WESTERN GOVERNORS UNIVERSITY | D604
Advanced Analytics
Task 1
Environmental Sound Classification
Using Deep Learning
Project focus: 50-class environmental audio classification with a convolutional neural network and
mel-spectrogram inputs.
Document scope: Sections A-K, reported empirical evidence, technical rationale, model evaluation,
business recommendations, source attribution, and full executable Python source code.
Prepared: October 27, 2026 | Course: D604 Advanced Analytics
Evidence statement. Numerical results and supplied figures are taken from the uploaded 2025 Task 1 report. The
original training logs, predictions, code execution and audio archive were not provided, and no model training was
performed while producing this document. The attached source-code file is a reproducibility implementation, not
proof that the reported metrics were reproduced.
Environmental sound classification • technical report 1
,WGU D604 | ADVANCED ANALYTICS | TASK 1
Contents and report scope
Contents and report scope 2
Executive summary 3
A. Research question and business problem 3
A1. Research question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
A2. Objectives and measurable goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
A3. Selected neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
A4. Justification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4
B. Image dataset 5
C. Video dataset 5
D. Audio data preparation 5
D1. Exploratory data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
D2. Spectrogram extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
D3. Audio signal creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6
D4. Signal padding and truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
D5. Spectrogram normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
D6. Feature extraction and tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
D7. Ordered data-preparation pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7
D8. Training, validation and test split . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
D9. Prepared audio data deliverables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
D10. Rationale and quality controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
E. Neural network architecture 8
E1. Model summary and reproducible design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
E2. Architectural justification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8
E3. Backpropagation and training hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
E4. Confusion matrix and error categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
F. Model evaluation 10
F1. Training progress and stopping criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
F2. Model fitness assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
F3. Predictive accuracy and operational significance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
G. Summary and recommendations 12
G1. Saving the trained network and reproducibility artifacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
G2. Assessment of neural-network functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
G3. Effectiveness relative to the business problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Environmental sound classification • technical report 2
, WGU D604 | ADVANCED ANALYTICS | TASK 1
G4. Lessons learned and prioritized improvements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
G5. Recommended course of action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
H. Code output and documentation 13
I. Sources for third-party code 14
J. Sources and citations 14
K. Professional communication 15
Appendix A. Complete executable Python implementation 15
Appendix B. Evidence and decision checklist 19
Section organization: A - Research Question and Business Problem; B - Image Dataset; C - Video
Dataset; D - Audio Data Preparation; E - Neural Network Architecture; F - Model Evaluation; G - Summary
and Recommendations; H - Code Output and Documentation; I - Sources for Third-Party Code; J - Sources
and Citations; K - Professional Communication; plus technical appendices.
Executive summary
This project examines whether a convolutional neural network (CNN) can identify environmental sound
events for a proposed smart-city monitoring workflow. The ESC-50 benchmark contains 2,000 labeled
five-second audio recordings distributed evenly among 50 classes. Waveforms are resampled to 22,050
Hz, standardized to five seconds, transformed into log-mel spectrograms with 128 frequency bands, and
used as two-dimensional CNN inputs. The project described in the provided Task 1 document uses a 70%
training, 15% validation and 15% test split.
The supplied report records training accuracy of 99.43%, validation accuracy of 55.67% and test
accuracy of 55.33%. The training-validation difference is 43.76 percentage points, indicating
substantial overfitting. Despite outperforming the 2% uniform-random baseline for a balanced 50-class
task, the model misses approximately 44.67% of test examples. It therefore does not meet the proposed
70% pilot accuracy goal or the stronger 85% acceptance criterion recommended before any operational
rollout.
The immediate recommendation is not to deploy. Priority actions are to validate with source-isolating
ESC-50 folds, reduce network capacity, augment only training recordings, compare a transfer-learning
baseline, and test external recordings representative of real operating conditions. This project is a
structured technical analysis of source-reported outcomes; reproducibility requires running the supplied
code on the actual ESC-50 files and replacing the historical figures and metrics with the resulting
evidence.
A. Research question and business problem
A1. Research question
Research question: How accurately and reliably can a convolutional neural network classify 50 classes
of environmental sound from the ESC-50 dataset, and does the resulting predictive performance justify
considering it as a supporting component in automated urban environmental monitoring?
Environmental sound classification • technical report 3
Advanced Analytics
Task 1
Environmental Sound Classification
Using Deep Learning
Project focus: 50-class environmental audio classification with a convolutional neural network and
mel-spectrogram inputs.
Document scope: Sections A-K, reported empirical evidence, technical rationale, model evaluation,
business recommendations, source attribution, and full executable Python source code.
Prepared: October 27, 2026 | Course: D604 Advanced Analytics
Evidence statement. Numerical results and supplied figures are taken from the uploaded 2025 Task 1 report. The
original training logs, predictions, code execution and audio archive were not provided, and no model training was
performed while producing this document. The attached source-code file is a reproducibility implementation, not
proof that the reported metrics were reproduced.
Environmental sound classification • technical report 1
,WGU D604 | ADVANCED ANALYTICS | TASK 1
Contents and report scope
Contents and report scope 2
Executive summary 3
A. Research question and business problem 3
A1. Research question . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
A2. Objectives and measurable goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
A3. Selected neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
A4. Justification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4
B. Image dataset 5
C. Video dataset 5
D. Audio data preparation 5
D1. Exploratory data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
D2. Spectrogram extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
D3. Audio signal creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6
D4. Signal padding and truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
D5. Spectrogram normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
D6. Feature extraction and tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
D7. Ordered data-preparation pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7
D8. Training, validation and test split . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
D9. Prepared audio data deliverables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
D10. Rationale and quality controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
E. Neural network architecture 8
E1. Model summary and reproducible design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
E2. Architectural justification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8
E3. Backpropagation and training hyperparameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
E4. Confusion matrix and error categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
F. Model evaluation 10
F1. Training progress and stopping criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
F2. Model fitness assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
F3. Predictive accuracy and operational significance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
G. Summary and recommendations 12
G1. Saving the trained network and reproducibility artifacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
G2. Assessment of neural-network functionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
G3. Effectiveness relative to the business problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Environmental sound classification • technical report 2
, WGU D604 | ADVANCED ANALYTICS | TASK 1
G4. Lessons learned and prioritized improvements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
G5. Recommended course of action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
H. Code output and documentation 13
I. Sources for third-party code 14
J. Sources and citations 14
K. Professional communication 15
Appendix A. Complete executable Python implementation 15
Appendix B. Evidence and decision checklist 19
Section organization: A - Research Question and Business Problem; B - Image Dataset; C - Video
Dataset; D - Audio Data Preparation; E - Neural Network Architecture; F - Model Evaluation; G - Summary
and Recommendations; H - Code Output and Documentation; I - Sources for Third-Party Code; J - Sources
and Citations; K - Professional Communication; plus technical appendices.
Executive summary
This project examines whether a convolutional neural network (CNN) can identify environmental sound
events for a proposed smart-city monitoring workflow. The ESC-50 benchmark contains 2,000 labeled
five-second audio recordings distributed evenly among 50 classes. Waveforms are resampled to 22,050
Hz, standardized to five seconds, transformed into log-mel spectrograms with 128 frequency bands, and
used as two-dimensional CNN inputs. The project described in the provided Task 1 document uses a 70%
training, 15% validation and 15% test split.
The supplied report records training accuracy of 99.43%, validation accuracy of 55.67% and test
accuracy of 55.33%. The training-validation difference is 43.76 percentage points, indicating
substantial overfitting. Despite outperforming the 2% uniform-random baseline for a balanced 50-class
task, the model misses approximately 44.67% of test examples. It therefore does not meet the proposed
70% pilot accuracy goal or the stronger 85% acceptance criterion recommended before any operational
rollout.
The immediate recommendation is not to deploy. Priority actions are to validate with source-isolating
ESC-50 folds, reduce network capacity, augment only training recordings, compare a transfer-learning
baseline, and test external recordings representative of real operating conditions. This project is a
structured technical analysis of source-reported outcomes; reproducibility requires running the supplied
code on the actual ESC-50 files and replacing the historical figures and metrics with the resulting
evidence.
A. Research question and business problem
A1. Research question
Research question: How accurately and reliably can a convolutional neural network classify 50 classes
of environmental sound from the ESC-50 dataset, and does the resulting predictive performance justify
considering it as a supporting component in automated urban environmental monitoring?
Environmental sound classification • technical report 3
