elizabetvaganova
commited on
Commit
·
6626f2f
1
Parent(s):
7b53c86
Update app.py
Browse files
app.py
CHANGED
@@ -2,70 +2,15 @@ import gradio as gr
|
|
2 |
import numpy as np
|
3 |
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 |
|
11 |
-
#
|
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 |
-
asr_pipe = pipeline("automatic-speech-recognition", model="new/asr/model", device=device)
|
18 |
-
|
19 |
-
embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
|
20 |
-
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
|
21 |
-
|
22 |
-
|
23 |
-
def translate(audio):
|
24 |
-
outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
|
25 |
-
return outputs["text"]
|
26 |
-
|
27 |
-
|
28 |
-
def synthesise(text):
|
29 |
-
inputs = processor(text=text, return_tensors="pt")
|
30 |
-
speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
|
31 |
-
return speech.cpu()
|
32 |
-
|
33 |
-
|
34 |
-
def speech_to_speech_translation(audio):
|
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
|
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 English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
|
44 |
-
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
|
45 |
-
|
46 |
-
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
|
47 |
-
"""
|
48 |
-
|
49 |
-
demo = gr.Blocks()
|
50 |
-
|
51 |
-
mic_translate = gr.Interface(
|
52 |
-
fn=speech_to_speech_translation,
|
53 |
-
inputs=gr.Audio(source="microphone", type="filepath"),
|
54 |
-
outputs=gr.Audio(label="Generated Speech", type="numpy"),
|
55 |
-
title=title,
|
56 |
-
description=description,
|
57 |
-
)
|
58 |
-
|
59 |
-
file_translate = gr.Interface(
|
60 |
-
fn=speech_to_speech_translation,
|
61 |
-
inputs=gr.Audio(source="upload", type="filepath"),
|
62 |
-
outputs=gr.Audio(label="Generated Speech", type="numpy"),
|
63 |
-
examples=[["./example.wav"]],
|
64 |
-
title=title,
|
65 |
-
description=description,
|
66 |
-
)
|
67 |
|
68 |
-
|
69 |
-
|
70 |
|
71 |
-
|
|
|
|
2 |
import numpy as np
|
3 |
import torch
|
4 |
from datasets import load_dataset
|
|
|
5 |
from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
|
6 |
|
|
|
7 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
8 |
|
9 |
+
# Load a lightweight automatic speech recognition model (vosk)
|
10 |
+
asr_pipe = pipeline("automatic-speech-recognition", model="alphacep/kaldi-ru", device=device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
+
# Load a lightweight text-to-speech checkpoint and speaker embeddings
|
13 |
+
processor = SpeechT5Processor.from_pretrained("ttskit/ttskit-tts-ljspeech")
|
14 |
|
15 |
+
model = SpeechT5ForTextToSpeech.from_pretrained("ttskit/ttskit-tts-ljspeech").to(device)
|
16 |
+
vocoder = SpeechT5HifiGan.from_pretrained("ljspeech/vocoder-cry
|