【摘要】：隨著深度學習的發(fā)展,深度神經(jīng)網(wǎng)絡(luò)(Deep Neural Network,DNN)與循環(huán)神經(jīng)網(wǎng)絡(luò)(Recurrent Neural Networks,RNN)已被成功應(yīng)用于英文語音識別和語音關(guān)鍵詞檢出系統(tǒng)。本文主要研究了分別用深度神經(jīng)網(wǎng)絡(luò)-隱馬爾科夫模型(Deep Neural Network-Hidden Markov Model,DNN-HMM)和帶有長短時記憶單元的循環(huán)神經(jīng)網(wǎng)絡(luò)(Long Short Term Memory RNN,LSTM-RNN)對中文聲韻母進行聲學建模,從而優(yōu)化現(xiàn)有中文語音關(guān)鍵詞檢出系統(tǒng)性能。本文首先介紹了連續(xù)語音識別的框架與原理,包括語音信號的特征提取、語音信號聲學建模技術(shù)、發(fā)音字典和語言模型以及基于加權(quán)有限狀態(tài)轉(zhuǎn)換器的語音解碼網(wǎng)絡(luò)。其中語音信號特征提取包括感知線性預(yù)測系數(shù)、梅爾頻率倒譜系數(shù)、濾波器組特征、基頻特征四種聲學特征。其次研究了基于連續(xù)語音識別器的語音關(guān)鍵詞檢出技術(shù),包括基于網(wǎng)格結(jié)構(gòu)建立索引、關(guān)鍵詞搜索方法、關(guān)鍵詞確認置信度以及語音關(guān)鍵詞檢出系統(tǒng)的評價指標。本文還研究了一種中文語音關(guān)鍵詞檢出系統(tǒng),此系統(tǒng)采用高識別率的聲韻母進行聲學建模和檢索,通過查表法將輸入漢字字符形式的關(guān)鍵字轉(zhuǎn)化為聲韻母進行關(guān)鍵詞檢出。本文分別訓練基于DNN-HMM的聲學模型和基于LSTM-RNN的聲學模型,搭建中文語音關(guān)鍵詞檢出系統(tǒng),各得到了73.32%和79.84%的召回率,說明使用LSTM-RNN聲學建�？梢詢�(yōu)化語音關(guān)鍵詞檢出系統(tǒng)性能。之后為搭建的中文語音關(guān)鍵詞檢出系統(tǒng)選取不同聲學特征進行性能分析,結(jié)果表明基頻特征可以一定程度上提高檢出性能;然后采用融合置信度優(yōu)化中文語音關(guān)鍵詞檢出系統(tǒng)性能;其次,對比兩個系統(tǒng)在不同規(guī)格訓練數(shù)據(jù)下的性能,討論了各自的應(yīng)用范圍;最后,提出了一種召回率更高的基于系統(tǒng)融合的中文語音關(guān)鍵詞檢出系統(tǒng)。
[Abstract]:With the development of deep learning, depth neural network (Deep Neural) and cyclic neural network (Recurrent Neural) have been successfully applied to English speech recognition and speech keyword detection systems. In this paper, the acoustic modeling of Chinese consonants is mainly studied by using the deep neural network-hidden Markov model (Deep Neural Network-Hidden Markov Model DNN-HMM) and the cyclic neural network (Long Short Term Memory RNNN LSTM-RNN) with long and short memory units. In order to optimize the existing Chinese voice keyword detection system performance. This paper first introduces the framework and principle of continuous speech recognition, including feature extraction of speech signal, acoustic modeling technology of speech signal, pronunciation dictionary and language model, and speech decoding network based on weighted finite state converter. The speech signal feature extraction includes four acoustic features: perceptual linear prediction coefficient, Mel frequency cepstrum number, filter bank feature and fundamental frequency feature. Secondly, the technology of speech keyword detection based on continuous speech recognizer is studied, including indexing based on grid structure, keyword search method, confidence of keyword confirmation and evaluation index of speech keyword detection system. This paper also studies a Chinese phonetic keyword detection system, which uses a high recognition rate phonetic mother for acoustic modeling and retrieval, and converts the key words in the Chinese character form to the consonant for keyword detection through the look-up table method. In this paper, the acoustic model based on DNN-HMM and the acoustic model based on LSTM-RNN are trained, and the Chinese voice keyword detection system is built, and the recall rates of 73.32% and 79.84% are obtained, respectively. It is shown that the performance of the system can be optimized by using LSTM-RNN acoustic modeling. Then the different acoustic features are selected for the Chinese speech keyword detection system. The results show that the fundamental frequency feature can improve the detection performance to some extent. Then the fusion confidence is used to optimize the performance of the Chinese voice keyword detection system. Secondly, the performance of the two systems under different specifications training data is compared, and their application scope is discussed. A Chinese speech keyword detection system based on system fusion with higher recall rate is proposed.
【學位授予單位】：南京理工大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TN912.3

【相似文獻】

相關(guān)重要報紙文章 前2條

1 葉向榮 鄭民軍;大破語音關(guān)[N];中國教育報;2001年

2 山東省陽信縣翟王鎮(zhèn)韓箔中心小學 張玉蘭;語文課堂 讓學生快樂表達[N];學知報;2010年

相關(guān)碩士學位論文 前2條

1 侯云飛;中文語音關(guān)鍵詞檢出技術(shù)研究[D];南京理工大學;2017年

2 朱林;語音關(guān)鍵詞檢測若干問題研究[D];北京郵電大學;2015年

，

本文編號：2172099