sentiment analysis added :4head:

app.py

@@ -37,14 +37,14 @@ from pydub import AudioSegment
 def predict(audio_file_path):
     if(audio_file_path == None):
         output = "Please record your voice using the record button before submitting :)"
-        return output, {}, ""
+        return output, {}, {}, ""
     
     input_data = open(audio_file_path, 'rb').read()
     print(len(input_data))
     
     if(len(input_data) == 88108 or len(input_data) == 94252):
         output = "It appears your recording device isn't supported by Hugging Face/Gradio yet (iOS and macOS are causing issues). Windows and android record properly, sorry for the temporary inconvenience!"
-        return output, {}, ""
+        return output, {}, {}, ""
 
     # Set the content type
     headers = {'Content-Type': 'application/json'}
@@ -61,15 +61,30 @@ def predict(audio_file_path):
             confArray = predictions[label]
             avg = sum(confArray) / len(confArray)
             confs[labels[label]] = avg
-
+        
+        sentiments = obj['toxicity_predictions']
+        sentiment_labels = {'toxicity':'Toxic', 'severe_toxicity':'Severe Toxicity', 'obscene':'Obscene', 'threat':'Threat', 'insult':'Insult', 'identity_attack':'Identity Hate', 'sexual_explicit':'Sexually Explicit'}
+        sentiment_confs = {}
+        detected_toxicity = False
+        for s in sentiment_labels.keys():
+            sentiment_conf = sentiments[s]
+            if float(sentiment_conf) > 0.01:
+                detected_toxicity = True
+            sentiment_confs[sentiment_labels[s]] = sentiment_conf
+        
+        if detected_toxicity:
+            sentiment_confs['Not Toxic'] = "0.0"
+        else:
+            sentiment_confs['Not Toxic'] = "0.99"
+        
         output = "Audio processed successfully."
-        return output, confs, obj['whisper'].get('text')
+        return output, confs, sentiment_confs, obj['whisper'].get('text')
     except JSONDecodeError as e:
         if "viable" in resp.text or "detected" in resp.text:
             output = "No viable audio detected within your clip! Make sure the clip you recorded is audible!"
         else:
             output = "Our servers are currently overloaded, try again in a few minutes."       
-    return output, {}, ""
+    return output, {}, {}, ""
 
 btn_label_dict = {'Child': 'child_unknown', 'Teen Female': 'teens_female', 'Teen Male':'teens_male', 'Adult Female':'twenties+_female', 'Adult Male':'twenties+_male'}
 
@@ -103,8 +118,9 @@ with gr.Blocks() as demo:
                 submit_btn = gr.Button("Submit")
         with gr.Column(scale=1):
             resp = gr.Textbox(label="Response")
-            labels = gr.Label(num_top_classes=5, label="Prediction confidences")
             words = gr.Textbox(label="Detected words")
+            labels2 = gr.Label(num_top_classes=7, label="Sentiment analysis")
+            labels = gr.Label(num_top_classes=5, label="Demographic confidences")
             flag_btn = gr.Button("Flag as incorrect", visible=False)
             with gr.Row(visible=False) as flag_options:
                 with gr.Row():
@@ -147,9 +163,9 @@ with gr.Blocks() as demo:
     def trigger_predict(audio):
         print("triggering prediction")
         # options = hide_flagging_options()
-        output, confs, words = predict(audio)
+        output, confs, sentiments, words = predict(audio)
         btn = show_main_flag_btn()
-        return output, confs, words, btn
+        return output, confs, sentiments, words, btn
     
     ex = gr.Examples(
             examples=example_list, 
@@ -160,7 +176,7 @@ with gr.Blocks() as demo:
     submit_btn.click(
         fn = trigger_predict,
         inputs=audio,
-        outputs=[resp, labels, words, flag_btn]
+        outputs=[resp, labels, labels2, words, flag_btn]
     )
     child_flag_btn.click(
         fn=send_flagged_feedback,