ai-agents-speaker-embedding-pe-user-wants-to-reuse-a-random-speaker-embedding-acr-7f3448ef

THEN Generate a speaker embedding using chat.sample_random_speaker(), then save it with torch.save() to a .pth file. For later use, load the embedding with torch.load() and pass it as spk_emb in the inference params. This preserves the exact voice timbre.

# Save
rand_spk = chat.sample_random_speaker()
torch.save(rand_spk, 'speaker/my_voice.pth')
# Load
rand_spk = torch.load('speaker/my_voice.pth')
params_infer_code = {'prompt':'[speed_5]', 'temperature':0.3, 'spk_emb': rand_spk}
wavs = chat.infer(text, use_decoder=True, params_infer_code=params_infer_code)