Update app.py

app.py

@@ -11,6 +11,8 @@ 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)
+translator = pipeline("translation", model="Helsinki-NLP/opus-mt-en-ru")
+
 
 model = VitsModel.from_pretrained('facebook/mms-tts-rus').to(device)
 tokenizer = VitsTokenizer.from_pretrained('facebook/mms-tts-rus')
@@ -18,8 +20,8 @@ tokenizer = VitsTokenizer.from_pretrained('facebook/mms-tts-rus')
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe","language":"russian"})
-    return outputs["text"]
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
+    return translator(outputs['text'])[0]['translation_text']
 
 
 def synthesise(text):
@@ -33,12 +35,14 @@ 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
+    return 16000, synthesised_speech[0]
 
 
 title = "Cascaded STST"
 description = """
-Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Russian"""
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Russian
+![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
+"""
 
 demo = gr.Blocks()