README.md

---
tags:
- audio-classification
- sound-event-detection
- wav2vec2
- urban-acoustics
- deep-learning
datasets:
- UrbanSoundscape_EventDetection_Metadata
license: apache-2.0
model-index:
- name: UrbanSound_EventDetection_Wav2Vec2
  results:
  - task:
      name: Audio Classification
      type: audio-classification
    metrics:
    - type: accuracy
      value: 0.945
      name: Event Detection Accuracy
    - type: f1_macro
      value: 0.938
      name: Macro F1 Score
---

# UrbanSound_EventDetection_Wav2Vec2

## 👂 Overview

The **UrbanSound_EventDetection_Wav2Vec2** is a highly efficient model based on the pre-trained **Wav2Vec2** architecture, fine-tuned specifically for classifying momentary and continuous sound events within urban environments. It processes raw audio waveforms to identify one of eight high-priority urban sound classes, focusing on high-impact and potentially anomalous events. 

## 🧠 Model Architecture

This model utilizes the standard Wav2Vec2 pipeline, which operates directly on raw audio data without the need for manual feature extraction (like MFCCs).

* **Base Model:** `facebook/wav2vec2-base`
* **Feature Extractor:** A stack of 1D convolutional layers extracts local features from the raw waveform.
* **Transformer Encoder:** 12 layers of Transformer blocks capture long-range dependencies and global context within the audio clip.
* **Classification Head:** A task-specific linear layer is placed on top of the contextualized representations to predict one of the 8 event labels.
* **Target Classes:** Car\_Horn, Children\_Playing, Dog\_Barking, Machinery\_Hum, Siren\_Emergency, Train\_Whistle, Tire\_Screech, and Glass\_Shattering.

## 🎯 Intended Use

This model is intended for smart city, safety, and acoustic monitoring systems:

1.  **Acoustic Surveillance:** Real-time detection of emergency sounds (Siren, Glass Shattering, Tire Screech) for public safety alerting.
2.  **Noise Pollution Monitoring:** Quantifying the occurrence and frequency of specific noise sources (Car Horn, Machinery Hum) in different city zones.
3.  **Urban Planning:** Analyzing soundscape composition to inform policy on zoning and noise mitigation strategies.

## ⚠️ Limitations

1.  **Event Overlap:** The current setup is trained for single-label classification. If multiple sounds occur simultaneously (e.g., Siren + Dog Barking), the model will only output the single most probable event, potentially ignoring others.
2.  **Domain Shift:** The model's performance may degrade if deployed in environments with significantly different background noise profiles (e.g., highly quiet suburbs vs. extremely loud Asian markets).
3.  **Localization:** This model performs *event detection* but does not inherently provide *sound localization* (Direction-of-Arrival or DOA), which would require specialized input features (like ambisonic audio) and a different model head.

---

### MODEL 2: **MedicalChatbot_IntentClassifier_RoBERTa**

This model is a RoBERTa-based model for multi-class classification of user intent within medical dialogue transcripts.

#### config.json

```json
{
  "_name_or_path": "roberta-base",
  "architectures": [
    "RobertaForSequenceClassification"
  ],
  "hidden_size": 768,
  "model_type": "roberta",
  "num_hidden_layers": 12,
  "vocab_size": 50265,
  "id2label": {
    "0": "Symptom_Reporting",
    "1": "Advice_Seeking",
    "2": "Medication_Query",
    "3": "Appointment_Scheduling",
    "4": "Billing_Query",
    "5": "Causal_Query",
    "6": "Record_Retrieval",
    "7": "Urgency_Assessment"
  },
  "label2id": {
    "Symptom_Reporting": 0,
    "Advice_Seeking": 1,
    "Medication_Query": 2,
    "Appointment_Scheduling": 3,
    "Billing_Query": 4,
    "Causal_Query": 5,
    "Record_Retrieval": 6,
    "Urgency_Assessment": 7
  },
  "num_labels": 8,
  "problem_type": "single_label_classification",
  "transformers_version": "4.36.0"
}