Update app.py

app.py

@@ -2,71 +2,87 @@ import gradio as gr
 import numpy as np
 import torch
 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)
-
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
-
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
-
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
-
-
-def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
-    return outputs["text"]
-
-
-def synthesise(text):
+# Load Whisper large-v2 model for multilingual speech translation
+asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-large-v2", device=device)
+
+# Load MMS TTS model for multilingual text-to-speech
+processor = SpeechT5Processor.from_pretrained("facebook/mms-tts-eng")
+model = SpeechT5ForTextToSpeech.from_pretrained("facebook/mms-tts-eng").to(device)
+vocoder = SpeechT5HifiGan.from_pretrained("facebook/mms-tts-eng").to(device)
+
+# Define supported languages
+LANGUAGES = {
+    "French": "fra", "German": "deu", "Spanish": "spa", "Italian": "ita",
+    "Portuguese": "por", "Dutch": "nld", "Russian": "rus", "Chinese": "cmn",
+    "Japanese": "jpn", "Korean": "kor"
+}
+
+def translate(audio, source_lang, target_lang):
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={
+        "task": "transcribe",
+        "language": source_lang,
+    })
+    transcription = outputs["text"]
+    
+    # Use Whisper for translation
+    translation = asr_pipe(transcription, max_new_tokens=256, generate_kwargs={
+        "task": "translate",
+        "language": target_lang,
+    })["text"]
+    
+    return translation
+
+def synthesise(text, target_lang):
     inputs = processor(text=text, return_tensors="pt")
-    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+    speech = model.generate_speech(inputs["input_ids"].to(device), vocoder=vocoder, language=target_lang)
     return speech.cpu()
 
-
-def speech_to_speech_translation(audio):
-    translated_text = translate(audio)
-    synthesised_speech = synthesise(translated_text)
+def speech_to_speech_translation(audio, source_lang, target_lang):
+    translated_text = translate(audio, source_lang, target_lang)
+    synthesised_speech = synthesise(translated_text, target_lang)
     synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
     return 16000, synthesised_speech
 
-
-title = "Cascaded STST"
+title = "Multilingual Speech-to-Speech Translation"
 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:
-
-![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
+Demo for multilingual speech-to-speech translation (STST), mapping from source speech in any supported language to target speech in any other supported language.
 """
 
 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"),
-    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"),
-    examples=[["./example.wav"]],
-    title=title,
-    description=description,
-)
-
 with demo:
-    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
-
-demo.launch()
+    gr.Markdown(f"# {title}")
+    gr.Markdown(description)
+    
+    with gr.Row():
+        source_lang = gr.Dropdown(choices=list(LANGUAGES.keys()), label="Source Language")
+        target_lang = gr.Dropdown(choices=list(LANGUAGES.keys()), label="Target Language")
+    
+    with gr.Tabs():
+        with gr.TabItem("Microphone"):
+            mic_input = gr.Audio(source="microphone", type="filepath")
+            mic_output = gr.Audio(label="Generated Speech", type="numpy")
+            mic_button = gr.Button("Translate")
+        
+        with gr.TabItem("Audio File"):
+            file_input = gr.Audio(source="upload", type="filepath")
+            file_output = gr.Audio(label="Generated Speech", type="numpy")
+            file_button = gr.Button("Translate")
+    
+    mic_button.click(
+        speech_to_speech_translation,
+        inputs=[mic_input, source_lang, target_lang],
+        outputs=mic_output
+    )
+    
+    file_button.click(
+        speech_to_speech_translation,
+        inputs=[file_input, source_lang, target_lang],
+        outputs=file_output
+    )
+
+demo.launch()