kosw/instructtts-qwen-eval

InstructTTSEval Qwen3-TTS Evaluation Set

A ready-to-use evaluation slice of InstructTTSEval: the first 30 English samples, with TTS audio generated by Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign and instruction-following judgments produced by Qwen/Qwen2.5-Omni-3B as an audio judge.

Each of the 30 source samples has three style-instruction variants, giving 90 rows (one row per sample × instruction type):

Column	Type	Description
id	string	Source sample id in InstructTTSEval (e.g. en_0)
instruction_type	string	APS (Acoustic-Parameter Specification), DSD (Descriptive-Style Directive), or RP (Role-Play scenario)
text	string	Text to synthesize
instruction	string	Natural-language style instruction
audio	Audio (24 kHz)	Speech generated by Qwen3-TTS-12Hz-1.7B-VoiceDesign for this (text, instruction) pair
qwen_score	bool	Whether the Qwen2.5-Omni-3B judge deemed the audio consistent with the instruction (may be null if judging failed)

Baseline results (Qwen3-TTS-12Hz-1.7B-VoiceDesign, judged by Qwen2.5-Omni-3B)

Instruction type	Consistent	Rate
APS	28/30	93.3%
DSD	23/30	76.7%
RP	17/30	56.7%

Macro average: 75.6%

Usage

Evaluate your own TTS model

Synthesize one wav per dataset row, named {id}_{instruction_type}.wav (e.g. en_0_APS.wav, ..., 90 files total), then judge them (needs a CUDA GPU, ~8 GB VRAM). Instructions and the judge prompt are fetched from this repo automatically.

Option A — pip package (UTMOSv2-style):

pip install git+https://github.com/Kosw/InstructTTSEval.git@qwen-omni-eval

instructtts-eval --wav_dir my_gen_wavs --resume

import instructtts_eval
out = instructtts_eval.judge(wav_dir="my_gen_wavs", resume=True)
print(out["stats"]["macro_avg"])

Faster judging via a vLLM server (continuous batching; Linux): vllm serve Qwen/Qwen2.5-Omni-3B --port 8000, then add --base_url http://localhost:8000/v1 --num_workers 8 (or judge(..., base_url=...)). Without base_url the judge runs locally via transformers.

Option B — standalone script (no package install, same code):

hf download kosw/instructtts-qwen-eval judge.py --repo-type dataset --local-dir .
pip install transformers accelerate datasets soundfile scipy tqdm   # plus torch matching your CUDA
python judge.py --wav_dir my_gen_wavs --resume

Per-type consistency rates (APS/DSD/RP + macro average) are printed at the end; per-sample verdicts land in my_gen_wavs_score.jsonl.

Load the dataset directly

from datasets import load_dataset

ds = load_dataset("kosw/instructtts-qwen-eval", split="test")
print(ds[0]["instruction_type"], ds[0]["instruction"][:80])
# ds[0]["audio"] -> generated reference audio (24 kHz)

Attribution

Source texts and instructions come from CaasiHUANG/InstructTTSEval (MIT license). Please cite the original paper:

@misc{huang2025instructttsevalbenchmarkingcomplexnaturallanguage,
      title={InstructTTSEval: Benchmarking Complex Natural-Language Instruction Following in Text-to-Speech Systems},
      author={Kexin Huang and Qian Tu and Liwei Fan and Chenchen Yang and Dong Zhang and Shimin Li and Zhaoye Fei and Qinyuan Cheng and Xipeng Qiu},
      year={2025},
      eprint={2506.16381},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.16381},
}