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
from cosyvoice.utils.file_utils import load_wav
t0 = time.time()
cv = CosyVoice2('pretrained_models/CosyVoice2-0.5B', load_jit=False, load_trt=False, fp16=False)
print('load', round(time.time() - t0, 1), 's', flush=True)
ref = sys.argv[1]
prompt_text = sys.argv[2]
text = sys.argv[3]
out = sys.argv[4]
t0 = time.time(); chunks = []
for r in cv.inference_zero_shot(text, prompt_text, ref, stream=False):
    chunks.append(r['tts_speech'])
wav = torch.cat(chunks, dim=1)
dur = wav.shape[1] / cv.sample_rate
el = time.time() - t0
torchaudio.save(out, wav, cv.sample_rate)
print(f'audio {dur:.1f}s, took {el:.1f}s, RTF {el/dur:.2f}')
