Update app.py

app.py

@@ -9,13 +9,12 @@ from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Proce
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 # load speech translation checkpoint
-asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-tiny", device=device) #тут изменил на tiny
+asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-tiny", device=device) #тут изменил на tiny, потому что это более компактная модель
 
-# load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+transl = pipeline("translation", model="Helsinki-NLP/opus-mt-en-ru")
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+model = VitsModel.from_pretrained("facebook/mms-tts-rus")
+token = AutoTokenizer.from_pretrained("facebook/mms-tts-rus")
 
 embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
 speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
@@ -23,13 +22,15 @@ speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze
 
 def translate(audio):
     outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
-    return outputs["text"]
+    return transl(outputs['text'])[0]['translation_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 synthesise(text: str, tokenizer: AutoTokenizer = token, model: VitsModel = model):
+    inputs = token(text, return_tensors="pt")
+    # print(inputs)
+    with torch.no_grad():
+        output = model(**inputs).waveform
+    return output.cpu()
 
 
 def speech_to_speech_translation(audio):