Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from typing import Text
|
3 |
+
import gradio as gr
|
4 |
+
import soundfile as sf
|
5 |
+
from transformers import pipeline
|
6 |
+
import numpy as np
|
7 |
+
import torch
|
8 |
+
import re
|
9 |
+
from speechbrain.pretrained import EncoderClassifier
|
10 |
+
|
11 |
+
|
12 |
+
def create_speaker_embedding(speaker_model, waveform: np.ndarray) -> np.ndarray:
|
13 |
+
with torch.no_grad():
|
14 |
+
speaker_embeddings = speaker_model.encode_batch(torch.tensor(waveform))
|
15 |
+
speaker_embeddings = torch.nn.functional.normalize(speaker_embeddings, dim=2)
|
16 |
+
if device.type != 'cuda':
|
17 |
+
speaker_embeddings = speaker_embeddings.squeeze().numpy()
|
18 |
+
else:
|
19 |
+
speaker_embeddings = speaker_embeddings.squeeze().cpu().numpy()
|
20 |
+
speaker_embeddings = torch.tensor(speaker_embeddings, dtype=dtype).unsqueeze(0).to(device)
|
21 |
+
return speaker_embeddings
|
22 |
+
|
23 |
+
|
24 |
+
def remove_special_characters_s(text: Text) -> Text:
|
25 |
+
chars_to_remove_regex = '[\…\–\"\“\%\‘\”\�\»\«\„\`\'́]'
|
26 |
+
# remove special characters
|
27 |
+
text = re.sub(chars_to_remove_regex, '', text)
|
28 |
+
text = re.sub("՚", "'", text)
|
29 |
+
text = re.sub("’", "'", text)
|
30 |
+
text = re.sub(r'ы', 'и', text)
|
31 |
+
text = text.lower()
|
32 |
+
return text
|
33 |
+
|
34 |
+
|
35 |
+
def cyrillic_to_latin(text: Text) -> Text:
|
36 |
+
replacements = [
|
37 |
+
('а', 'a'),
|
38 |
+
('б', 'b'),
|
39 |
+
('в', 'v'),
|
40 |
+
('г', 'h'),
|
41 |
+
('д', 'd'),
|
42 |
+
('е', 'e'),
|
43 |
+
('ж', 'zh'),
|
44 |
+
('з', 'z'),
|
45 |
+
('и', 'y'),
|
46 |
+
('й', 'j'),
|
47 |
+
('к', 'k'),
|
48 |
+
('л', 'l'),
|
49 |
+
('м', 'm'),
|
50 |
+
('н', 'n'),
|
51 |
+
('о', 'o'),
|
52 |
+
('п', 'p'),
|
53 |
+
('р', 'r'),
|
54 |
+
('с', 's'),
|
55 |
+
('т', 't'),
|
56 |
+
('у', 'u'),
|
57 |
+
('ф', 'f'),
|
58 |
+
('х', 'h'),
|
59 |
+
('ц', 'ts'),
|
60 |
+
('ч', 'ch'),
|
61 |
+
('ш', 'sh'),
|
62 |
+
('щ', 'sch'),
|
63 |
+
('ь', "'"),
|
64 |
+
('ю', 'ju'),
|
65 |
+
('я', 'ja'),
|
66 |
+
('є', 'je'),
|
67 |
+
('і', 'i'),
|
68 |
+
('ї', 'ji'),
|
69 |
+
('ґ', 'g')
|
70 |
+
]
|
71 |
+
|
72 |
+
for src, dst in replacements:
|
73 |
+
text = text.replace(src, dst)
|
74 |
+
return text
|
75 |
+
|
76 |
+
|
77 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
78 |
+
|
79 |
+
dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
|
80 |
+
|
81 |
+
spk_model_name = "speechbrain/spkrec-xvect-voxceleb"
|
82 |
+
|
83 |
+
speaker_model = EncoderClassifier.from_hparams(
|
84 |
+
source=spk_model_name,
|
85 |
+
run_opts={"device": device},
|
86 |
+
savedir=os.path.join("/tmp", spk_model_name)
|
87 |
+
)
|
88 |
+
|
89 |
+
waveform, samplerate = sf.read("speaker.wav")
|
90 |
+
|
91 |
+
speaker_embeddings = create_speaker_embedding(speaker_model, waveform)
|
92 |
+
|
93 |
+
transcriber = pipeline("text-to-speech", model="Oysiyl/speecht5_tts_common_voice_uk")
|
94 |
+
|
95 |
+
def transcribe(text: Text) -> tuple((int, np.ndarray)):
|
96 |
+
text = remove_special_characters_s(text)
|
97 |
+
text = cyrillic_to_latin(text)
|
98 |
+
out = transcriber(text, forward_params={"speaker_embeddings": speaker_embeddings})
|
99 |
+
audio, sr = out["audio"], out["sampling_rate"]
|
100 |
+
return sr, audio
|
101 |
+
|
102 |
+
|
103 |
+
demo = gr.Interface(
|
104 |
+
transcribe,
|
105 |
+
gr.Textbox(),
|
106 |
+
outputs="audio",
|
107 |
+
title="Text to Speech for Ukrainian language demo",
|
108 |
+
description="Click on the example below or type text!",
|
109 |
+
examples=[["Держава-агресор Росія закуповує комунікаційне обладнання, зокрема супутникові інтернет-термінали Starlink, для використання у війні в арабських країнах"],
|
110 |
+
["Доброго вечора, ми з України!"]],
|
111 |
+
cache_examples=True
|
112 |
+
)
|
113 |
+
|
114 |
+
demo.launch()
|