Update app.py

app.py

@@ -1,60 +1,195 @@
 import gradio as gr
 import numpy as np
 import torch
-
-from transformers import (
-    VitsModel,
-    VitsTokenizer,
-    pipeline
-)
+from datasets import load_dataset
+from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 # load speech translation checkpoint
-asr_pipe = pipeline(
-    "automatic-speech-recognition",
-    model="openai/whisper-base",
-    device=device
-)
-
-model = VitsModel.from_pretrained("Matthijs/mms-tts-deu")
-tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-deu")
-
-
-def translate(audio):
-    outputs = asr_pipe(
-        audio,
-        max_new_tokens=256,
-        generate_kwargs={"task": "transcribe", "language": "de"}
-    )
+asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-large-v2", device=device)
+greek_translation_pipe = pipeline("translation", model="Helsinki-NLP/opus-mt-en-el")
+
+# load text-to-speech checkpoint and speaker embeddings
+model_id = "microsoft/speecht5_tts"  # update with your model id
+# pipe = pipeline("automatic-speech-recognition", model=model_id)
+model = SpeechT5ForTextToSpeech.from_pretrained(model_id)
+vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+speaker_embeddings = torch.tensor(embeddings_dataset[7440]["xvector"]).unsqueeze(0)
+
+processor = SpeechT5Processor.from_pretrained(model_id)
+
+model_id_greek = "Sandiago21/speecht5_finetuned_google_fleurs_greek"
+model_greek = SpeechT5ForTextToSpeech.from_pretrained(model_id_greek)
+processor_greek = SpeechT5Processor.from_pretrained(model_id_greek)
+
+replacements = [
+    ("á", "a"),
+    ("â", "a"),
+    ("ã", "a"),
+    ("í", "i"),
+    ("á", "a"),
+    ("í", "i"),
+    ("ñ", "n"),
+    ("ó", "o"),
+    ("ú", "u"),
+    ("ü", "u"),
+    ("á", "a"),
+    ("ç", "c"),
+    ("è", "e"),
+    ("ì", "i"),
+    ("í", "i"),
+    ("ò", "o"),
+    ("ó", "o"),
+    ("ù", "u"),
+    ("ú", "u"),
+    ("š", "s"),
+    ("ï", "i"),
+    ("à", "a"),
+    ("â", "a"),
+    ("ç", "c"),
+    ("è", "e"),
+    ("ë", "e"),
+    ("î", "i"),
+    ("ï", "i"),
+    ("ô", "o"),
+    ("ù", "u"),
+    ("û", "u"),
+    ("ü", "u"),
+    ("ου", "u"),
+    ("αυ", "af"),
+    ("ευ", "ef"),
+    ("ει", "i"),
+    ("οι", "i"),
+    ("αι", "e"),
+    ("ού", "u"),
+    ("εί", "i"),
+    ("οί", "i"),
+    ("αί", "e"),
+    ("Ά", "A"),
+    ("Έ", "E"),
+    ("Ή", "H"),
+    ("Ί", "I"),
+    ("Ό", "O"),
+    ("Ύ", "Y"),
+    ("Ώ", "O"),
+    ("ΐ", "i"),
+    ("Α", "A"),
+    ("Β", "B"),
+    ("Γ", "G"),
+    ("Δ", "L"),
+    ("Ε", "Ε"),
+    ("Ζ", "Z"),
+    ("Η", "I"),
+    ("Θ", "Th"),
+    ("Ι", "I"),
+    ("Κ", "K"),
+    ("Λ", "L"),
+    ("Μ", "M"),
+    ("Ν", "N"),
+    ("Ξ", "Ks"),
+    ("Ο", "O"),
+    ("Π", "P"),
+    ("Ρ", "R"),
+    ("Σ", "S"),
+    ("Τ", "T"),
+    ("Υ", "Y"),
+    ("Φ", "F"),
+    ("Χ", "X"),
+    ("Ω", "O"),
+    ("ά", "a"),
+    ("έ", "e"),
+    ("ή", "i"),
+    ("ί", "i"),
+    ("α", "a"),
+    ("β", "v"),
+    ("γ", "g"),
+    ("δ", "d"),
+    ("ε", "e"),
+    ("ζ", "z"),
+    ("η", "i"),
+    ("θ", "th"),
+    ("ι", "i"),
+    ("κ", "k"),
+    ("λ", "l"),
+    ("μ", "m"),
+    ("ν", "n"),
+    ("ξ", "ks"),
+    ("ο", "o"),
+    ("π", "p"),
+    ("ρ", "r"),
+    ("ς", "s"),
+    ("σ", "s"),
+    ("τ", "t"),
+    ("υ", "i"),
+    ("φ", "f"),
+    ("χ", "h"),
+    ("ψ", "ps"),
+    ("ω", "o"),
+    ("ϊ", "i"),
+    ("ϋ", "i"),
+    ("ό", "o"),
+    ("ύ", "i"),
+    ("ώ", "o"),
+    ("í", "i"),
+    ("õ", "o"),
+    ("Ε", "E"),
+    ("Ψ", "Ps"),
+]
+
+def cleanup_text(text):
+    for src, dst in replacements:
+        text = text.replace(src, dst)
+    return text
+
+
+def synthesize_speech(text):
+    text = cleanup_text(text)
+    inputs = processor(text=text, return_tensors="pt")
+    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+
+    return gr.Audio.update(value=(16000, speech.cpu().numpy()))
+
+
+def translate_to_english(audio):
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate", "language": "english"})
     return outputs["text"]
 
 
-def synthesise(text):
-    if len(text.strip()) == 0:
-        return (16000, np.zeros(0).astype(np.int16))
+def synthesise_from_english(text):
+    text = cleanup_text(text)
+    inputs = processor(text=text, return_tensors="pt")
+    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+    return speech.cpu().numpy()
+
 
-    inputs = tokenizer(text, return_tensors="pt")
-    input_ids = inputs["input_ids"]
+def translate_from_english_to_greek(text):
+    return greek_translation_pipe(text)[0]["translation_text"]
 
-    with torch.no_grad():
-        outputs = model(input_ids)
 
-    speech = outputs.audio[0]
+def synthesise_from_greek(text):
+    text = cleanup_text(text)
+    inputs = processor_greek(text=text, return_tensors="pt")
+    speech = model_greek.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
     return speech.cpu()
 
 
 def speech_to_speech_translation(audio):
-    translated_text = translate(audio)
-    synthesised_speech = synthesise(translated_text)
+    translated_text = translate_to_english(audio)
+    translated_text = translate_from_english_to_greek(translated_text)
+#     synthesised_speech = synthesise_from_english(translated_text)
+#     translated_text = translate_from_english_to_greek(synthesised_speech)
+    synthesised_speech = synthesise_from_greek(translated_text)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
-    return 16000, synthesised_speech
+    return ((16000, synthesised_speech), translated_text)
+
 
 title = "Cascaded STST"
 description = """
-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
-[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Greek. Demo uses OpenAI's [Whisper Large v2](https://huggingface.co/openai/whisper-large-v2) model for speech translation, and [Sandiago21/speecht5_finetuned_google_fleurs_greek](https://huggingface.co/Sandiago21/speecht5_finetuned_google_fleurs_greek) checkpoint for text-to-speech, which is based on Microsoft's
+[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech, fine-tuned in Greek Audio dataset:
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 """
 
@@ -63,14 +198,15 @@ demo = gr.Blocks()
 mic_translate = gr.Interface(
     fn=speech_to_speech_translation,
     inputs=gr.Audio(source="microphone", type="filepath"),
-    outputs=gr.Audio(label="Generated Speech", type="numpy"),
+    outputs=[gr.Audio(label="Generated Speech", type="numpy"), gr.outputs.Textbox()],
     title=title,
     description=description,
 )
+
 file_translate = gr.Interface(
     fn=speech_to_speech_translation,
     inputs=gr.Audio(source="upload", type="filepath"),
-    outputs=gr.Audio(label="Generated Speech", type="numpy"),
+    outputs=[gr.Audio(label="Generated Speech", type="numpy"), gr.outputs.Textbox()],
     examples=[["./example.wav"]],
     title=title,
     description=description,