Update app.py
Browse filesUsage of belarusan language translation
app.py
CHANGED
@@ -1,3 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
import numpy as np
|
3 |
import torch
|
@@ -9,22 +15,43 @@ from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Proce
|
|
9 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
10 |
|
11 |
# load speech translation checkpoint
|
12 |
-
asr_pipe = pipeline("automatic-speech-recognition", model="
|
13 |
|
14 |
# load text-to-speech checkpoint and speaker embeddings
|
15 |
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
|
16 |
|
17 |
-
model = SpeechT5ForTextToSpeech.from_pretrained("
|
18 |
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
|
19 |
|
20 |
embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
|
21 |
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
|
22 |
|
23 |
|
24 |
-
def translate(audio):
|
25 |
-
outputs =
|
26 |
-
return outputs["text"]
|
27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
29 |
def synthesise(text):
|
30 |
inputs = processor(text=text, return_tensors="pt")
|
@@ -32,10 +59,14 @@ def synthesise(text):
|
|
32 |
return speech.cpu()
|
33 |
|
34 |
|
|
|
|
|
|
|
|
|
35 |
def speech_to_speech_translation(audio):
|
36 |
-
translated_text = translate(audio)
|
37 |
synthesised_speech = synthesise(translated_text)
|
38 |
-
synthesised_speech = (synthesised_speech.numpy() *
|
39 |
return 16000, synthesised_speech
|
40 |
|
41 |
|
@@ -61,7 +92,7 @@ file_translate = gr.Interface(
|
|
61 |
fn=speech_to_speech_translation,
|
62 |
inputs=gr.Audio(source="upload", type="filepath"),
|
63 |
outputs=gr.Audio(label="Generated Speech", type="numpy"),
|
64 |
-
examples=[["./example.wav"]],
|
65 |
title=title,
|
66 |
description=description,
|
67 |
)
|
|
|
1 |
+
!apt-get install -y perl
|
2 |
+
!wget https://www.isi.edu/~ulf/uroman/downloads/uroman-v1.2.7.tar.gz
|
3 |
+
!mkdir uroman
|
4 |
+
!tar -zxvf ./uroman-v1.2.7.tar.gz -C ./uroman
|
5 |
+
!chmod +x ./uroman/bin/uroman.pl
|
6 |
+
|
7 |
import gradio as gr
|
8 |
import numpy as np
|
9 |
import torch
|
|
|
15 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
16 |
|
17 |
# load speech translation checkpoint
|
18 |
+
asr_pipe = pipeline("automatic-speech-recognition", model="KoRiF/whisper-small-be", device=device)
|
19 |
|
20 |
# load text-to-speech checkpoint and speaker embeddings
|
21 |
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
|
22 |
|
23 |
+
model = SpeechT5ForTextToSpeech.from_pretrained("KoRiF/speecht5_finetuned_common_voice_be").to(device)
|
24 |
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
|
25 |
|
26 |
embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
|
27 |
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
|
28 |
|
29 |
|
30 |
+
def translate(audio, transliteration = lambda txt: txt):
|
31 |
+
outputs = pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "be"})#larusian
|
32 |
+
return transliteration(outputs["text"])
|
33 |
+
|
34 |
+
import subprocess
|
35 |
+
|
36 |
+
def transliterate_text(text, lang_code=None, use_chart=False, use_cache=True):
|
37 |
+
command = ['perl', './uroman/bin/uroman.pl']
|
38 |
+
if lang_code:
|
39 |
+
command.extend(['-l', lang_code])
|
40 |
+
if use_chart:
|
41 |
+
command.append('--chart')
|
42 |
+
if not use_cache:
|
43 |
+
command.append('--no-cache')
|
44 |
+
|
45 |
+
process = subprocess.Popen(command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
|
46 |
+
universal_newlines=True)
|
47 |
+
output, error = process.communicate(input=text)
|
48 |
+
if (error):
|
49 |
+
print(f"Error: >>> {error}")
|
50 |
+
return output.strip()
|
51 |
+
|
52 |
+
language = 'bel'
|
53 |
+
def transliterate(text):
|
54 |
+
return transliterate_text(text, language)
|
55 |
|
56 |
def synthesise(text):
|
57 |
inputs = processor(text=text, return_tensors="pt")
|
|
|
59 |
return speech.cpu()
|
60 |
|
61 |
|
62 |
+
|
63 |
+
target_dtype = np.int16
|
64 |
+
max_range = np.iinfo(target_dtype).max
|
65 |
+
|
66 |
def speech_to_speech_translation(audio):
|
67 |
+
translated_text = translate(audio, transliterate)#
|
68 |
synthesised_speech = synthesise(translated_text)
|
69 |
+
synthesised_speech = (synthesised_speech.numpy() * max_range).astype(np.int16)
|
70 |
return 16000, synthesised_speech
|
71 |
|
72 |
|
|
|
92 |
fn=speech_to_speech_translation,
|
93 |
inputs=gr.Audio(source="upload", type="filepath"),
|
94 |
outputs=gr.Audio(label="Generated Speech", type="numpy"),
|
95 |
+
#examples=[["./example.wav"]],
|
96 |
title=title,
|
97 |
description=description,
|
98 |
)
|