import sys, time, torch, torchaudio
sys.path.insert(0, 'CosyVoice'); sys.path.insert(0, 'CosyVoice/third_party/Matcha-TTS')
torch.set_num_threads(12)
from cosyvoice.cli.cosyvoice import CosyVoice2
cv = CosyVoice2('pretrained_models/CosyVoice2-0.5B', load_jit=False, load_trt=False, fp16=False)
REFS = {'hc': ("大家好,今天我們要學習如何使用終端機,查詢主機的基本資訊,包括處理器和記憶體的大小。", 'ref-hc.wav')}
S = ["大家好,我是方舟老師,歡迎來到這門課。", "機器人其實就是三件事,感測、思考、行動。", "通電之前,一定要先確認電池的接線。"]
for tag, (pt, wav) in REFS.items():
    for i, t in enumerate(S):
        t0 = time.time(); ch = [r['tts_speech'] for r in cv.inference_zero_shot(t, pt, wav, stream=False, text_frontend=True)]
        w = torch.cat(ch, dim=1); torchaudio.save(f'out/r{tag}-{i}.wav', w, cv.sample_rate)
        print(tag, i, f'{w.shape[1]/cv.sample_rate:.1f}s, {time.time()-t0:.0f}s', flush=True)
