资源稀缺蒙语语音识别研究

Research on Low-resource Mongolian Speech Recognition

随着语音识别技术的发展,资源稀缺语言的语音识别系统的研究吸引了更广泛的关注。以蒙语为目标识别语言,研究了在资源稀缺的情况下(如仅有10小时的带标注的语音)如何利用其他多语言信息提高识别系统的性能。借助基于多语言深度神经网络的跨语言迁移学习和基于多语言深度Bottleneck神经网络的抽取特征可以获得更具有区分度的声学模型。通过搜索引擎以及网络爬虫的定向抓取获得大量的网页数据,有助于获得文本数据,以增强语言模型的性能。融合多个不同识别结果以进一步提高识别精度。与基线系统相比,多种系统融合的识别绝对错误率减少12%。

With the development of speech recognition technology,the research on low-resource speech recognition has gained extensive attention.Taking the Mongolian as the target language,we studied how to use the multilingual information to improve the performance of speech recognition in the low-resource condition,for example,only 10 hours of transcribed speech data are used for acoustic modeling.More discriminative acoustic model can be gotten by using crosslingual transfer of multilingual deep neural network and multilingual deep bottleneck features.Large amount of web pages can be gotten by using the web search engine and Web crawler,which can help to get large amount of text data for improving the performance of language model.It can further improve the recognition results by fusing different number of recognition results from different recognizers.Comparing the fusion recognition result with the recognition result of baseline system,there are nearly 12% absolute word error rate（WER）reductions.