Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -11,7 +11,7 @@ device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
|
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 = "
|
15 |
# pipe = pipeline("automatic-speech-recognition", model=model_id)
|
16 |
model = SpeechT5ForTextToSpeech.from_pretrained(model_id)
|
17 |
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
|
@@ -21,44 +21,31 @@ speaker_embeddings = torch.tensor(embeddings_dataset[7440]["xvector"]).unsqueeze
|
|
21 |
processor = SpeechT5Processor.from_pretrained(model_id)
|
22 |
|
23 |
replacements = [
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
("ó", "o"),
|
31 |
-
("ù", "u"),
|
32 |
-
("ú", "u"),
|
33 |
-
("š", "s"),
|
34 |
-
("ï", "i"),
|
35 |
-
("ñ", "n"),
|
36 |
-
("ü", "u"),
|
37 |
]
|
38 |
|
|
|
39 |
def cleanup_text(text):
|
40 |
for src, dst in replacements:
|
41 |
text = text.replace(src, dst)
|
42 |
return text
|
43 |
|
44 |
-
def synthesize_speech(text):
|
45 |
-
text = cleanup_text(text)
|
46 |
-
inputs = processor(text=text, return_tensors="pt")
|
47 |
-
speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
|
48 |
-
|
49 |
-
return gr.Audio.update(value=(16000, speech.cpu().numpy()))
|
50 |
-
|
51 |
def translate(audio):
|
52 |
-
outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "
|
53 |
return outputs["text"]
|
54 |
|
55 |
|
|
|
56 |
def synthesise(text):
|
57 |
text = cleanup_text(text)
|
58 |
inputs = processor(text=text, return_tensors="pt")
|
59 |
speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
|
60 |
-
return speech.cpu()
|
61 |
-
|
62 |
|
63 |
def speech_to_speech_translation(audio):
|
64 |
translated_text = translate(audio)
|
@@ -69,8 +56,8 @@ def speech_to_speech_translation(audio):
|
|
69 |
|
70 |
title = "Cascaded STST"
|
71 |
description = """
|
72 |
-
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in
|
73 |
-
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech, fine-tuned in
|
74 |
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
|
75 |
"""
|
76 |
|
|
|
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 = "ckandemir/speecht5_finetuned_voxpopuli_fr" # 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")
|
|
|
21 |
processor = SpeechT5Processor.from_pretrained(model_id)
|
22 |
|
23 |
replacements = [
|
24 |
+
("à", "a"), ("â", "a"),
|
25 |
+
("ç", "c"),
|
26 |
+
("é", "e"), ("è", "e"), ("ê", "e"), ("ë", "e"),
|
27 |
+
("î", "i"), ("ï", "i"),
|
28 |
+
("ô", "o"),
|
29 |
+
("ù", "u"), ("û", "u"),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
]
|
31 |
|
32 |
+
|
33 |
def cleanup_text(text):
|
34 |
for src, dst in replacements:
|
35 |
text = text.replace(src, dst)
|
36 |
return text
|
37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
def translate(audio):
|
39 |
+
outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "french"})
|
40 |
return outputs["text"]
|
41 |
|
42 |
|
43 |
+
|
44 |
def synthesise(text):
|
45 |
text = cleanup_text(text)
|
46 |
inputs = processor(text=text, return_tensors="pt")
|
47 |
speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
|
48 |
+
return gr.Audio.update(value=(16000, speech.cpu().numpy()))
|
|
|
49 |
|
50 |
def speech_to_speech_translation(audio):
|
51 |
translated_text = translate(audio)
|
|
|
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 French. Demo uses OpenAI's [Whisper Large v2](https://huggingface.co/openai/whisper-large-v2) model for speech translation, and [ckandemir/speecht5_finetuned_voxpopuli_fr"](https://huggingface.co/ckandemir/speecht5_finetuned_voxpopuli_fr) checkpoint for text-to-speech, which is based on Microsoft's
|
60 |
+
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech, fine-tuned in French Audio dataset:
|
61 |
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
|
62 |
"""
|
63 |
|