Update app.py

app.py

@@ -2,31 +2,54 @@ import gradio as gr
 import numpy as np
 import torch
 from datasets import load_dataset
-
 from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
 
 
 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("Ellight/speecht5_finetuned_voxpopuli_nl")
-
-model = SpeechT5ForTextToSpeech.from_pretrained("Ellight/speecht5_finetuned_voxpopuli_nl").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
-
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation",trust_remote_code=True)
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+model_id = "Ellight/speecht5_finetuned_voxpopuli_nl"  # update with your model id
+# pipe = pipeline("automatic-speech-recognition", model=model_id)
+model = SpeechT5ForTextToSpeech.from_pretrained(model_id)
+vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+speaker_embeddings = torch.tensor(embeddings_dataset[7440]["xvector"]).unsqueeze(0)
+
+processor = SpeechT5Processor.from_pretrained(model_id)
+
+replacements = [
+    ("à", "a"),
+    ("ç", "c"),
+    ("è", "e"),
+    ("ë", "e"),
+    ("í", "i"),
+    ("ï", "i"),
+    ("ö", "o"),
+    ("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": "translate"})
+    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "Dutch"})
     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()
@@ -41,9 +64,8 @@ def speech_to_speech_translation(audio):
 
 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 Dutch. Demo uses OpenAI's [Whisper Large v2](https://huggingface.co/openai/whisper-large-v2) model for speech translation, and [Sandiago21/speecht5_finetuned_voxpopuli_it](https://huggingface.co/Sandiago21/speecht5_finetuned_voxpopuli_it) checkpoint for text-to-speech, which is based on Microsoft's
+[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech, fine-tuned in Dutch Audio dataset:
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 """
 
@@ -69,4 +91,4 @@ file_translate = gr.Interface(
 with demo:
     gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
 
-demo.launch()
+demo.launch()