hewliyang
commited on
Commit
•
0323180
1
Parent(s):
dbfdf1a
use whisper-large-v3 & mms-tts-zlm
Browse files- .gitignore +2 -0
- README.md +7 -0
- app.py +36 -20
- dev.ipynb +0 -0
- examples/anime-jap.mp3 +0 -0
- examples/english-1.wav +0 -0
- examples/english-2.wav +0 -0
- examples/spanish.wav +0 -0
- requirements.txt +1 -1
.gitignore
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
__pycache__
|
2 |
+
.venv
|
README.md
CHANGED
@@ -9,4 +9,11 @@ app_file: app.py
|
|
9 |
pinned: false
|
10 |
---
|
11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
9 |
pinned: false
|
10 |
---
|
11 |
|
12 |
+
Part of the HuggingFace Audio Processing course.
|
13 |
+
|
14 |
+
This is a Gradio wrapper around a (X -> Malay) speech2speech pipeline, where **X** is any language supported by
|
15 |
+
`openai/whisper-base`.
|
16 |
+
|
17 |
+
The TTS model used is `facebook/mms-tts-zlm`, a pretrained checkpoint for speech in Malay which is part of their **Massively Multilingual Speech** project. The underlying architecture is based on VITS, which generates waveforms directly and does not need a seperate vocoder.
|
18 |
+
|
19 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.py
CHANGED
@@ -1,47 +1,62 @@
|
|
|
|
1 |
import gradio as gr
|
2 |
import numpy as np
|
3 |
-
import torch
|
4 |
-
from datasets import load_dataset
|
5 |
|
6 |
-
from transformers import
|
|
|
|
|
|
|
|
|
7 |
|
8 |
|
9 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
|
|
|
|
10 |
|
11 |
# load speech translation checkpoint
|
12 |
-
asr_pipe = pipeline(
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
-
# load text-to-speech checkpoint
|
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 |
-
|
25 |
-
|
26 |
-
return outputs["text"]
|
27 |
|
|
|
|
|
28 |
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
|
|
|
|
|
|
|
|
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() * 32767).astype(np.int16)
|
39 |
-
return 16000, synthesised_speech
|
40 |
|
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 |
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
|
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")
|
@@ -61,9 +76,10 @@ 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=
|
65 |
title=title,
|
66 |
description=description,
|
|
|
67 |
)
|
68 |
|
69 |
with demo:
|
|
|
1 |
+
import torch
|
2 |
import gradio as gr
|
3 |
import numpy as np
|
|
|
|
|
4 |
|
5 |
+
from transformers import (
|
6 |
+
VitsModel,
|
7 |
+
VitsTokenizer,
|
8 |
+
pipeline,
|
9 |
+
)
|
10 |
|
11 |
|
12 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
13 |
+
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
14 |
+
print(f"Using {device} with fp {torch_dtype}")
|
15 |
|
16 |
# load speech translation checkpoint
|
17 |
+
asr_pipe = pipeline( # noqa: F821
|
18 |
+
"automatic-speech-recognition",
|
19 |
+
model="openai/whisper-large-v3",
|
20 |
+
device=device,
|
21 |
+
torch_dtype=torch_dtype,
|
22 |
+
)
|
23 |
|
24 |
+
# load text-to-speech checkpoint
|
|
|
25 |
|
26 |
+
model = VitsModel.from_pretrained("facebook/mms-tts-zlm")
|
27 |
+
tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-zlm")
|
28 |
|
|
|
|
|
29 |
|
30 |
+
def synthesise(text):
|
31 |
+
inputs = tokenizer(text=text, return_tensors="pt")
|
32 |
+
input_ids = inputs["input_ids"]
|
33 |
|
34 |
+
with torch.no_grad():
|
35 |
+
outputs = model(input_ids)
|
|
|
36 |
|
37 |
+
speech = outputs["waveform"]
|
38 |
+
return speech
|
39 |
|
40 |
+
|
41 |
+
def translate(audio):
|
42 |
+
outputs = asr_pipe(
|
43 |
+
audio,
|
44 |
+
max_new_tokens=256,
|
45 |
+
generate_kwargs={"task": "transcribe", "language": "ms"},
|
46 |
+
)
|
47 |
+
return outputs["text"]
|
48 |
|
49 |
|
50 |
def speech_to_speech_translation(audio):
|
51 |
translated_text = translate(audio)
|
52 |
synthesised_speech = synthesise(translated_text)
|
53 |
synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
|
54 |
+
return 16000, synthesised_speech.T
|
55 |
|
56 |
|
57 |
title = "Cascaded STST"
|
58 |
description = """
|
59 |
+
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in **Malay**. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
|
60 |
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
|
61 |
|
62 |
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
|
|
|
76 |
fn=speech_to_speech_translation,
|
77 |
inputs=gr.Audio(source="upload", type="filepath"),
|
78 |
outputs=gr.Audio(label="Generated Speech", type="numpy"),
|
79 |
+
examples="./examples",
|
80 |
title=title,
|
81 |
description=description,
|
82 |
+
live=True,
|
83 |
)
|
84 |
|
85 |
with demo:
|
dev.ipynb
ADDED
The diff for this file is too large to render.
See raw diff
|
|
examples/anime-jap.mp3
ADDED
Binary file (608 kB). View file
|
|
examples/english-1.wav
ADDED
Binary file (465 kB). View file
|
|
examples/english-2.wav
ADDED
Binary file (426 kB). View file
|
|
examples/spanish.wav
ADDED
Binary file (263 kB). View file
|
|
requirements.txt
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
torch
|
2 |
-
git+https://github.com/huggingface/transformers
|
3 |
datasets
|
4 |
sentencepiece
|
|
|
|
1 |
torch
|
|
|
2 |
datasets
|
3 |
sentencepiece
|
4 |
+
transformers
|