Update app.py

Added translation to french and a french speech synthetizer.

app.py

@@ -5,31 +5,52 @@ 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"]
 
 
+# load translator to french
+en_fr_translator = pipeline("translation_en_to_fr")
+
+
+# load text-to-speech
+model = VitsModel.from_pretrained("facebook/mms-tts-fra")
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-fra")
+
 def synthesise(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()
+    translation_to_french = en_fr_translator(text)
+    french_text = translation_to_french[0]['translation_text']
+
+    inputs = tokenizer(french_text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
+    
+    with torch.no_grad():
+        outputs = model(input_ids)
+    
+    speech = outputs["waveform"]
+    return speech
+
+
+# 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 synthesise(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()
 
 
 def speech_to_speech_translation(audio):
@@ -41,8 +62,10 @@ def speech_to_speech_translation(audio):
 
 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 French. 
+Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation,
+Google's [T5](https://huggingface.co/t5-base) for translating from English to French 
+and Facebook's [Massive Multilingual Speech (MMS)](https://huggingface.co/facebook/mms-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")
 """