Upload 2 files

app.py

@@ -0,0 +1,136 @@
+import gradio as gr
+import torch
+import torchaudio
+from transformers import AutoModelForCTC, AutoProcessor, pipeline
+from pydub import AudioSegment
+import numpy as np
+import librosa
+import io
+import tempfile
+
+# Load ASR model and processor for Hindi speech recognition
+print("Loading ASR model...")
+asr_processor = AutoProcessor.from_pretrained("ai4bharat/indicwav2vec-hindi")
+asr_model = AutoModelForCTC.from_pretrained("ai4bharat/indicwav2vec-hindi")
+
+# Load sentiment analysis pipeline for Hindi text
+print("Loading sentiment analysis model...")
+sentiment_pipeline = pipeline(
+    "text-classification",
+    model="LondonStory/txlm-roberta-hindi-sentiment",
+    return_all_scores=True
+)
+
+# Move models to appropriate device (CPU for free Hugging Face Space)
+device = "cuda" if torch.cuda.is_available() else "cpu"
+asr_model.to(device)
+print(f"Models loaded on device: {device}")
+
+def predict(audio_filepath):
+    """
+    Main prediction function that processes Hindi audio and returns sentiment analysis.
+    
+    Args:
+        audio_filepath: Path to the uploaded audio file
+    
+    Returns:
+        Dictionary with sentiment labels and confidence scores
+    """
+    try:
+        # Load and preprocess audio
+        print(f"Processing audio file: {audio_filepath}")
+        
+        # Load audio using librosa and resample to 16kHz as required by the ASR model
+        audio_array, sample_rate = librosa.load(audio_filepath, sr=16000)
+        
+        # Ensure audio is in the correct format
+        if len(audio_array.shape) > 1:
+            audio_array = np.mean(audio_array, axis=1)
+        
+        # Process audio with ASR processor
+        inputs = asr_processor(
+            audio_array,
+            sampling_rate=16000,
+            return_tensors="pt",
+            padding=True
+        )
+        
+        # Move inputs to device
+        inputs = {k: v.to(device) for k, v in inputs.items()}
+        
+        # Transcribe audio to Hindi text
+        with torch.no_grad():
+            logits = asr_model(**inputs).logits
+        
+        # Get predicted token IDs
+        predicted_ids = torch.argmax(logits, dim=-1)
+        
+        # Decode the transcription
+        transcription = asr_processor.batch_decode(predicted_ids)[0]
+        
+        print(f"Transcribed text: {transcription}")
+        
+        # Handle empty transcription
+        if not transcription.strip():
+            return {"error": "Could not transcribe audio. Please ensure you're speaking in Hindi."}
+        
+        # Perform sentiment analysis on the transcribed text
+        sentiment_results = sentiment_pipeline(transcription)
+        
+        # Format results for Gradio
+        result_dict = {}
+        for result in sentiment_results[0]:
+            label = result['label']
+            score = result['score']
+            result_dict[label] = float(score)
+        
+        # Add transcription to results for reference
+        result_dict['Transcription'] = transcription
+        
+        return result_dict
+        
+    except Exception as e:
+        print(f"Error processing audio: {str(e)}")
+        return {"error": f"Error processing audio: {str(e)}"}
+
+# Create Gradio interface
+demo = gr.Interface(
+    fn=predict,
+    inputs=gr.Audio(
+        type="filepath", 
+        label="Upload Hindi Speech",
+        sources=["upload", "microphone"]
+    ),
+    outputs=gr.Label(
+        label="Sentiment Analysis Result",
+        num_top_classes=3
+    ),
+    title="🎤 Hindi Speech Sentiment Analysis",
+    description="""
+    ### Upload or record Hindi audio to analyze sentiment
+    
+    This app performs the following steps:
+    1. **Speech Recognition**: Converts your Hindi speech to text using AI4Bharat's IndicWav2Vec model
+    2. **Sentiment Analysis**: Analyzes the emotional tone using a specialized Hindi sentiment model
+    
+    **Instructions**:
+    - Upload an audio file or record directly using the microphone
+    - Speak clearly in Hindi for best results
+    - The app will show both the transcribed text and sentiment scores
+    
+    **Supported sentiments**: Positive, Negative, and Neutral with confidence scores
+    """,
+    examples=None,
+    theme=gr.themes.Soft(),
+    allow_flagging="never"
+)
+
+# Launch the app
+if __name__ == "__main__":
+    # Launch with share=True for public access, queue for handling multiple requests
+    demo.launch(
+        share=False,  # Set to True if you want a public link for testing
+        server_name="0.0.0.0",  # Required for Hugging Face Spaces
+        server_port=7860,  # Default port for Hugging Face Spaces
+        show_error=True
+    )

requirements.txt

@@ -0,0 +1,6 @@
+transformers
+torch
+torchaudio
+datasets
+gradio
+pydub