查詢結果分析
來源資料
相關文獻
- 應用特徵偏移向量的線性調整作環境調適之語音辨認
- 結合虛擬實境技術之環境適應步行及平衡復健訓練系統之實體開發
- 淺談中文字及其輸入、辨識之比較
- 淺談中文字及其輸入、辨識之比較
- 漫談智慧財權之保護
- 原住民學生與非原住民學生入學適應輔導研究初探
- 中文語音辨識系統
- Wavelet Transforms for Speech Signal Processing
- Towards a Pure Chinese Spoken Dialogue System
- Prototype LVQ Based Computerized Tool for Accent Diagnosis among Chinese Speakers of English as a Foreign Language
頁籤選單縮合
題 名 | 應用特徵偏移向量的線性調整作環境調適之語音辨認=Linear Adaptation of Feature Deviation Vector for Environment Adaptive Speech Recognition |
---|---|
作 者 | 李立民; 林耕弘; | 書刊名 | 明志學報 |
卷 期 | 35:1 2003.06[民92.06] |
頁 次 | 頁1-8 |
分類號 | 312.23 |
關鍵詞 | 語音辨識; 環境適應; 外加雜訊; 特徵偏移向量; Speech recognition; Environment adaptation; Additive noise; Feature deviation vector; |
語 文 | 中文(Chinese) |
中文摘要 | 在安靜環境中所發展出來的語音辨識系統移植到雜訊環境時,其辨認率將會急遽下降,故雜訊效應的補償方法將為語音辨認之重要課題。在本文中,我們對語音之逆頻譜特徵係數在外加雜訊下的行為進行探討。我們發現逆頻譜向量的值會隨著外加雜訊的增加而漸縮小,當環境為白色雜訊時,其變動的方向會與乾淨語音的逆頻譜向量的方向相反,且會縮小到原點,但是在非白色雜訊下,變動的方向則往純雜訊與乾淨語音間的特徵向量差值方向行進,基於這麼行為,我們提出特徵偏移向量,將其引入參考模型中,我們利用少量的混雜語音與參考模型中每個狀態之逆頻譜均值間的差值,求得特徵偏移向量,但考慮到求出特徵偏移向量所使用的混雜語音,可能與實際使用時的環境並不匹配,所以我們再利用一最佳純量係數,乘上特徵偏移向量作為最佳線性調整量,加至原始的語音參考模型,使模型能夠更適應的在雜訊環境下使用,實驗結果顯示這個方法能夠有效的使用在白色雜訊與其他非白色雜訊環境下。 |
英文摘要 | When a speech recognition system in queit environment is moved to a noisy environment, the recognition rate drops drastically. The compensation of noise effect becomes an important task for noisy speech recognition. In this study, we investigate the behavior of speech cepstral vector due to additive noise. We find that the cepstral vectors deviates as the level of additive noise increases. In the case of white noise, the direction of cepstral vector deviation is approximately opposite to the direction of the cepstral vector of the clean speech. As power level of the white noise increases, the cepstral vector of the noisy speech will converge to the zero vector. However, for other types of noise, the change of cepstral vector is approximately at the direction of the difference vector of the noise cepstral vector and clean speech cepstral vector. Base on this behavior, we include a feature deviation vector into the reference model to compensate for the noise effect. The deviation vector is calculated according to the difference value of the cepstral vector of a few noisy speech and the corresponding model state cepstral mean vector. During the pattern matching phase, an optimally scaled deviation vector is added to the state mean vector of the clean speech model so that the clean speech model is adapted to the noisy environment. Experimental results show that the proposed method is effective for white noise and color noises. |
本系統中英文摘要資訊取自各篇刊載內容。