使用卷积循环神经网络的多语言识别
计算与语言
2017-05-19 v2
摘要
语言识别作为自动说话人识别的一个重要方面，在过去十年中出现了许多改进性能的变化和新方法。我们比较了使用对数刻度音频频谱与使用从原始音频样本中提取的多音声音序列来训练神经网络并将语音分类为英语或西班牙语的性能。为此，我们采用了一种新颖的方法，即使用卷积循环神经网络，采用长短期记忆（LSTM）或门控循环单元（GRU）进行神经网络的前向传播。我们的假设是，使用多音声音序列作为特征，并以LSTM和GRU作为神经网络的门控机制，其性能优于使用单向深度神经网络的传统MFCC特征。
引用
@article{arxiv.1611.04010,
title = {Multi-Language Identification Using Convolutional Recurrent Neural Network},
author = {Vrishabh Ajay Lakhani and Rohan Mahadev},
journal= {arXiv preprint arXiv:1611.04010},
year = {2017}
}
备注
Further experiments were performed on the model using LibriVox speech dataset and it was found that a Time Distributed CRNN model performed better and represented our initial ideas about the speaker recognition task better. The dataset contains speech in three languages - English, Spanish and Czech. A report on our findings along with experimental results will be published soon