Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -6,7 +6,7 @@ import torch
|
|
6 |
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
|
7 |
|
8 |
|
9 |
-
checkpoint = "
|
10 |
processor = SpeechT5Processor.from_pretrained(checkpoint)
|
11 |
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
|
12 |
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
|
@@ -31,23 +31,7 @@ def predict(text, speaker):
|
|
31 |
input_ids = inputs["input_ids"]
|
32 |
input_ids = input_ids[..., :model.config.max_text_positions]
|
33 |
|
34 |
-
|
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 |
-
else:
|
50 |
-
speaker_embedding = np.load(speaker_embeddings[speaker[:3]])
|
51 |
|
52 |
speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
|
53 |
|
@@ -91,12 +75,10 @@ article = """
|
|
91 |
"""
|
92 |
|
93 |
examples = [
|
94 |
-
["
|
95 |
-
["
|
96 |
-
["
|
97 |
-
["
|
98 |
-
["A synonym for cinnamon is a cinnamon synonym.", "BDL (male)"],
|
99 |
-
["How much wood would a woodchuck chuck if a woodchuck could chuck wood? He would chuck, he would, as much as he could, and chuck as much wood as a woodchuck would if a woodchuck could chuck wood.", "CLB (female)"],
|
100 |
]
|
101 |
|
102 |
gr.Interface(
|
@@ -104,14 +86,10 @@ gr.Interface(
|
|
104 |
inputs=[
|
105 |
gr.Text(label="Input Text"),
|
106 |
gr.Radio(label="Speaker", choices=[
|
107 |
-
"BDL (male)",
|
108 |
"CLB (female)",
|
109 |
-
|
110 |
-
"RMS (male)",
|
111 |
-
"SLT (female)",
|
112 |
-
"Surprise Me!"
|
113 |
],
|
114 |
-
value="
|
115 |
],
|
116 |
outputs=[
|
117 |
gr.Audio(label="Generated Speech", type="numpy"),
|
|
|
6 |
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
|
7 |
|
8 |
|
9 |
+
checkpoint = "GreenCounsel/speecht5_tts_common_voice_5_sv"
|
10 |
processor = SpeechT5Processor.from_pretrained(checkpoint)
|
11 |
model = SpeechT5ForTextToSpeech.from_pretrained(checkpoint)
|
12 |
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
|
|
|
31 |
input_ids = inputs["input_ids"]
|
32 |
input_ids = input_ids[..., :model.config.max_text_positions]
|
33 |
|
34 |
+
speaker_embedding = np.load(speaker_embeddings[speaker[:3]])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
|
36 |
speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
|
37 |
|
|
|
75 |
"""
|
76 |
|
77 |
examples = [
|
78 |
+
["GreenCounsel grundades i Malmö för sex år sedan.", "CLB (female)"],
|
79 |
+
["Med hjälp av maskininlärning kan mycket av juridiken automatiseras samtidigt som juristerna fokuserar på frågor där de ger störst värde.GreenCounsel har byggt en chatbott som kan förstå frågor på många olika språk och ge kvalitetssäkrade svar.She sells seashells by the seashore. I saw a kitten eating chicken in the kitchen.", "CLB (female)"],
|
80 |
+
["GreenCounsel har byggt en chatbott som kan förstå frågor på många olika språk och ge kvalitetssäkrade svar.", "CLB (female)"],
|
81 |
+
["Vi har också byggt ett system för att automatisera arbetsflöden för juridiska tjänster via internet.", "CLB (female)"],
|
|
|
|
|
82 |
]
|
83 |
|
84 |
gr.Interface(
|
|
|
86 |
inputs=[
|
87 |
gr.Text(label="Input Text"),
|
88 |
gr.Radio(label="Speaker", choices=[
|
|
|
89 |
"CLB (female)",
|
90 |
+
|
|
|
|
|
|
|
91 |
],
|
92 |
+
value="CLB (female)"),
|
93 |
],
|
94 |
outputs=[
|
95 |
gr.Audio(label="Generated Speech", type="numpy"),
|