Robust voice activity detection using DNN approaches
Voice activity detection (VAD) is a pivotal component in various speech processing applications, playing a crucial role in tasks such as speech recognition, speaker diarization, and noise suppression. Recognizing its significance, this thesis delves into the exploration of advancements in single-cha...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
Nanyang Technological University
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/175226 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | Voice activity detection (VAD) is a pivotal component in various speech processing applications, playing a crucial role in tasks such as speech recognition, speaker diarization, and noise suppression. Recognizing its significance, this thesis delves into the exploration of advancements in single-channel VAD systems, leveraging
the power of deep learning techniques. Through meticulous experimentation and
analysis, we undertake comprehensive evaluations of three prominent VAD models:
Pyannote, Silero, and MarbleNet, across a spectrum of conditions and scenarios. Our investigations encompass a nuanced examination of varying parameters such as chunk sizes, strides, and prediction thresholds, aiming to discern their nuanced impacts on model performance. From our findings, we discern Pyannote as the standout performer exhibiting superior accuracy compared to Silero by approximately 16.87% and MarbleNet by approximately 25.97% on the DIHARD III dataset. Consequently,
we pivot our focus towards enhancing Pyannote’s capabilities. In the process
of enhancing, we looked into how different parameters affect the performance of
Pyannote and trained models on varying chunk sizes and stride to deduce the same.
With this, we were able to conclude that models trained on small chunk size and
strides do not necessarily perform well during inference with small chunks and strides.
Additionally, we delve into the realm of scalability and production readiness, exploring strategies facilitated by the Open Neural Network Exchange (ONNX) framework.
These efforts provide important insights that can enhance the field of VAD, leading to the development of more robust and efficient voice activity detection systems capable of meeting the needs of modern speech processing applications |
|---|