Update app.py

app.py

@@ -3,7 +3,7 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+from transformers import VitsModel, VitsTokenizer, pipeline
 
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
@@ -11,14 +11,11 @@ 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")
+# load text-to-speech checkpoint
+tokenizer = VitsTokenizer.from_pretrained("Matthijs/mms-tts-deu")
 
-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)
+model = VitsModel.from_pretrained("Matthijs/mms-tts-deu")
+model.to(device)
 
 
 def translate(audio):
@@ -26,23 +23,29 @@ def translate(audio):
     return outputs["text"]
 
 
-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 synthesize(text):
+    input = tokenizer(text, return_tensors="pt")
+    with torch.no_grad():
+        output = model(input['input_ids'].to(device))
+    
+    return output.audio[0].cpu()  
+
 
+target_dtype = np.int16 # output audio file format expected by Gradio
+max_range = np.iinfo(target_dtype).max
 
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
-    synthesised_speech = synthesise(translated_text)
-    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
-    return 16000, synthesised_speech
+    synthesized_speech = synthesize(translated_text)
+    # normalize audio array by dynamic range of target dtype for Gradio
+    synthesized_speech = (synthesized_speech.numpy() * max_range).astype(target_dtype)
+    return 16000, synthesized_speech
 
 
 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 German. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Facebook's
+[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")
 """