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("omthkkr/speecht5_finetuned_voxpopuli_sl")
@@ -20,6 +20,18 @@ vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(devic
 embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
 speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
 
+replacements = [
+    ("č", "c"),
+    ("š", "s"),
+    ("ž", "z"),
+]
+
+
+def cleanup_text(inputs):
+    for src, dst in replacements:
+        inputs = inputs.replace(src, dst)
+    return inputs
+
 
 def translate(audio):
     outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "sl"})
@@ -27,6 +39,7 @@ def translate(audio):
 
 
 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()