{
  "voice_clone": {
    "default": {
      "language": "en",
      "texts": [
        "Hello from OmniVoice. This benchmark checks that the native runtime produces stable speech across repeated requests.",
        "The final request returns to a medium sentence so we can compare output stability after reuse."
      ]
    },
    "long_short_long": {
      "language": "en",
      "texts": [
        "OmniVoice should keep its graph capacity and memory behavior stable even when a long request is followed by a short request and then another long request. This paragraph is intentionally longer so the runtime has to handle a meaningfully larger text prompt, build the corresponding internal token layout once, and then keep that prepared capacity available for later reuse without falling into a rebuild staircase or wasting extra memory across the same long lived session.",
        "This shorter sentence should reuse the previously prepared runtime state without forcing a fresh rebuild.",
        "We now return to another long request to verify that the reused context still behaves correctly, that the speech stays clear, and that the runtime does not grow memory in a staircase pattern when it revisits a larger prompt after handling a smaller one in between."
      ]
    },
    "short_pair": {
      "language": "en",
      "texts": [
        "The first short request establishes a light prompt path for parity.",
        "The second short request checks repeatability with nearly identical graph pressure."
      ]
    }
  },
  "voice_design": {
    "default": {
      "language": "en",
      "voice_design_instruct": "female, high pitch",
      "texts": [
        "Hello from OmniVoice. This is a voice design parity check without a reference speaker.",
        "The second request keeps the same design instruction so we can compare repeatability within the same session."
      ]
    }
  },
  "auto_voice": {
    "default": {
      "language": "en",
      "texts": [
        "Hello from OmniVoice. This auto voice scenario checks the language-only path without reference audio or style instruction.",
        "The second auto voice request confirms that the same session can keep producing stable speech without clone or design conditioning."
      ]
    },
    "nonverbal_en": {
      "language": "en",
      "whisper_model": "small",
      "texts": [
        "[laughter] You really got me there. I did not see that coming at all.",
        "[sigh] Well, that explains everything. [question-en] Are you serious?"
      ]
    },
    "nonverbal_zh": {
      "language": "zh",
      "whisper_model": "small",
      "whisper_language": "Chinese",
      "asr_compact_lcs_min": 0.65,
      "texts": [
        "[laughter] 真没想到会这样。 [question-en] 你现在明白了吗？",
        "[sigh] 这下总算解释清楚了。 [question-en] 你是不是也同意？"
      ]
    },
    "single_nonverbal_zh": {
      "language": "zh",
      "whisper_model": "small",
      "whisper_language": "Chinese",
      "asr_compact_lcs_min": 0.65,
      "texts": [
        "[sigh] 这下总算解释清楚了。"
      ]
    },
    "single_laughter_zh": {
      "language": "zh",
      "whisper_model": "small",
      "whisper_language": "Chinese",
      "asr_compact_lcs_min": 0.65,
      "texts": [
        "[laughter] 真没想到会这样。"
      ]
    },
    "single_question_en_zh": {
      "language": "zh",
      "whisper_model": "small",
      "whisper_language": "Chinese",
      "asr_compact_lcs_min": 0.65,
      "texts": [
        "[question-en] 你现在明白了吗？"
      ]
    }
  }
}