Alonso Cárdenas
f2b43a4bb6
Add model folders with early instructions and a main.py file for testing. Add output folder for wav files generated. Add asset folder with manito.wav inside. Add env.example. Update readme.md for clear instructions to follow.
42 lines
1.2 KiB
Python
42 lines
1.2 KiB
Python
import time
|
|
import numpy as np
|
|
import soundfile as sf
|
|
import sounddevice as sd
|
|
import torch
|
|
from qwen_tts import Qwen3TTSModel
|
|
|
|
model = Qwen3TTSModel.from_pretrained(
|
|
"Qwen/Qwen3-TTS-12Hz-0.6B-Base",
|
|
device_map="cpu",
|
|
dtype=torch.float32,
|
|
)
|
|
|
|
prompt = model.create_voice_clone_prompt(
|
|
ref_audio="../../assets/manito.wav",
|
|
ref_text="¡Hola! Me llamo Manito y me encanta aprender contigo. ¿Jugamos un rato? Vamos a descubrir colores, formas y números juntos.",
|
|
)
|
|
|
|
texto = "¡Hola! Me llamo Manito y me encanta aprender contigo. ¿Jugamos un rato? Vamos a descubrir colores, formas y números juntos."
|
|
t0 = time.time()
|
|
primero = None
|
|
chunks = []
|
|
|
|
for chunk, sr in model.stream_generate_voice_clone(
|
|
text=texto,
|
|
language="Spanish",
|
|
voice_clone_prompt=prompt,
|
|
emit_every_frames=4,
|
|
decode_window_frames=80,
|
|
overlap_samples=0,
|
|
):
|
|
if primero is None:
|
|
primero = time.time() - t0
|
|
chunks.append(np.asarray(chunk, dtype=np.float32))
|
|
|
|
total = time.time() - t0
|
|
audio = np.concatenate(chunks)
|
|
dur = len(audio) / sr
|
|
sf.write("../../outputs/qwen3-tts-6b/qwen_stream.wav", audio, sr)
|
|
print(f"primer chunk {primero:.2f}s | total {total:.2f}s | audio {dur:.2f}s | RTF {total/dur:.2f} | chunks {len(chunks)}")
|
|
sd.play(audio, sr)
|
|
sd.wait() |