"""web/simon 语音批量生成（Qwen3-TTS CustomVoice，Serena 音色）。

生成后写入 audio/voice/wav_tmp/*.wav，由 shell 统一 ffmpeg 转 ogg/m4a。
每条断言：时长 > 0.3s（不足时尾部补静音到 0.35s）、峰值 > 0.05（统一峰值归一 0.8）。
"""
import os
import numpy as np
import soundfile as sf
import torch
from qwen_tts import Qwen3TTSModel

OUT = os.path.join(os.path.dirname(__file__), "..", "audio", "voice", "wav_tmp")
os.makedirs(OUT, exist_ok=True)

LINES = {
    # 开场 + 教学
    "intro": "欢迎来到音序回响。我会用六种声音演奏一段序列，请你记住顺序，再用同样的手势复现出来。先来认识这六种声音。",
    "teach_up": "向上滑，是上行琶音。",
    "teach_down": "向下滑，是低沉的下行音。",
    "teach_left": "向左滑，声音在你的左边。",
    "teach_right": "向右滑，声音在你的右边。",
    "teach_tap": "单击，是短促的鼓点。",
    "teach_doubletap": "双击，是两声连响。",
    "teach_end": "好了，记住这些声音。序列会越来越长，看你能坚持几轮。",
    # 流程提示
    "listen": "听好了。",
    "your_turn": "轮到你了。",
    # 反馈
    "correct": "答对了！",
    "wrong": "错了。",
    "new_record": "新纪录！",
    "gameover": "游戏结束。",
    "replay": "单击屏幕，再来一局。",
    # 播报拼接片段
    "score_pre": "你坚持了",
    "best_pre": "历史最高纪录，",
    "di": "第",
    "lun": "轮",
    "num_shi": "十",
    "num_0": "零", "num_1": "一", "num_2": "二", "num_3": "三", "num_4": "四",
    "num_5": "五", "num_6": "六", "num_7": "七", "num_8": "八", "num_9": "九",
}

model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
    device_map="cuda:0",
    dtype=torch.bfloat16,
)

names = list(LINES.keys())
texts = [LINES[n] for n in names]
wavs, sr = model.generate_custom_voice(
    text=texts,
    language=["Chinese"] * len(names),
    speaker="Serena",
    instruct="用清晰、沉稳而有引导感的声音，语速适中，像一个耐心的游戏主持人",
)

for name, wav in zip(names, wavs):
    wav = np.asarray(wav, dtype=np.float32)
    peak = np.max(np.abs(wav))
    assert peak > 0.05, f"{name} 峰值过小: {peak}"
    wav = wav / peak * 0.8
    dur = len(wav) / sr
    if dur < 0.35:  # 单字片段补静音，保证可听清且满足时长断言
        wav = np.concatenate([wav, np.zeros(int((0.35 - dur) * sr), dtype=np.float32)])
        dur = len(wav) / sr
    assert dur > 0.3, f"{name} 时长过短: {dur}"
    path = os.path.join(OUT, name + ".wav")
    sf.write(path, wav, sr)
    print(f"{name}: {dur:.2f}s peak={np.max(np.abs(wav)):.2f}")
print("done", len(names), "clips")