ykirpichev
commited on
Commit
•
8e92972
1
Parent(s):
6ae1b69
Update app.py
Browse filesUse:
generate_kwargs={"task": "transcribe", "language": "de"}) to translate in German
Use:
Matthijs/mms-tts-deu for TTS
app.py
CHANGED
@@ -4,7 +4,7 @@ import torch
|
|
4 |
from datasets import load_dataset
|
5 |
|
6 |
from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
|
7 |
-
|
8 |
|
9 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
10 |
|
@@ -12,24 +12,35 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
|
12 |
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
|
13 |
|
14 |
# load text-to-speech checkpoint and speaker embeddings
|
15 |
-
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
|
|
|
|
|
|
|
|
|
16 |
|
17 |
-
model =
|
18 |
-
|
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 = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "
|
26 |
return outputs["text"]
|
27 |
|
28 |
|
29 |
def synthesise(text):
|
30 |
-
inputs =
|
31 |
-
|
32 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
|
34 |
|
35 |
def speech_to_speech_translation(audio):
|
@@ -41,8 +52,8 @@ def speech_to_speech_translation(audio):
|
|
41 |
|
42 |
title = "Cascaded STST"
|
43 |
description = """
|
44 |
-
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in
|
45 |
-
[
|
46 |
|
47 |
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
|
48 |
"""
|
|
|
4 |
from datasets import load_dataset
|
5 |
|
6 |
from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
|
7 |
+
from transformers import VitsModel, VitsTokenizer
|
8 |
|
9 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
10 |
|
|
|
12 |
asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
|
13 |
|
14 |
# load text-to-speech checkpoint and speaker embeddings
|
15 |
+
# processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
|
16 |
+
|
17 |
+
# model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
|
18 |
+
# vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
|
19 |
+
|
20 |
|
21 |
+
model = VitsModel.from_pretrained("Matthijs/mms-tts-deu").to(device)
|
22 |
+
tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-deu")
|
23 |
|
24 |
+
# embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
|
25 |
+
# speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
|
26 |
|
27 |
|
28 |
def translate(audio):
|
29 |
+
outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "de"})
|
30 |
return outputs["text"]
|
31 |
|
32 |
|
33 |
def synthesise(text):
|
34 |
+
inputs = tokenizer(text_example, return_tensors="pt")
|
35 |
+
input_ids = inputs["input_ids"]
|
36 |
+
|
37 |
+
|
38 |
+
with torch.no_grad():
|
39 |
+
outputs = model(input_ids)
|
40 |
+
return outputs.cpu()
|
41 |
+
# inputs = processor(text=text, return_tensors="pt")
|
42 |
+
# speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
|
43 |
+
# return speech.cpu()
|
44 |
|
45 |
|
46 |
def speech_to_speech_translation(audio):
|
|
|
52 |
|
53 |
title = "Cascaded STST"
|
54 |
description = """
|
55 |
+
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in German. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and
|
56 |
+
[Massive Multilingual Speech (MMS) TTS](https://huggingface.co/learn/audio-course/chapter6/pre-trained_models#massive-multilingual-speech-mms) model for text-to-speech:
|
57 |
|
58 |
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
|
59 |
"""
|