Systematic Literature Review of Speaker Diarization Techniques : Toward Bridging Gaps in Low-resourced Languages using Machine Learning
Speaker diarization, the process of segmenting audio into speaker-specific regions, plays a critical role in various speech technologies by determining "who spoke when" in a conversation. This technique is particularly valuable for enhancing automatic speech recognition (ASR) and conversat...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
ARQII Publication
2025
|
| Subjects: | |
| Online Access: | http://ir.unimas.my/id/eprint/47225/2/801-2473-10%20%281%29.pdf http://ir.unimas.my/id/eprint/47225/ https://arqiipubl.com/ojs/index.php/AMS_Journal/article/view/801 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Speaker diarization, the process of segmenting audio into speaker-specific regions, plays a critical role in various speech technologies by determining "who spoke when" in a conversation. This technique is particularly valuable for enhancing automatic speech recognition (ASR) and conversational artificial intelligent systems. However, its application to low-resourced languages remains underexplored, limiting not only the performance of speaker diarization among low-resourced languages, but also stagnating the advancements of ASR to low-resourced languages. This is due to the fact that speaker diarization enables speaker adaptation in ASR, crucial for maximizing the performance of ASR itself. This lack of digital resources of speaker diarization to low-resourced languages, as well as the scarcity of its implementation presents a gap between low-resourced languages and popular languages in terms of the advancements of speech technologies involving the particular languages. This paper focuses on Sarawak Malay, a low-resourced language, and presents conversational data collected through a crowd-sourced approach, which needs speaker turns and transcripts. These missing annotations create challenges for building accurate acoustic models. To address this, we conducted a systematic review of recent speaker diarization research and related machine learning techniques. Using the PRISMA methodology, we reviewed 42 articles published between 2018 and 2023. Our findings identify key machine learning models, such as i-vectors and x-vectors, and open-source tools like Pyannote, which offer promising advancements in diarization performance. Besides that, these tools have shown potential to be implemented in developing speaker diarization models for low-resourced language. By highlighting the gaps in current research for low-resourced languages, we provide a pathway for improving speaker diarization models in these underrepresented languages through machine learning techniques. |
|---|