VoiceDesign | 34.3 | |
| 7 | MOSS-TTS-2.0 Preview | 34.2 | |
| 8 | FireRedTTS3 | 30.9 | |
| 9 | VoxCPM2 | 28.9 | |
| 10 | Higgs Audio V2 | 26.4 |
Interactive version: theaggregate.ai/benchmark?slug=drama-eval-interaction-naturalness-english · How It Works · Data refreshed daily, snapshot 2026-11-13.