The North System for Formosa Speech Recognition Challenge 2023

The North System for Formosa Speech Recognition Challenge 2023 Li-Wei Chen 陳力瑋 North Co., Ltd., Taiwan Kai-Chen Cheng 鄭楷蓁 North Co., Ltd., Taiwan Hung-Shin Lee 李鴻欣 North Co., Ltd., Taiwan Abstract This report provides a concise overview of the proposed North system, which aims to achieve automatic word/syllable recognition for Tai- wanese Hakka (Sixian). The report outlines three key components of the system: the acqui- sition, composition, and utilization of the train- ing data; the architecture of the model; and the hardware specifications and operational statis- tics. The demonstration of the system has been made public1. Keywords: Hakka, Sixian, speech recognition 1 Introduction This document furnishes a succinct yet compre- hensive overview of the proposed North System, a technologically advanced mechanism designed with the primary objective of achieving automatic recognition of words and syllables specific to the Taiwanese Hakka language, with a particular fo- cus on the Sixian dialect. The report meticulously delineates three pivotal components integral to the effective functionality and operation of the system, as enumerated below: 1. Acquisition, Composition, and Utilization of Training Data: • Acquisition: The systematic collection and sourcing of relevant linguistic data pertinent to the Taiwanese Hakka lan- guage, ensuring a robust and representa- tive dataset. • Composition: The strategic assembly and organization of the acquired data, en- suring it is structured in a manner con- ducive to effective machine learning. 1https://shorturl.at/mzGL7 • Utilization: The application of the com- posed data in training the system, ensur- ing it accurately and efficiently recog- nizes and processes linguistic elements of the Sixian dialect. 2. Architectural Framework of the Model: • A detailed exposition of the structural and operational framework of the model, elucidating the technological and algo- rithmic methodologies employed to facil- itate accurate linguistic recognition and processing. • An exploration of the model’s design principles, underlying algorithms, and computational processes that enable it to effectively learn, recognize, and interpret the linguistic nuances of the Taiwanese Hakka language. 3. Hardware Specifications and Operational Met- rics: • Hardware Specifications: A thorough breakdown of the technological infras- tructure supporting the system, detailing the hardware components and specifica- tions that underpin its operation. • Operational Metrics: An analytical overview of the system’s performance metrics, providing insights into its opera- tional efficiency, accuracy, and reliability in real-world applications. 2 Data Table 1 meticulously presents a detailed inventory of our principal corpus sources, accompanied by their respective statistical details, providing an in- depth insight into the voluminous data utilized in our research and development endeavors. The 35th Conference on Computational Linguistics and Speech Processing (ROCLING 2023) Taipei City, Taiwan, October 20-21, 2023. The Association for Computational Linguistics and Chinese Language Processing In addition to the primary corpus, our research team has assiduously gathered an extensive col- lection of Hakka (Sixian dialect)-related speech data, exceeding 100 hours, from a variety of online platforms. This includes, but is not limited to: • YouTube: A prominent video-sharing plat- form where a myriad of Hakka (Sixian) lin- guistic content, ranging from casual conversa- tions to …

Content truncated to 500 words