Sandiago21 commited on
Commit
974f879
1 Parent(s): c54a4dd

Upload folder using huggingface_hub

Browse files
Files changed (5) hide show
  1. README.md +3 -9
  2. app.py +98 -0
  3. example.wav +0 -0
  4. packages.txt +2 -0
  5. requirements.txt +6 -0
README.md CHANGED
@@ -1,12 +1,6 @@
1
  ---
2
- title: Speech To Speech Translation Italian
3
- emoji: 🌖
4
- colorFrom: indigo
5
- colorTo: yellow
6
- sdk: gradio
7
- sdk_version: 3.36.1
8
  app_file: app.py
9
- pinned: false
 
10
  ---
11
-
12
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
1
  ---
2
+ title: speech-to-speech-translation-italian
 
 
 
 
 
3
  app_file: app.py
4
+ sdk: gradio
5
+ sdk_version: 3.36.0
6
  ---
 
 
app.py ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ import numpy as np
3
+ import torch
4
+ from datasets import load_dataset
5
+ from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
6
+
7
+
8
+ device = "cuda:0" if torch.cuda.is_available() else "cpu"
9
+
10
+ # load speech translation checkpoint
11
+ asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-large-v2", device=device)
12
+
13
+ # load text-to-speech checkpoint and speaker embeddings
14
+ model_id = "Sandiago21/speecht5_finetuned_voxpopuli_it" # update with your model id
15
+ # pipe = pipeline("automatic-speech-recognition", model=model_id)
16
+ model = SpeechT5ForTextToSpeech.from_pretrained(model_id)
17
+ vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
18
+ embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
19
+ speaker_embeddings = torch.tensor(embeddings_dataset[7440]["xvector"]).unsqueeze(0)
20
+
21
+ checkpoint = "microsoft/speecht5_tts"
22
+ processor = SpeechT5Processor.from_pretrained(checkpoint)
23
+
24
+ replacements = [
25
+ ("á", "a"),
26
+ ("ç", "c"),
27
+ ("è", "e"),
28
+ ("ì", "i"),
29
+ ("í", "i"),
30
+ ("ò", "o"),
31
+ ("ó", "o"),
32
+ ("ù", "u"),
33
+ ("ú", "u"),
34
+ ("š", "s"),
35
+ ("ï", "i"),
36
+ ]
37
+
38
+ def cleanup_text(text):
39
+ for src, dst in replacements:
40
+ text = text.replace(src, dst)
41
+ return text
42
+
43
+ def synthesize_speech(text):
44
+ text = cleanup_text(text)
45
+ inputs = processor(text=text, return_tensors="pt")
46
+ speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
47
+
48
+ return gr.Audio.update(value=(16000, speech.cpu().numpy()))
49
+
50
+ def translate(audio):
51
+ outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
52
+ return outputs["text"]
53
+
54
+
55
+ def synthesise(text):
56
+ text = cleanup_text(text)
57
+ inputs = processor(text=text, return_tensors="pt")
58
+ speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
59
+ return speech.cpu()
60
+
61
+
62
+ def speech_to_speech_translation(audio):
63
+ translated_text = translate(audio)
64
+ synthesised_speech = synthesise(translated_text)
65
+ synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
66
+ return 16000, synthesised_speech
67
+
68
+
69
+ title = "Cascaded STST"
70
+ description = """
71
+ 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
72
+ [SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
73
+ ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
74
+ """
75
+
76
+ demo = gr.Blocks()
77
+
78
+ mic_translate = gr.Interface(
79
+ fn=speech_to_speech_translation,
80
+ inputs=gr.Audio(source="microphone", type="filepath"),
81
+ outputs=gr.Audio(label="Generated Speech", type="numpy"),
82
+ title=title,
83
+ description=description,
84
+ )
85
+
86
+ file_translate = gr.Interface(
87
+ fn=speech_to_speech_translation,
88
+ inputs=gr.Audio(source="upload", type="filepath"),
89
+ outputs=gr.Audio(label="Generated Speech", type="numpy"),
90
+ examples=[["./example.wav"]],
91
+ title=title,
92
+ description=description,
93
+ )
94
+
95
+ with demo:
96
+ gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
97
+
98
+ demo.launch()
example.wav ADDED
Binary file (263 kB). View file
 
packages.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ ffmpeg
2
+
requirements.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ torch
2
+ git+https://github.com/huggingface/transformers
3
+ datasets
4
+ torchaudio
5
+ sentencepiece
6
+