High accuracy transcription, cleaning and fast term detection for spoken documents

研究了语音自动识别、口语风格转换为书面风格和口语文本的快速特征提取等问题。为了准确的自动语音识别，我们提出了一种新的语音识别模型，超越了传统的HMM，并显示了有效性。为了清理语音识别结果，我们提出了一个概率模型，清理从口语风格识别错误的书面风格，并显示了有效性。最后，本文提出了一种基于三元组索引的语音文本词检测方法，并验证了该方法的有效性。

We studied on automatic speech recognition, written style transformation from spoken style transcription and fast term detection for spoken documents. For accurately automatic speech recognition, we proposed a novel speech recognition model beyond the conventional HMM, and showed the effectiveness. For cleaning speech recognition results, we proposed a probabilistic model of cleaning from spoken style with recognition errors to written style and showed the effectiveness. Finally, we proposed a fast term detection method based on tri-gram indexes from the transcription of spoken documents and showed the effectiveness.