heisenberg3376
commited on
Commit
•
bfd1577
1
Parent(s):
1864517
Update app.py
Browse files
app.py
CHANGED
@@ -11,6 +11,8 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
|
11 |
|
12 |
# load speech translation checkpoint
|
13 |
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
|
|
|
|
|
14 |
|
15 |
model = VitsModel.from_pretrained('facebook/mms-tts-rus').to(device)
|
16 |
tokenizer = VitsTokenizer.from_pretrained('facebook/mms-tts-rus')
|
@@ -18,8 +20,8 @@ tokenizer = VitsTokenizer.from_pretrained('facebook/mms-tts-rus')
|
|
18 |
|
19 |
|
20 |
def translate(audio):
|
21 |
-
outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "
|
22 |
-
return outputs[
|
23 |
|
24 |
|
25 |
def synthesise(text):
|
@@ -33,12 +35,14 @@ def speech_to_speech_translation(audio):
|
|
33 |
translated_text = translate(audio)
|
34 |
synthesised_speech = synthesise(translated_text)
|
35 |
synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
|
36 |
-
return 16000, synthesised_speech
|
37 |
|
38 |
|
39 |
title = "Cascaded STST"
|
40 |
description = """
|
41 |
-
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Russian
|
|
|
|
|
42 |
|
43 |
demo = gr.Blocks()
|
44 |
|
|
|
11 |
|
12 |
# load speech translation checkpoint
|
13 |
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
|
14 |
+
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-en-ru")
|
15 |
+
|
16 |
|
17 |
model = VitsModel.from_pretrained('facebook/mms-tts-rus').to(device)
|
18 |
tokenizer = VitsTokenizer.from_pretrained('facebook/mms-tts-rus')
|
|
|
20 |
|
21 |
|
22 |
def translate(audio):
|
23 |
+
outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
|
24 |
+
return translator(outputs['text'])[0]['translation_text']
|
25 |
|
26 |
|
27 |
def synthesise(text):
|
|
|
35 |
translated_text = translate(audio)
|
36 |
synthesised_speech = synthesise(translated_text)
|
37 |
synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
|
38 |
+
return 16000, synthesised_speech[0]
|
39 |
|
40 |
|
41 |
title = "Cascaded STST"
|
42 |
description = """
|
43 |
+
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Russian
|
44 |
+
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
|
45 |
+
"""
|
46 |
|
47 |
demo = gr.Blocks()
|
48 |
|