Spaces:
Sleeping
Sleeping
Create audio_model.py
Browse files- audio_model.py +55 -0
audio_model.py
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoProcessor, BlipForConditionalGeneration, AutoTokenizer,SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
|
2 |
+
import librosa
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
|
6 |
+
#CONSTANTS
|
7 |
+
speaker_embeddings = {
|
8 |
+
"BDL": "spkemb/cmu_us_bdl_arctic-wav-arctic_a0009.npy",
|
9 |
+
"CLB": "spkemb/cmu_us_clb_arctic-wav-arctic_a0144.npy",
|
10 |
+
"RMS": "spkemb/cmu_us_rms_arctic-wav-arctic_b0353.npy",
|
11 |
+
"SLT": "spkemb/cmu_us_slt_arctic-wav-arctic_a0508.npy",
|
12 |
+
}
|
13 |
+
|
14 |
+
# Carga el modelo de clasificación de tetxo a audio speech
|
15 |
+
checkpoint = "microsoft/speecht5_tts"
|
16 |
+
processor = SpeechT5Processor.from_pretrained(checkpoint)
|
17 |
+
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
|
18 |
+
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
|
19 |
+
|
20 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
21 |
+
|
22 |
+
### TEXT TO AUDIO SPEECH MODEL 2
|
23 |
+
# Define la función que convierte texto en voz
|
24 |
+
def text_to_speech(text,speaker):
|
25 |
+
# Genera el audio utilizando el modelo
|
26 |
+
if len(text.strip()) == 0:
|
27 |
+
return (16000, np.zeros(0).astype(np.int16))
|
28 |
+
inputs = processor(text=text, return_tensors="pt")
|
29 |
+
|
30 |
+
# limit input length
|
31 |
+
input_ids = inputs["input_ids"]
|
32 |
+
input_ids = input_ids[..., :model.config.max_text_positions]
|
33 |
+
|
34 |
+
if speaker == "Surprise Me!":
|
35 |
+
# load one of the provided speaker embeddings at random
|
36 |
+
idx = np.random.randint(len(speaker_embeddings))
|
37 |
+
key = list(speaker_embeddings.keys())[idx]
|
38 |
+
speaker_embedding = np.load(speaker_embeddings[key])
|
39 |
+
|
40 |
+
# randomly shuffle the elements
|
41 |
+
np.random.shuffle(speaker_embedding)
|
42 |
+
|
43 |
+
# randomly flip half the values
|
44 |
+
x = (np.random.rand(512) >= 0.5) * 1.0
|
45 |
+
x[x == 0] = -1.0
|
46 |
+
speaker_embedding *= x
|
47 |
+
|
48 |
+
#speaker_embedding = np.random.rand(512).astype(np.float32) * 0.3 - 0.15
|
49 |
+
speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
|
50 |
+
|
51 |
+
speech = model.generate_speech(input_ids, speaker_embedding, vocoder=vocoder)
|
52 |
+
|
53 |
+
speech = (speech.numpy() * 32767).astype(np.int16)
|
54 |
+
return (16000, speech)
|
55 |
+
### END TEXT TO AUDIO SPEECH MODEL 2
|