Create app.py

app.py

@@ -0,0 +1,110 @@
+from datasets import load_dataset
+from transformers import pipeline
+import soundfile as sf
+import torch
+import gradio as gr
+import numpy as np
+
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+
+model_name = "Helsinki-NLP/opus-mt-en-ar"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
+
+def translate_to_arabic(text):
+    inputs = tokenizer(text, return_tensors="pt", padding=True)
+    outputs = model.generate(**inputs, max_length=100)
+    translated = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return translated
+
+# Test it
+print(translate_to_arabic("Hello, how are you?"))
+
+def predict_image(image):
+  pipe = pipeline("image-classification", model="google/vit-base-patch16-224")
+  ClassifedImage=pipe(image)
+
+  result=ClassifedImage[0]['label']
+  return result
+
+def translate_to_arabic(text):
+  pipe = pipeline("translation", model="Helsinki-NLP/opus-mt-en-ar")
+  result=pipe(text , max_length=100)
+  return result[0]['translation_text']
+
+# Use a pipeline as a high-level helper
+# Warning: Pipeline type "translation" is no longer supported in transformers v5.
+# You must load the model directly (see below) or downgrade to v4.x with:
+# 'pip install "transformers<5.0.0'
+from transformers import pipeline
+
+pipe = pipeline("translation", model="Helsinki-NLP/opus-mt-en-ar")
+
+def translate_to_arabic(text):
+    inputs = tokenizer(text, return_tensors="pt", padding=True)
+    outputs = model.generate(**inputs, max_length=100)
+    translated = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return translated
+
+"""✅ Because SpeechT5 requires a speaker embedding
+
+The model MBZUAI/speecht5_tts_clartts_ar is based on SpeechT5, and SpeechT5 needs a speaker embedding to generate speech.
+
+Think of it like this:
+
+The text tells the model what to say.
+
+The speaker embedding tells the model what the voice should sound like.
+
+Without a speaker embedding, the model does not know what voice to use → and the pipeline will fail or produce wrong audio.
+
+What is the dataset?
+
+herwoww/arabic_xvector_embeddings contains pre-computed speaker embeddings (x-vectors) for different Arabic speakers.
+
+Each embedding is like a "voice fingerprint".
+
+You pick one of them to generate speech in that voice.
+
+In your code, you choose speaker number 105:
+"""
+
+def text_to_speech(text):
+  pipe = pipeline("text-to-speech", model="MBZUAI/speecht5_tts_clartts_ar")
+  embedding_dataset=load_dataset("herwoww/arabic_xvector_embeddings" , split="validation") #Those embeddings represent a speaker’s voice characteristics
+  '''
+   This is a 1D tensor.
+   But most models expect: (batch_size, input_size)
+   torch.Size([1, 3]) 1 sample (batch size = 1)
+   (784,)>> (1, 784)>> (batch_size, 784)
+  '''
+  speaker_embedding=torch.tensor(embedding_dataset[100]['speaker_embeddings']).unsqueeze(0) ##It becomes a 2-D tensor
+  speech=pipe(text , forward_params={'speaker_embeddings':speaker_embedding})
+
+  return (speech['sampling_rate'],np.array(speech['audio'], dtype=np.float32))
+
+from PIL import Image
+with gr.Blocks() as app:
+  gr.Markdown("Image Classification, Arabic Translation, TTS")
+
+  with gr.Row():
+    with gr.Column():
+      image_input=gr.Image(type="pil",label="Upload the Image to classify it" )
+      classify_image=gr.Button("Classify the Image")
+      pred=gr.Textbox(label="Classifcation Result")
+
+    classify_image.click(fn=predict_image , inputs=image_input , outputs=pred)
+
+  with gr.Row():
+    translated_output=gr.Textbox(label="Translated Text")
+    translate_btn=gr.Button("Translate to Arabic")
+
+    translate_btn.click(fn=translate_to_arabic , inputs=pred , outputs=translated_output)
+
+  with gr.Row():
+    tts_btn=gr.Button("Convert to Speech")
+    audio_output=gr.Audio(label="Audio Output")
+
+    tts_btn.click(fn=text_to_speech , inputs=translated_output , outputs=audio_output)
+
+app.launch()