Create app.py

app.py

@@ -0,0 +1,335 @@
+# app.py - Audio & Text Sentiment Analyzer
+# Transcription: openai/whisper-base.en (official HF version)
+# Sentiment: nlptown/bert-base-multilingual-uncased-sentiment (5-star accurate model)
+
+import gradio as gr
+import torch
+import numpy as np
+import librosa
+from transformers import (
+    AutoProcessor,
+    AutoModelForSpeechSeq2Seq,
+    AutoTokenizer,
+    AutoModelForSequenceClassification
+)
+import torch.nn.functional as F
+
+print("Loading models... Please wait.")
+
+# === Load Whisper exactly as requested ===
+processor = AutoProcessor.from_pretrained("openai/whisper-base.en")
+whisper_model = AutoModelForSpeechSeq2Seq.from_pretrained("openai/whisper-base.en")
+whisper_model.eval()
+device = "cuda" if torch.cuda.is_available() else "cpu"
+whisper_model.to(device)
+
+# === Load Sentiment model exactly as requested ===
+sentiment_tokenizer = AutoTokenizer.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')
+sentiment_model = AutoModelForSequenceClassification.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')
+sentiment_model.eval()
+sentiment_model.to(device)
+
+print("All models loaded successfully!")
+
+
+# Transcribe audio using official Whisper
+def transcribe_audio(audio_path):
+    if audio_path is None:
+        return ""
+
+    try:
+        # Load and resample to 16kHz
+        speech, _ = librosa.load(audio_path, sr=16000)
+
+        # Process input
+        input_features = processor(speech, sampling_rate=16000, return_tensors="pt").input_features
+        input_features = input_features.to(device)
+
+        # Generate transcription
+        with torch.no_grad():
+            predicted_ids = whisper_model.generate(input_features)
+
+        transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
+        return transcription.strip()
+
+    except Exception as e:
+        print(f"Transcription error: {e}")
+        return "[Transcription failed]"
+
+
+# Sentiment analysis with 5-star rating and confidence
+def analyze_sentiment(text):
+    if not text.strip():
+        return "⭐⭐⭐ Neutral", "0%"
+
+    inputs = sentiment_tokenizer(text[:512], return_tensors="pt", truncation=True).to(device)
+
+    with torch.no_grad():
+        logits = sentiment_model(**inputs).logits
+
+    probabilities = F.softmax(logits, dim=-1)[0]
+    predicted_class = torch.argmax(probabilities).item() + 1  # 1 to 5
+    confidence = probabilities[predicted_class - 1].item() * 100
+    conf_str = f"{confidence:.1f}%"
+
+    stars = "⭐" * predicted_class
+    if predicted_class == 1:
+        level = f"{stars} Very Negative"
+    elif predicted_class == 2:
+        level = f"{stars} Negative"
+    elif predicted_class == 3:
+        level = f"{stars} Neutral"
+    elif predicted_class == 4:
+        level = f"{stars} Positive"
+    else:
+        level = f"{stars} Very Positive"
+
+    return level, conf_str
+
+
+# Main unified function
+def analyze_input(audio_path, input_text):
+    # Use typed text if provided
+    if input_text and input_text.strip():
+        final_text = input_text.strip()
+
+    # Otherwise transcribe audio
+    elif audio_path is not None:
+        print("Transcribing audio...")
+        final_text = transcribe_audio(audio_path)
+        if not final_text or "failed" in final_text.lower():
+            return "Transcription failed or no speech detected.", "", "", "Please try again with clearer English audio."
+
+    else:
+        return "No input provided.", "", "", "Please type text or record/upload audio."
+
+    # Sentiment analysis
+    level, confidence = analyze_sentiment(final_text)
+    final_result = f"{level} (Confidence: {confidence})"
+
+    return final_text, level, confidence, final_result
+
+
+# Gradio Interface
+with gr.Blocks(title="Audio & Text Sentiment Analyzer", theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# 🎤✍️ Audio to Text + 5-Star Sentiment Analyzer")
+    gr.Markdown("""
+    - **Transcription**: OpenAI Whisper-base.en (excellent English accuracy)  
+    - **Sentiment**: Multilingual BERT fine-tuned on reviews → accurate **1–5 star** ratings  
+    - Record/upload audio **or** type text directly
+    """)
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            audio_input = gr.Audio(
+                sources=["microphone", "upload"],
+                type="filepath",
+                label="Record or Upload Audio (English recommended)"
+            )
+
+            gr.Markdown("**OR**")
+
+            text_input = gr.Textbox(
+                label="Type or Paste Text",
+                placeholder="Enter your review, feedback, or transcribed text...",
+                lines=6
+            )
+
+    btn = gr.Button("Transcribe & Analyze Sentiment", variant="primary", size="lg")
+
+    with gr.Column():
+        gr.Markdown("### 📝 Transcribed / Entered Text")
+        text_display = gr.Textbox(label="Text", lines=8, interactive=False)
+
+        gr.Markdown("### 🌟 Sentiment Result")
+        with gr.Row():
+            level_out = gr.Textbox(label="Sentiment Level", scale=2)
+            conf_out = gr.Textbox(label="Confidence", scale=1)
+
+        result_out = gr.Textbox(label="Final Verdict", lines=2, interactive=False)
+
+    btn.click(
+        fn=analyze_input,
+        inputs=[audio_input, text_input],
+        outputs=[text_display, level_out, conf_out, result_out]
+    )
+
+    gr.Markdown("""
+    ### Notes
+    - Best performance with **clear English speech**
+    - Sentiment model excels at review-style language (opinions, experiences)
+    - Confidence >80% = very reliable prediction
+    - Runs completely locally — perfect for privacy
+    - Built with ❤️ in Accra by Chris (@chrisbekor99)
+    """)
+
+# Launch
+if __name__ == "__main__":
+    demo.launch(
+        server_name="127.0.0.1",
+        server_port=7860,
+        share=False  # Change to True for public link via ngrok
+    )# app.py - Audio & Text Sentiment Analyzer
+# Transcription: openai/whisper-base.en (official HF version)
+# Sentiment: nlptown/bert-base-multilingual-uncased-sentiment (5-star accurate model)
+
+import gradio as gr
+import torch
+import numpy as np
+import librosa
+from transformers import (
+    AutoProcessor,
+    AutoModelForSpeechSeq2Seq,
+    AutoTokenizer,
+    AutoModelForSequenceClassification
+)
+import torch.nn.functional as F
+
+print("Loading models... Please wait.")
+
+# === Load Whisper exactly as requested ===
+processor = AutoProcessor.from_pretrained("openai/whisper-base.en")
+whisper_model = AutoModelForSpeechSeq2Seq.from_pretrained("openai/whisper-base.en")
+whisper_model.eval()
+device = "cuda" if torch.cuda.is_available() else "cpu"
+whisper_model.to(device)
+
+# === Load Sentiment model exactly as requested ===
+sentiment_tokenizer = AutoTokenizer.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')
+sentiment_model = AutoModelForSequenceClassification.from_pretrained('nlptown/bert-base-multilingual-uncased-sentiment')
+sentiment_model.eval()
+sentiment_model.to(device)
+
+print("All models loaded successfully!")
+
+# Transcribe audio using official Whisper
+def transcribe_audio(audio_path):
+    if audio_path is None:
+        return ""
+
+    try:
+        # Load and resample to 16kHz
+        speech, _ = librosa.load(audio_path, sr=16000)
+
+        # Process input
+        input_features = processor(speech, sampling_rate=16000, return_tensors="pt").input_features
+        input_features = input_features.to(device)
+
+        # Generate transcription
+        with torch.no_grad():
+            predicted_ids = whisper_model.generate(input_features)
+
+        transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
+        return transcription.strip()
+
+    except Exception as e:
+        print(f"Transcription error: {e}")
+        return "[Transcription failed]"
+
+# Sentiment analysis with 5-star rating and confidence
+def analyze_sentiment(text):
+    if not text.strip():
+        return "⭐⭐⭐ Neutral", "0%"
+
+    inputs = sentiment_tokenizer(text[:512], return_tensors="pt", truncation=True).to(device)
+
+    with torch.no_grad():
+        logits = sentiment_model(**inputs).logits
+
+    probabilities = F.softmax(logits, dim=-1)[0]
+    predicted_class = torch.argmax(probabilities).item() + 1  # 1 to 5
+    confidence = probabilities[predicted_class - 1].item() * 100
+    conf_str = f"{confidence:.1f}%"
+
+    stars = "⭐" * predicted_class
+    if predicted_class == 1:
+        level = f"{stars} Very Negative"
+    elif predicted_class == 2:
+        level = f"{stars} Negative"
+    elif predicted_class == 3:
+        level = f"{stars} Neutral"
+    elif predicted_class == 4:
+        level = f"{stars} Positive"
+    else:
+        level = f"{stars} Very Positive"
+
+    return level, conf_str
+
+# Main unified function
+def analyze_input(audio_path, input_text):
+    # Use typed text if provided
+    if input_text and input_text.strip():
+        final_text = input_text.strip()
+
+    # Otherwise transcribe audio
+    elif audio_path is not None:
+        print("Transcribing audio...")
+        final_text = transcribe_audio(audio_path)
+        if not final_text or "failed" in final_text.lower():
+            return "Transcription failed or no speech detected.", "", "", "Please try again with clearer English audio."
+
+    else:
+        return "No input provided.", "", "", "Please type text or record/upload audio."
+
+    # Sentiment analysis
+    level, confidence = analyze_sentiment(final_text)
+    final_result = f"{level} (Confidence: {confidence})"
+
+    return final_text, level, confidence, final_result
+
+# Gradio Interface
+with gr.Blocks(title="Audio & Text Sentiment Analyzer", theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# 🎤✍️ Audio to Text + 5-Star Sentiment Analyzer")
+    gr.Markdown("""
+    - **Transcription**: OpenAI Whisper-base.en (excellent English accuracy)  
+    - **Sentiment**: Multilingual BERT fine-tuned on reviews → accurate **1–5 star** ratings  
+    - Record/upload audio **or** type text directly
+    """)
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            audio_input = gr.Audio(
+                sources=["microphone", "upload"],
+                type="filepath",
+                label="Record or Upload Audio (English recommended)"
+            )
+
+            gr.Markdown("**OR**")
+
+            text_input = gr.Textbox(
+                label="Type or Paste Text",
+                placeholder="Enter your review, feedback, or transcribed text...",
+                lines=6
+            )
+
+    btn = gr.Button("Transcribe & Analyze Sentiment", variant="primary", size="lg")
+
+    with gr.Column():
+        gr.Markdown("### 📝 Transcribed / Entered Text")
+        text_display = gr.Textbox(label="Text", lines=8, interactive=False)
+
+        gr.Markdown("### 🌟 Sentiment Result")
+        with gr.Row():
+            level_out = gr.Textbox(label="Sentiment Level", scale=2)
+            conf_out = gr.Textbox(label="Confidence", scale=1)
+
+        result_out = gr.Textbox(label="Final Verdict", lines=2, interactive=False)
+
+    btn.click(
+        fn=analyze_input,
+        inputs=[audio_input, text_input],
+        outputs=[text_display, level_out, conf_out, result_out]
+    )
+
+    gr.Markdown("""
+    ### Notes
+    - Best performance with **clear English speech**
+    - Sentiment model excels at review-style language (opinions, experiences)
+    - Confidence >80% = very reliable prediction
+    - Runs completely locally — perfect for privacy
+    - Built with ❤️ in Accra by Chris (@chrisbekor99)
+    """)
+
+# Run app
+if __name__ == "__main__":
+    demo.launch()