Research On Event-Detection Based Phone Recognition

Speech event is acoustic cues sequences where all kinds of speech attributes or featureschange abruptly, we can make speech recognition by detecting the events with the assistant ofspeech knowledge. In various kinds of speech events, the most widely used phonologicalattributes is a characteristics parameters reflecting the vocal process, and the phonologicalattributes detection based method has become state-of-the-art technique in Automatic SpeechRecognition. This paper focus on phonological attributes detection and phone modeling process,containing the following aspects:As speech signal has strong correlation in time series, to introduce the neighbor framesinformation into system will improve the recognition performance. A phonological attributesdetection method based on long-term information is proposed in this article, which comprises oftwo layers Recurrency and Time-delayed neural networks (RTDNN). The low-level RTDNNcarries out phonological attributes detection on short-term features. Based on the output oflow-level RTDNN, the high-level RTDNN considering long-term information which fully tapsthe relation between speech signals in time and make more context dependent feature intorecognition. The experiments results show that，compared with the detection using short-termfeatures, the phonological detection rate improves a lot based on long-term features, in addition,about2%in white and pink noise conditions.This paper proposes a phone boundary detection method based on the correlation betweenadjacent frames, and taking these features and boundary information in Conditional RandomFields based phone recognition system. Firstly, calculate the angles between posterior probabilityvectors of adjacent frames, and then mark the frames with maximum angle as the boundarycandidates. Secondly, select the true boundaries through three restrictions in these boundarycandidates. In the percent of detected boundaries within20ms from the manually placedboundaries reached78.3%. Finally, the combination of phonological attributes and phoneboundaries are presented as the observation vectors of Conditional Random Fields to makephone recognition, experimental results show that the accuracy of phoneme recognition rateimproves1.3%.On account of the differences in speaking rate may impair the adaptation ability ofacoustical models, a novel adaptation algorithm is proposed in this paper, which adjusts theframe and step size in the front end of the system with the cell of one utterance. After adaptation,the speaking rate consistent with the average rate of the speech corpus and decreasing its effectin model training. The algorithm was used in the pre-processing stage before the phonologicalattributes features detection, and then with the nonlinear transformation, we put them as theobservation of Hidden Markov Models and Conditional Random Fields based phone recognitionsystems. After the adaptation approach, the average frame of one phone in an utterance becomesconstant and the dynamic range decreases, therefore the phoneme classification rate increasesabout0.8%.In current event-detection based automatic speech recognition system, two kinds of mistakes are familiar, namely the detection errors and the asynchronous problems betweenphonological attributes and phone boundaries. Aiming at these problems, an adjustment methodbased on the prior knowledge of corpus is proposed in this paper. In this method, priorknowledge of training set and the detection result from experiment are combined at first, then theasynchronies in the phone boundary area is compensated and the detection errors will becorrected to some extent, which may improve the model precision at little cost.