Spontaneous speech recognition

对学术讲座任务中的自发语音识别进行了研究，得到了以下结果：(1)基于发音变体建模和无监督自适应的演讲语音识别研究了自发语音中观察到的语音变异。针对语音变体对语境的依赖性，提出了一种新的基于语音变体形态分析数据的语言建模方法。该方法在自发日语语料库(CSJ)上进行了测试，词错率(WER)下降了4.74%。此外，为了进一步提高识别性能，还引入了声学模型和语言模型的无监督自适应。(2)基于离散混合隐马尔可夫模型的演讲语音识别研究了离散混合隐马尔可夫模型(…)在含噪语音识别中的应用发现在环境噪声和脉冲噪声条件下，DMHMM的性能优于连续混合隐马尔可夫模型。然而，目前还不清楚这种方法在清洁条件下是否有效。这项调查的目的是评估DMHMM系统在清洁条件下的性能。在评价中，我们决定使用“自发日语语料库”(CSJ)，因为我们想要将我们的系统的性能与其他使用普通语料库的识别系统的性能进行比较，并阐明在这样一个更困难的任务中的性能。在识别实验中，使用3000个状态的DMHMM(每个状态16个混合成分)作为声学模型。利用CSJ的2668场讲座中的686万个单词对代表发音变化的语言模型进行训练，并用于识别。实验结果表明，该系统对10个男性演讲者的学术演讲获得了20.30%的WER，验证了该方法的有效性。较少

We investigated spontaneous speech recognition on academic lecture task and obtained the following results.(1) Lecture speech recognition using pronunciation variant modeling and unsupervised adaptationWe focus on the pronunciation variations observed in spontaneous speech. Aiming to introduce the context-dependence of pronunciation variants, we propose a new method of language modeling based on morphological analysis data designed for pronunciation variant. The proposed method was evaluated on the Corpus of Spontaneous Japanese (CSJ) and achieved the decrease in word error rate (WER) by 4.74% absolute. In addition, unsupervised adaptation of both acoustic and language models was introduced to improve the recognition performance further. The results showed the decrease in WER from 19.96% without adaptation to 15.41% with unsupervised adaptation.(2) Lecture speech recognition using discrete-mixture HMMsWe have investigated noisy speech recognition by using discrete-mixture HMM (DMHMM), … More and found that the performance of DMHMM overcame that of continuous-mixture HMM under environmental noise conditions or impulsive noise conditions. However, it is not clear whether this method is effective in clean conditions. The aim of this investigation is to evaluate the performance of the DMHMM system in clean conditions. In evaluation, we decided to use the "Corpus of Spontaneous Japanese" (CSJ) because we want to compare the performance of our system with that of other recognition systems with common speech corpus, and clarify the performance in such a more difficult task. In the recognition experiments, 3000-state DMHMMs (16 mixture components per state) were used as acoustic models. The language model which represents the pronunciation variety was trained by using 6.86 million words from 2668 lectures in CSJ and was used for recognition. As a result, the system obtained 20.30% WER for 10 academic lectures uttered by male speakers and demonstrated the effectiveness of the proposed method. Less