Update app.py

app.py

@@ -1,7 +1,162 @@
 import gradio as gr
+import whisper
+from transformers import pipeline
+import pandas as pd, numpy as np
 import os
+import torchaudio
+import librosa
+from scipy.io.wavfile import write
+import shutil
+import soundfile as sf
+import noisereduce as nr
+from scipy.stats import skew
+from tqdm import tqdm
+import requests
+import pickle
+import dash
+import dash_bootstrap_components as dbc
+from dash import html
 
-hf_token = os.environ.get('HF_TOKEN')
+sr = 8000
 
-iface = gr.load(name="aslanovaf/Sentiment_Analysis_Azerbaijani", hf_token=hf_token, src="spaces")
-iface.queue(api_open=False).launch(show_api=False)
+url = "https://huggingface.co/spaces/aslanovaf/Sentiment_Analysis_Azerbaijani/resolve/main/sentiment_model_8000.pickle"
+hf_token = os.environ.get("HF_TOKEN")
+headers = {"Authorization": f"Bearer {hf_token}"}
+
+response = requests.get(url, headers=headers) 
+if response.status_code == 200:
+    model = pickle.loads(response.content)
+else:
+    st.markdown(f"Failed to download TTS from {url} (Status code: {response.status_code})")
+
+
+def split_full_audio_15_sec(audio_file):
+    audio, orig_sr = sf.read(audio_file)
+    audio = librosa.resample(y=audio, orig_sr=orig_sr, target_sr=sr)
+    
+    chunk_length = 15 * sr
+    total_length = len(audio)
+    start_index = 0
+    end_index = min(chunk_length, total_length)
+    f = 0
+    chunks = []
+
+    while start_index < total_length:
+        chunk = audio[start_index:end_index]
+        chunk_name = f"example_{f}.wav"
+        chunk_duration = len(chunk)/sr
+        if chunk_duration<3:
+            break
+        chunks.append(chunk)
+        start_index = end_index
+        end_index = min(end_index + chunk_length, total_length)
+        f+=1
+    return chunks
+
+def get_mfcc(name):
+    resampled_audio = name
+    
+    try:
+        reduced_noise = nr.reduce_noise(resampled_audio, sr=sr)
+        ft1 = librosa.feature.mfcc(y=reduced_noise, sr = sr, n_mfcc=16)
+        ft2 = librosa.feature.zero_crossing_rate(reduced_noise)[0]
+        ft3 = librosa.feature.spectral_rolloff(y=reduced_noise)[0]
+        ft4 = librosa.feature.spectral_centroid(y=reduced_noise)[0]
+        ft1_trunc = np.hstack((np.mean(ft1, axis=1), np.std(ft1, axis=1), skew(ft1, axis = 1), np.max(ft1, axis = 1), np.min(ft1, axis = 1)))
+        ft2_trunc = np.hstack((np.mean(ft2), np.std(ft2), skew(ft2), np.max(ft2), np.min(ft2)))
+        ft3_trunc = np.hstack((np.mean(ft3), np.std(ft3), skew(ft3), np.max(ft3), np.min(ft3)))
+        ft4_trunc = np.hstack((np.mean(ft4), np.std(ft4), skew(ft4), np.max(ft4), np.min(ft4)))
+        return pd.Series(np.hstack((ft1_trunc, ft2_trunc, ft3_trunc, ft4_trunc)))
+    except:
+        print('bad file')
+        return pd.Series([0]*95)
+
+    
+def analyze_sentiment(audio):
+    chunks = split_full_audio_15_sec(audio)
+    chunked_df = pd.DataFrame(data={'Chunk_order': [f'Chunk_{i+1}' for i in range(len(chunks))], 'Data': chunks})
+    df_features = chunked_df['Data'].apply(get_mfcc)
+    df = pd.concat([chunked_df, df_features], axis=1)
+    df = df.drop(columns=['Data'])
+    df.columns = ['Chunk_order']+[f'Feature_{i+1}' for i in range(95)]
+    df['Prediction'] = model.predict(df.drop(columns=['Chunk_order']))
+
+    df['Prediction'] = df['Prediction'].map({
+    'pozitive_normal':'Normal',
+    'scope':'Silence',
+    'neqativ':'Negative'
+    })
+
+    clean_df = df[['Chunk_order', 'Prediction']]
+    predictions = df['Prediction'].tolist()
+    final_prediction = 'Negative' if 'Negative' in predictions else 'Normal' if 'Normal' in predictions else 'Silence'
+    final_prediction_2x = 'Negative' if predictions.count('Negative')>1 else 'Normal' if 'Normal' in predictions else 'Silence'
+
+    color_map = {
+        'Normal':'success',
+        'Silence': 'warning',
+        'Negative': 'danger'
+    }
+    
+    return (', '.join(predictions), final_prediction)
+
+
+title = """<h1 align="center">🎤 Azerbaijani Audio Speech Sentiment Analysis 💬</h1>"""
+image_path = "thmbnail.jpg"
+description = """
+💻 This demo showcases a general-purpose sentiment analysis process. It is trained on a collection of audio calls from banking/fintech industries based on audio features. The main analysis predicts one of the categories (Normal/Negative/Silence) for each 15-second bucket in the audio. The final category for the whole audio is also estimated. 
+<br>
+⚙️ Components of the tool:<br>
+<br>
+&nbsp;&nbsp;&nbsp;&nbsp; - Sentiment analysis directly of the audios.<br>
+<br>
+❓ Use the microphone for real-time audio recording.<br>
+↑ Or upload an audio file.<br>
+<br>
+
+⚡️ The model will extract audio features and perform sentiment analysis on the audio.<br>
+
+"""
+
+custom_css = """
+#banner-image {
+    display: block;
+    margin-left: auto;
+    margin-right: auto;
+}
+#chat-message {
+    font-size: 14px;
+    min-height: 300px;
+}
+"""
+
+block = gr.Blocks(css=custom_css)
+
+
+with block:
+    gr.HTML(title)
+
+    with gr.Row():
+        with gr.Column():
+            gr.HTML(description)
+        with gr.Column():
+            gr.Image(image_path, elem_id="banner-image", show_label=False)
+
+    gr.Interface(
+        fn=analyze_sentiment,
+        inputs=[
+            gr.Audio(sources=["upload", "microphone"], type="filepath", label="Input Audio"),
+        ],
+        outputs=[gr.Textbox(label="Sentiment Analysis Results of 15-second buckets"),gr.Textbox(label="Final Prediction")],
+        # layout="vertical",
+        # theme="huggingface",
+        examples=[
+            ["./Recording_1.wav", "analyze_sentiment"],
+            ["./Recording_2.wav", "analyze_sentiment"],
+        ],
+        cache_examples=True,
+        allow_flagging="never",
+    )
+    # gr.TabbedInterface([mic, file], ["Audio from Microphone", "Audio from File"])
+
+block.launch()