"""Qwen3-TTS CustomVoice 冒烟测试：预置音色 + 情感指令生成中文对白。"""
import os
import time

os.environ.setdefault("HF_HOME", "/home/ubuntu/SoundGame/models/hf")
os.environ.setdefault("HF_ENDPOINT", "https://hf-mirror.com")

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

t0 = time.time()
model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    # 未安装 flash-attn，使用默认注意力实现
)
print(f"model loaded in {time.time() - t0:.1f}s")
print("speakers:", model.get_supported_speakers())
print("languages:", model.get_supported_languages())

wavs, sr = model.generate_custom_voice(
    text="快躲开！它来了！",
    language="Chinese",
    speaker="Vivian",
    instruct="用惊恐、急促的语气大声喊出来",
)

out = "/home/ubuntu/SoundGame/tools/samples/tts_test.wav"
sf.write(out, wavs[0], sr)
dur = len(wavs[0]) / sr
print(f"saved: {out}  sr={sr}  duration={dur:.2f}s  samples={len(wavs[0])}")
assert dur > 0.3, "audio too short, something went wrong"