The Atayal language is an Austronesian language spoken by the Atayal people, one of Taiwan’s indigenous groups, in which Squliq and C’uli’ are two major dialects. However, a significant decline in population has led to severe interruptions in language transmission. This study leverages machine learning techniques to develop an integrated system that combines text-to-speech (TTS), automatic speech recognition (ASR), and bidirectional machine translation (MT). In this study, we construct a parallel corpus consisting of Atayal speech, Atayal Romanized script (the standardized writing system of the Atayal language is the Latin-based alphabets), and Mandarin Chinese translations. The integrated system and parallel corpus serve as an effective tool for language preservation. The TTS module converts Atayal Romanized script into natural-sounding Atayal speech, simulating native intonations and prosody, thereby assisting learners in acquiring accurate pronunciation. The ASR module transcribes Atayal speech into Romanized script, supporting speech input applications and lowering the barrier for text entry. The MT module enables real-time translation between Atayal Romanized script and Mandarin Chinese text, expanding the practical use and scope of the Atayal language. A web-based platform was developed to integrate TTS, ASR, and MT functionalities and The website provides open access, allowing users to utilize each function, thereby further facilitating the preservation and dissemination of Atayal. Performance evaluation was conducted using the TTS module for speech synthesis, applying voiceprint recognition models to extract voice features for similarity comparison. The ASR and MT module was evaluated based on word error rate (WER). Experimental results indicate that the TTS module achieved a \(78.14\%\) accuracy in voice similarity verification. The ASR model attained a WER of \(1.01\%\) . For translation tasks, the WER from Mandarin Chinese to Romanized Atayal was \(51.98\%\) , while the WER from Romanized Atayal to Mandarin Chinese was \(47.71\%\) . These results demonstrate the potential of the integrated system to support the revitalization of the Atayal language.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Advancing Atayal Language Preservation with AI-Driven Multimodal Speech and Text Processing

  • Jia-Lien Hsu,
  • Wei-Yuan Li

摘要

The Atayal language is an Austronesian language spoken by the Atayal people, one of Taiwan’s indigenous groups, in which Squliq and C’uli’ are two major dialects. However, a significant decline in population has led to severe interruptions in language transmission. This study leverages machine learning techniques to develop an integrated system that combines text-to-speech (TTS), automatic speech recognition (ASR), and bidirectional machine translation (MT). In this study, we construct a parallel corpus consisting of Atayal speech, Atayal Romanized script (the standardized writing system of the Atayal language is the Latin-based alphabets), and Mandarin Chinese translations. The integrated system and parallel corpus serve as an effective tool for language preservation. The TTS module converts Atayal Romanized script into natural-sounding Atayal speech, simulating native intonations and prosody, thereby assisting learners in acquiring accurate pronunciation. The ASR module transcribes Atayal speech into Romanized script, supporting speech input applications and lowering the barrier for text entry. The MT module enables real-time translation between Atayal Romanized script and Mandarin Chinese text, expanding the practical use and scope of the Atayal language. A web-based platform was developed to integrate TTS, ASR, and MT functionalities and The website provides open access, allowing users to utilize each function, thereby further facilitating the preservation and dissemination of Atayal. Performance evaluation was conducted using the TTS module for speech synthesis, applying voiceprint recognition models to extract voice features for similarity comparison. The ASR and MT module was evaluated based on word error rate (WER). Experimental results indicate that the TTS module achieved a \(78.14\%\) accuracy in voice similarity verification. The ASR model attained a WER of \(1.01\%\) . For translation tasks, the WER from Mandarin Chinese to Romanized Atayal was \(51.98\%\) , while the WER from Romanized Atayal to Mandarin Chinese was \(47.71\%\) . These results demonstrate the potential of the integrated system to support the revitalization of the Atayal language.