Update app.py

app.py

@@ -1,35 +1,76 @@
+import gradio as gr
 import spaces
-from faster_whisper import WhisperModel
+import torch
+from transformers import pipeline
+import os
 
-# 1. Initialize the model on CPU globally. 
-# ZeroGPU boots your script on a CPU-only node first.
+# आपका पसंदीदा और सटीक हिंदी मॉडल
+MODEL_ID = "collabora/faster-whisper-large-v2-hindi"
 
-
-
-# 2. Wrap the inference logic with the ZeroGPU decorator.
-# You can increase the duration (in seconds) if you process long audio files.
-@spaces.GPU(duration=90)
-def transcribe_audio(audio_path):
-    # Dynamically change the device to CUDA during the ZeroGPU allocation window
-    model = WhisperModel(
-        "collabora/faster-whisper-large-v2-hindi", 
-        device="cuda", 
-        compute_type="float32"
+@spaces.GPU(duration=120)
+def transcribe_hindi(audio):
+    if audio is None:
+        return "कृपया ऑडियो प्रदान करें।", None
+    
+    # Standard Transformers Pipeline - ZeroGPU के CUDA ड्राइवर के साथ 100% अनुकूल
+    pipe = pipeline(
+        "automatic-speech-recognition",
+        model=MODEL_ID,
+        torch_dtype=torch.float16,
+        device="cuda"
     )
     
-    segments, info = model.transcribe(audio_path)
+    # ट्रांसक्रिप्शन जनरेट करें
+    result = pipe(audio, generate_kwargs={"language": "hindi"})
+    text_output = result["text"].strip()
     
-    # We consume the generator inside the GPU function to prevent losing the GPU context
-    results = []
-    for segment in segments:
-        results.append("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
+    # डाउनलोड के लिए .txt फ़ाइल बनाना
+    file_path = "transcription.txt"
+    with open(file_path, "w", encoding="utf-8") as f:
+        f.write(text_output)
         
-    # Optional: Clean up and move model back to CPU when done
-    model.model.to("cpu")
-    return results
+    return text_output, file_path
+
+custom_css = """
+footer {visibility: hidden}
+.gradio-container {background-color: #fcfcfc}
+#header {text-align: center; margin-bottom: 20px}
+"""
 
-# 3. Trigger your transcription
-output_segments = transcribe_audio("Mytest000.m4a")
+with gr.Blocks(title="IndicWhisper Collabora GPU") as demo:
+    gr.HTML("<div id='header'><h1>🎙️ Hindi Whisper (Collabora v2 - ZeroGPU)</h1></div>")
+    
+    with gr.Row():
+        with gr.Column():
+            audio_input = gr.Audio(
+                sources=["microphone", "upload"], 
+                type="filepath", 
+                label="ऑडियो रिकॉर्ड करें या अपलोड करें"
+            )
+            submit_btn = gr.Button("अनुवाद करें (Transcribe)", variant="primary")
+            
+        with gr.Column():
+            output_text = gr.Textbox(
+                label="Transcription Output", 
+                lines=10, 
+                placeholder="आपका टेक्स्ट यहाँ दिखाई देगा..."
+            )
+            # डाउनलोड लिंक के लिए Gradio का File कंपोनेंट
+            download_file = gr.File(
+                label="टैक्स्ट फ़ाइल डाउनलोड करें",
+                visible=True
+            )
+            
+    gr.Markdown("""
+    ---
+    **सुझाव:** यहाँ से प्राप्त आउटपुट को कॉपी करें या फ़ाइल डाउनलोड करके अपने **Gemini Gem** में पेस्ट करें ताकि **पञ्चमाक्षर नियमों** (ङ्, ञ्, ण्, न्, म्) के अनुसार शुद्धिकरण किया जा सके।
+    """)
+
+    submit_btn.click(
+        fn=transcribe_hindi, 
+        inputs=audio_input, 
+        outputs=[output_text, download_file]
+    )
 
-for line in output_segments:
-    print(line)
+if __name__ == "__main__":
+    demo.launch(css=custom_css)