update text

app.py

@@ -152,16 +152,54 @@ def predict(audio, start, end, model_name="BirdAST_Seq"):
 
 
 DESCRIPTION = """
+
+<div align="center">
+<b>Team Members: </b>
+
+Amro Abdrabo [[email protected] | [LinkedIn](https://www.linkedin.com/in/amroabdrabo/)]
+
+Shiyi Li  [[email protected] | [LinkedIn](www.linkedin.com/in/shiyili01)]
+
+Thomas Radinger [ [email protected] | [LinkedIn](https://www.linkedin.com/in/thomas-radinger-743958142/) ]
+</div>
+
 # Introduction 
 
-It is esimated that 50% of the global economy is threatened by biodiversity loss [2]. As such, intensive efforts have been concerted into estimating bird biodiversity, as birds are a top indicator of biodiversity in the region. One of these efforts is 
-finding the bird species in a region using bird species audio classification. 
+Birds are key indicators of ecosystem health and play pivotal roles in maintaining biodiversity [1]. To monitor and protect bird species, automatic bird sound recognition systems are essential. These systems can help in identifying bird species, monitoring their populations, and understanding their behavior. However, building such systems is challenging due to the diversity of bird sounds, complex acoustic interference and limited labeled data. 
+
+To tackle these challenges, we expored the potential of deep learning models for bird sound recognition. In our work, we developed two Audio Spectrogram Transformer (AST) based models: BirdAST and BirdAST_Seq, to predict bird species from audio recordings. We evaluated the models on a dataset of 728 bird species and achieved promising results. As the field-recordings may contain various types of audio rather than only bird songs/calls, we also employed an Audio Masked AutoEncoder (AudioMAE) model to pre-classify audio clips into bird, insects, rain, environmental noise, and other types. Details of the models and evaluation results are provided in the table below.
+
+Our contributions have shown the potential of deep learning models for bird sound recognition. We hope that our work can contribute to the development of automatic bird sound recognition systems and help in monitoring and protecting bird species.
+
+
+<div align="center">
+
+| Model name       | Architecture                   | ROC-AUC Score |
+| ---------------  |:------------------------------:|:-------------:|
+| BirdAST          | AST* + MLP                     | 0.6825        |
+| BirdAST_Seq      | AST* + Sequence Pooling + MLP  | 0.7335        |
+
+</div>
+
+# How to use the space:
+
+1. Choose a model from the dropdown list. It will download the model weights automatically if not already downloaded (~30 seconds).
+2. Upload an audio clip and specify the start and end time for prediction. 
+3. Click on the "Predict" button to get the predictions.
+4. In the output, you will get the audio type classification (e.g., bird, insects, rain, etc.) in the panel "Class Prediction" and the predicted bird species in the panel "Species Prediction". 
+- The audio types are predicted as multi-lable classification based on the AudioMAE model [2]. The predicted classes indicate the possible presence of different types of audio in the recording.
+- The bird species are predicted as a multi-class classification using the selected model. The predicted classes indicate the most possible bird species present in the recording. 
+5. The waveform and spectrogram of the audio clip are displayed in the respective panels.
+
+Notes:
+- For an unknown bird species, the model may predict the most similar bird species based on the training data.
+- If an audio clip contains non-bird sounds (predicted by the AudioMAE), the bird species prediction may not be accurate.
 
-# Solution
+**Disclaimer**: The model predictions are based on the training data and may not be accurate for all audio clips. The model is trained on a dataset of 728 bird species and may not generalize well to all bird species.
 
-To tackle this problem, we propose VOJ. It first preprocesses an audio signal using a bandpass filter (1K - 8K) and then applies downsampling to 16K Hz. Afterwards, we input the signal into AudioMAE (Audio Masked AutoEncoder by Meta [1]) which extracts relevant features even in the presence of corruptions to the signal spectrogram. 
-The AudioMAE is also trained on 527 types of audio that comprise bird, silence, environmental noise, and other types. The purpose of this initial inference stage is to provide an initial sense of the audio. If the AudioMAE outputs silence, we can expect low species prediction confidence, or if the output is insect, it may not be worth labelling. 
-Next, we train BirdAST, which has Audio Spectrogram Transformer (AST) as backbone, followed by an attention pooling and dense layer. We also train EfficientB0 on the melspectrogram, and finally, we train a model using Wav2Vec pretrained on 50 bird species [3].   
+<div align="center">
+  <b>Enjoy the Bird Songs! 🐦🎶
+</div>
 """
 
 
@@ -255,11 +293,11 @@ function createGradioAnimation() {
 """
 
 REFERENCES = """
-References
+# References
 
-[1] Huang, P.-Y., Xu, H., Li, J., Baevski, A., Auli, M., Galuba, W., Metze, F., & Feichtenhofer, C. (2022). Masked Autoencoders that Listen. In NeurIPS.
+[1] Torkington, S. (2023, February 7). 50% of the global economy is under threat from biodiversity loss. World Economic Forum. Retrieved from https://www.weforum.org/agenda/2023/02/biodiversity-nature-loss-cop15/. 
 
-[2] Torkington, S. (2023, February 7). 50% of the global economy is under threat from biodiversity loss. World Economic Forum. Retrieved from https://www.weforum.org/agenda/2023/02/biodiversity-nature-loss-cop15/. 
+[2] Huang, P.-Y., Xu, H., Li, J., Baevski, A., Auli, M., Galuba, W., Metze, F., & Feichtenhofer, C. (2022). Masked Autoencoders that Listen. In NeurIPS.
 
 [3] https://www.kaggle.com/code/dima806/bird-species-by-sound-detection
 """
@@ -286,7 +324,7 @@ def handle_model_selection(model_name, download_status):
                 break
             
         if download_flag:
-            download_status =  f"Model {model_name} is ready for prediction!"
+            download_status =  f"Model <{model_name}> is ready for prediction!"
         else:
             download_status = f"An error occurred while downloading model weights."
             
@@ -303,7 +341,7 @@ with gr.Blocks(theme = seafoam, css = css, js = js) as demo:
     gr.Markdown(DESCRIPTION)
     
     # add dropdown for model selection
-    model_names = ['BirdAST', 'BirdAST_Seq', 'EfficientNet']
+    model_names = ['BirdAST', 'BirdAST_Seq'] #, 'EfficientNet']
     model_dropdown = gr.Dropdown(label="Choose a model", choices=model_names)
     download_status = gr.Textbox(label="Model Status", lines=3, value='', interactive=False) # Non-interactive textbox for status