DRAMA-Eval-Results

Benchmark outputs and evaluation results for DRAMA-Eval

Paper: https://arxiv.org/abs/2610.31796
Project: https://github.com/ASLP-lab/DRAMA-Eval
Demo: https://ASLP-lab.github.io/DRAMA-Eval-Demo
Dataset (GT / prompt audio): https://huggingface.co/datasets/ASLP-lab/DRAMA-Eval

This repository stores model-generated audio and evaluation artifacts for the DRAMA-Eval benchmark (context-aware, multi-turn, multi-speaker speech generation). Ground-truth audio and prompt wavs are hosted separately in ASLP-lab/DRAMA-Eval.

Repository Structure

model_audio/
  en/                  # English model generations (per model folder)
  zh/                  # Chinese model generations (per model folder)
objective/
  wer/                 # Textual fidelity (WER/CER) summaries and per-utt results
  sim/                 # Speaker similarity summaries and per-utt results
lalm-res/
  local/               # Local LALM AB tests (B1/B2/B3), run_1..run_3
  global/              # Global LALM AB tests (C1/C2/C3), run_1..run_3

What Each Part Contains

model_audio

Generated speech for each evaluated system, split by language:

• model_audio/en/<ModelName>/
• model_audio/zh/<ModelName>/

Example models: SeedAudio, StepAudio-3-Gen, MiniMax-H3, Qwen3-TTS-VD, FireRedTTS2, FireRedTTS3, VoxCPM2, MiMo-Audio-7B, MOSS-TTSD, SoulX-Podcast, VibeVoice, Higgs-Audio-V2.

objective

Objective metrics aligned with the paper:

• WER / CER under objective/wer/<ModelName>/{en,zh}/
  • results.summary.json — mean WER (and related counts)
  • results.jsonl, manifest.jsonl — per-utterance details
• Speaker similarity (SIM) under objective/sim/<ModelName>/{en,zh}/
  • main_sim.summary.json — mean_of_means and mean_of_pair_stds (reported as mean ± std)
  • main_sim.jsonl — per-utterance details

lalm-res

LLM-as-a-Judge pairwise (A/B) results against ground-truth audio:

• Local (B-level) in lalm-res/local/<ModelName>/{en,zh}/run_*.jsonl
  • B1 Instruction Following
  • B2 Prosody & Fluency
  • B3 Local Expressiveness
• Global (C-level) in lalm-res/global/<ModelName>/{en,zh}/run_*.jsonl
  • C1 Interaction Naturalness
  • C2 Global Flow
  • C3 Scene Immersion

Scores in the paper are tie-adjusted win rates:

(N_win + 0.5 * N_tie) / N_comparisons

usually averaged over three independent runs (mean ± run-to-run std).

Related Releases

• Evaluation code / project: https://github.com/ASLP-lab/DRAMA-Eval
• GT audio and prompt wavs: https://huggingface.co/datasets/ASLP-lab/DRAMA-Eval
• Demo page: https://ASLP-lab.github.io/DRAMA-Eval-Demo

Citation

If you use these results, please cite:

@article{hu2026dramaeval,
  title   = {DRAMA-Eval: A Benchmark for Context-Aware, Multi-Turn, Multi-Speaker Speech Generation},
  author  = {Hu, Jingbin and Yang, Zihang and Peng, Xiaoge and Zhang, Haoyu and Zhang, Ziyu and Zhan, Qirui and Chen, Huakang and Guo, Zhao and Tian, Wenjie and Xia, Kangxiang and Hou, Wenkang and You, Ruonan and Liu, Zihan and Liao, Chenyi and Lyu, Hongyu and Bai, Zihao and Liang, Jinrui and Lin, Shengming and Chen, Yunxiang and Liu, Houdun and Feng, Su and Wu, Bengu and Xue, Liumeng and Xie, Lei},
  journal = {arXiv preprint arXiv:2610.31796},
  year    = {2026}
}

Contact

Audio, Speech and Language Processing Group (ASLP@NPU), Northwestern Polytechnical University / ASLP-lab.