seed-tts-eval-subset

A 100-prompt English subset of zhaochenyang20/seed-tts-eval (which is itself a HuggingFace mirror of BytedanceSpeech/seed-tts-eval), used by vllm-omni nightly perf CI for Qwen3-TTS Base voice_clone evaluation.

Layout

en/
├── meta.lst              # 100 lines, 4-field pipe-delimited:
│                         # utt_id|ref_text|prompt_wav_rel|target_text
└── prompt-wavs/          # WAVs referenced by meta.lst[2]

Usage

The bench reads this dataset via --dataset-path <repo_id>. See vllm_omni/benchmarks/data_modules/seed_tts_dataset.py.

vllm bench serve \
    --omni --backend openai-audio-speech \
    --endpoint /v1/audio/speech \
    --dataset-name seed-tts \
    --dataset-path linyueqian/seed-tts-eval-subset \
    --seed-tts-locale en \
    --num-prompts 80 --max-concurrency 8

Selection

100 rows from upstream en/meta.lst after a deterministic shuffle (random.Random(seed=42).shuffle). Each row's referenced WAV under prompt-wavs/ is copied alongside.

License

CC-BY-4.0, inherited from the upstream zhaochenyang20/seed-tts-eval mirror. Reference audio remains under original speakers' rights; benchmarking research use only.