查詢結果分析
相關文獻
- 各種連續語音辨認法則應用在國語數字上性能之比較
- Tone Recognition for Continuous Mandarin Speech with Limited Training Data Using Selected Context-Dependent Hidden Markov Models
- 會聽國語的電腦
- Recognition of Mandarin Connected Digits
- A Comparative Study on the Performance of Several Speech Recognition Techniques Applied on the Highly Confusing Mandarin Syllables
- 類神經網路做國語數字辨認
- 如何向不同背景的學生教華語語音
- 電腦聲音控制介面之研究
- A Phonetic Labeling Method for Mat Database Processing
- An Overview of RNN-Based Mandarin Speech Recognition Approaches
頁籤選單縮合
| 題 名 | 各種連續語音辨認法則應用在國語數字上性能之比較 |
|---|---|
| 作 者 | 黃嘉若; | 書刊名 | 電信研究 |
| 卷 期 | 19:3 1989.09[民78.09] |
| 頁 次 | 頁303-312 |
| 專 輯 | 語音辨認技術專輯.1 |
| 分類號 | 312.23 |
| 關鍵詞 | 國語; 語音; 數字; 辨認; |
| 語 文 | 中文(Chinese) |
| 中文摘要 | 為使語音辨認達到實用化的周的,連續語音辨認是必面對的問題,最主要的困難就是自然流利的連續語音所引起的連音(Coarticulation)效應,使得前後發音不但會互相影響,且有時會難以區分相鄰兩單音的邊界,所以我們無法以一個兩段式的辦認方法(先確定單音的位置,而後以不連續語音的辨認方法進行辦認)來進行辦認。本文將分析比較三種不同的連續語音辦認技術,即二階段式動態時間比對,建階式動態時間校準及一階式動態規正這三種,並引用段開式k—中心法則(Segmental k-means Algorithm)來建立參考樣本,以進語連續數字串的辦認比較。 由實驗結果得知,建階式時間校準(已知數字串個數)較一階式動態規正的辦認率約高出10%,但當未知個數時,則兩者辦認率的差異不大,此外經開式k—中心法則處理後的參考樣本對辦認率有顯著的改善,即建階式時間校準(未知個數)提昇約10%,建階式時間校準(已知個數)及一階式動態規正時則提昇了30%。 |
| 英文摘要 | In order to bring the speech recognition into real life, we must focus on connected word recognition, however there are difficulties such as coarticulation and ambiguities between word boundaries. These considerations suggest the desirability of not making any segmentation prior to recognition. In this paper, we give a brief introduction to several connected word recognition algorithms, such as two-level dynamic time warping, level-building dynamic time warping and one-stage dynamic programming, and then use segmental k-means training procedure to generate reference template with coarticulaion information. A series of tests using Manadarin connected digit strings. Experimental results shown, level-building dynamic time warping (known string length) has 10% higher recognition rate than one-stage dynamic programming, but when string length is unknown, then there is no difference. Also, the recognition rate can be highly improved (30% in some cases) is we use reference template trained after segmental k-means procedure. |
本系統中英文摘要資訊取自各篇刊載內容。