Update app.py

app.py

@@ -9,7 +9,7 @@ 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-base", device=device)
+asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-large-v2", device=device)
 
 # load text-to-speech checkpoint and speaker embeddings
 processor = SpeechT5Processor.from_pretrained("ckandemir/speecht5_finetuned_voxpopuli_fr")
@@ -21,12 +21,41 @@ embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validat
 speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
 
+replacements = [
+     ("á", "a"),
+     ("ç", "c"),
+     ("è", "e"),
+     ("ì", "i"),
+     ("í", "i"),
+     ("ò", "o"),
+     ("ó", "o"),
+     ("ù", "u"),
+     ("ú", "u"),
+     ("š", "s"),
+     ("ï", "i"),
+     ("ñ", "n"),
+     ("ü", "u"),
+]
+
+def cleanup_text(text):
+    for src, dst in replacements:
+        text = text.replace(src, dst)
+    return text
+
+def synthesize_speech(text):
+    text = cleanup_text(text)
+    inputs = processor(text=text, return_tensors="pt")
+    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
+
+    return gr.Audio.update(value=(16000, speech.cpu().numpy()))
+
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "fr"})
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "french"})
     return outputs["text"]
 
 
 def synthesise(text):
+    text = cleanup_text(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()