5.

1 Introduction

This chapter brings the entire project to a close. It summarizes what was done, presents the main
conclusions drawn from the results, highlights the challenges encountered during the
implementation, and gives recommendations for improvements and future work. The project
focused on developing a voice-based system using the ESP32-S3 that can verify a registered user
and then execute simple commands (“ON” and “OFF”). The system uses MFCC for extracting
speech features and Dynamic Time Warping (DTW) for comparing the input voice with stored
voice templates.

5.2 Summary of the Research Work

The main aim of this project was to build a voice-controlled system that can recognize
authorized users and allow them control a device using voice commands. To achieve this, the
following steps were carried out:

 Voice samples were collected from 30 different participants, with three repetitions
recorded from each person.
 The speech signals were processed, and MFCC features were extracted because MFCCs
represent speech in a compact form and work well for recognition tasks.
 A voice template was created and stored for each registered user.
 During testing, the system extracted MFCC features from a new voice input and used
DTW to compare it with the stored templates.
 If the voice matched an enrolled user (based on the DTW distance and threshold), the
system granted access and accepted voice commands.
 The user could then say “ON” or “OFF”, and the system executed the command by
controlling the output device.
 Finally, the system performance was evaluated using False Acceptance Rate (FAR) and
False Rejection Rate (FRR).

Overall, the project showed that voice recognition can be implemented successfully on a
microcontroller without using heavy deep learning models that require large datasets.
5.3 Conclusion

Based on the design, implementation, and testing carried out in this project, it can be concluded
that the objectives of the work were achieved. The system was able to:

 Verify a speaker using their voice and prevent unauthorized users from executing
commands.
 Extract meaningful voice features using MFCC, which helped improve recognition
performance.
 Use DTW effectively to match voice patterns even when the user spoke at different
speeds.
 Execute the intended commands (ON/OFF) once the user was verified.

In summary, the project confirmed that MFCC + DTW is a suitable approach for embedded
voice recognition systems, especially when working with limited memory and processing power
like the ESP32-S3.

5.4 Limitations of the Study

Even though the project worked as intended, some limitations were observed:

1. Small dataset size: Only 30 participants were used with three repetitions each. This is
reasonable for an embedded project but may not cover all voice variations.
2. Effect of noise: Background noise during recording or testing can affect MFCC
extraction and reduce matching accuracy.
3. Memory constraints: Since ESP32-S3 has limited memory, storing many templates or
large features is challenging.
4. Threshold sensitivity: The system’s FAR and FRR depend strongly on the chosen DTW
threshold. A strict threshold increases FRR, while a loose threshold increases FAR.
5. Limited commands: The system only recognizes “ON” and “OFF,” so it cannot handle
more complex voice interactions.
 6. Microphone quality: Differences in microphone sensitivity and distance from the
speaker can affect the voice signal and recognition outcome.

5.5 Recommendations

To improve the system and make it more robust, the following recommendations are suggested:

1. Collect more voice samples: Increasing the number of participants and repetitions will
improve system reliability.
2. Use multiple templates per user: Instead of storing one template per user, storing 2–3
templates can help handle voice variations.
3. Add noise filtering: Applying noise reduction techniques before feature extraction will
improve recognition in real environments.
4. Use adaptive thresholding: Setting a different threshold for each user may reduce false
rejections and improve overall accuracy.
5. Expand command vocabulary: Future versions can support more commands beyond
ON/OFF.
6. Improve hardware: Using a higher quality microphone and stable sampling settings can
enhance feature extraction quality.

5.6 Suggestions for Future Work

Future work can build on this project by:

 Implementing more advanced but lightweight machine learning approaches such as kNN,
SVM, or GMM, which may improve accuracy while still being microcontroller-friendly.
 Adding an enrollment interface (mobile app or web interface) to register users more
easily.
 Testing the system in more environments (indoor, outdoor, noisy spaces) to measure
robustness.
 Introducing security features like access logs and user identification records.
  Exploring a hybrid approach where DTW is used for quick matching while a small
classifier handles final decisions.

5.7 Contribution to Knowledge

This project contributes to knowledge by showing that a functional voice-based access control
system can be built on an embedded platform using MFCC and DTW, without depending on
deep learning models or large datasets. The work provides a practical method for developing
affordable voice-controlled systems that can be applied in smart homes, offices, and security-
based applications. The research established a voice recognition technology specifically to
control electrical switches, showcasing a practical implementation of AI in everyday life, also
established a system for reducing the likelihood of accidental switch activation and potentially
increase security by ensuring it responds only to authorized users and it established a technology
that can significantly aid individuals with physical disabilities to control electrical switches
without needing to physically interact with them.