Create app.py

app.py

@@ -0,0 +1,50 @@
+import gradio as gr
+import torch
+from transformers import WhisperProcessor, WhisperForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM
+
+# Load Whisper Odia ASR model (fine-tuned)
+asr_model = WhisperForConditionalGeneration.from_pretrained("ashutoshpattnaik50/whisper-small-odia-finetuned")
+processor = WhisperProcessor.from_pretrained("openai/whisper-small")
+
+# Load IndicTrans2 Odia-to-English model
+translator = AutoModelForSeq2SeqLM.from_pretrained("ai4bharat/indictrans2-en")
+translator_tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indictrans2-en")
+
+# Transcription + Translation function
+def transcribe_and_translate(audio_path):
+    # Load and preprocess audio
+    from datasets import load_dataset, Audio
+    import numpy as np
+    import torchaudio
+
+    speech_array, sampling_rate = torchaudio.load(audio_path)
+    if sampling_rate != 16000:
+        resampler = torchaudio.transforms.Resample(sampling_rate, 16000)
+        speech_array = resampler(speech_array)
+    
+    input_features = processor(speech_array.squeeze().numpy(), sampling_rate=16000, return_tensors="pt").input_features
+
+    predicted_ids = asr_model.generate(input_features)
+    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
+
+    # Translate to English
+    inputs = translator_tokenizer(transcription, return_tensors="pt")
+    output_tokens = translator.generate(**inputs)
+    translated = translator_tokenizer.batch_decode(output_tokens, skip_special_tokens=True)[0]
+
+    return transcription.strip(), translated.strip()
+
+# Gradio Interface
+interface = gr.Interface(
+    fn=transcribe_and_translate,
+    inputs=gr.Audio(type="filepath", label="🎤 Record or Upload Odia Audio"),
+    outputs=[
+        gr.Textbox(label="📝 Odia Transcription"),
+        gr.Textbox(label="🌐 English Translation")
+    ],
+    title="Odia Whisper ASR + Translator",
+    description="Speak in Odia and get instant transcription + English translation using Whisper and IndicTrans2.",
+)
+
+if __name__ == "__main__":
+    interface.launch()