Continuous/Spontaneous Taiwanese Speech Recognition

This project intends to continue and extend the long-term National Science Council sponsored projects executed in Chang Gung University about Taiwanese speech recognition for further studies, mainly from the "acoustic model" to the "language model", in order to make the Taiwanese continuous speech recognition technology complete. In this plan, we intend to gradually achieve a few milestones on speech recognition of the spontaneous broadcast television news programs. The final goal is the Taiwanese speech transcription system from speech to syllable. Such a system can help generate the text of Taiwanese speech. It can further assist computer-aided Taiwanese language education, Taiwanese Digital Archives and other tasks. We plan to adopt the following steps: - Collation of existing reading-style speech of Taiwanese information and written materials - Use of several new algorithms to train acoustic model and language model - Collection of Taiwanese broadcast television news speech database and text data - Automatic segmentation, grouping of broadcast speech - Use of lightly Supervised training algorithm to train the acoustic model and language model - The use of a number of advanced speech recognition technology to improve the recognition rate We will seek to achieve the following objectives: Year 1 • Continuous Taiwanese read speech recognition Objectives: syllable error rate of 15% Year 2 •Spontaneous Taiwanese Broadcast TV News speech recognition Objectives: syllable error rate of 25% Year 3 •Improvement of Spontaneous Taiwanese Broadcast TV News speech recognition Objectives: syllable error rate of 15%, implementation of anautomatic Taiwanese speech transcription system

Project ID:PB9907-12640
External Project ID:NSC99-2221-E182-029-MY3