1. Introduction As more and more computing power is put into today's processors, many resource intensive natural language input and processing methods can now be realized in a practical manner. Speech recognition ...
详细信息
1. Introduction As more and more computing power is put into today's processors, many resource intensive natural language input and processing methods can now be realized in a practical manner. Speech recognition technology is particularly important for Chinese. since. instead of clumsy and not very reliable key board input methods, it adopts the natural and effective means of communication used among people.
Numerous studies have demonstrated effectiveness of the multilayer networks with time sequences as inputs to the networks. A common design approach used in these neural networks(NN) is incorporation of short delays, o...
详细信息
Numerous studies have demonstrated effectiveness of the multilayer networks with time sequences as inputs to the networks. A common design approach used in these neural networks(NN) is incorporation of short delays, of temporal integration, or of recurrent connections. Spectral inputs are applied to input nodes sequentially, one frame at a time, and their corresponding input matrix can be formed. The NNs can thus be integrated into real time speech recognizer because only short delays are used.
In this paper, the neural network approaches for speech processing are reviewed briefly. The emphasis is placed on the automatic speech recognition, speech coding and speech synthesis using artificial neural networks ...
详细信息
In this paper, the neural network approaches for speech processing are reviewed briefly. The emphasis is placed on the automatic speech recognition, speech coding and speech synthesis using artificial neural networks (ANNs). It can be used to solve the sub-task of speech recognition problems because of their salient advantages, such as signal processing and feature extraction, and time alignment and pattern matching. The application of ANNs to speech coding lies in vector quantization and non-linear prediction of speech parameters.
Context-transparent word can improve the predictive ability of a language model. A method to automatically find context-transparent words is discussed. With this method, we find 6 such words in Chinese. We use these c...
详细信息
Context-transparent word can improve the predictive ability of a language model. A method to automatically find context-transparent words is discussed. With this method, we find 6 such words in Chinese. We use these context-transparent words in Chinese language models and get improvements both in perplexity and in decoding accuracy.
Current continuous speech recognition system requires punctuation marks being spoken during dictation. However it is difficult to do so in some cases. We propose a novel method to add punctuation marks automatically b...
详细信息
Current continuous speech recognition system requires punctuation marks being spoken during dictation. However it is difficult to do so in some cases. We propose a novel method to add punctuation marks automatically by using underlying language and acoustic information. When the location of a punctuation mark is given by the speaker,the method suggests the most likely punctuation mark(?) when no location information is *** determines both the most likely location and the most likely punctuation mark. Experiments show the effectiveness of our method.
This paper gives an overview of IBM Large Vocabulary Continues Speech Recognition System of Mandarin, which was used for transcription of broadcast news. It describes the acoustic and language models, segmentation and...
详细信息
This paper gives an overview of IBM Large Vocabulary Continues Speech Recognition System of Mandarin, which was used for transcription of broadcast news. It describes the acoustic and language models, segmentation and unsupervised adaptation techniques. Experiments results, analysis and discussions are reported also.
暂无评论