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)
PA = "所以今天的課程其實非常簡單,就是我們可以查得到系統的版本資訊,CPU的資訊,以及記憶體的大小,IP位置以及硬碟的大小。"
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。"]
def run(tag, fn):
    for i, t in enumerate(S):
        t0 = time.time(); ch = [r['tts_speech'] for r in fn(t)]
        w = torch.cat(ch, dim=1); torchaudio.save(f'out/{tag}-{i}.wav', w, cv.sample_rate)
        print(tag, i, f'{w.shape[1]/cv.sample_rate:.1f}s audio, {time.time()-t0:.0f}s', flush=True)
run('zs-fe', lambda t: cv.inference_zero_shot(t, PA, 'ref-a.wav', stream=False, text_frontend=True))
run('zs-nofe', lambda t: cv.inference_zero_shot(t, PA, 'ref-a.wav', stream=False, text_frontend=False))
run('xl', lambda t: cv.inference_cross_lingual(t, 'ref-a.wav', stream=False, text_frontend=False))
