Unsupervised Learning of Audio Segment Representations using Sequence-to-sequence Recurrent Neural Networks

Representing audio segments expressed with variable-length acoustic feature sequences as fixed-length feature vectors is usually needed in many speech applications, including speaker identification, audio emotion classification and spoken term detection (STD). In this paper, we apply and extend sequence-to-sequence learning framework to learn representations for audio segments without any supervision. The model we used is called Sequence-to-sequence Autoencoder (SA), which consists of two RNNs equipped with Long Short-Term Memory (LSTM) units: the first RNN acts as an encoder that maps the input sequence into a vector representation of fixed dimensionality, and the second RNN acts as a decoder that maps the representation back to the input sequence. The two RNNs are then jointly trained by minimizing the reconstruction error. We further propose Denoising Sequence-to-sequence Autoencoder (DSA) that improves the learned representations. The vector representations learned by SA and DSA are shown to be very helpful for query-by-example STD. The experimental results have shown that the proposed models achieved better retrieval performance than using audio segment representation designed heuristically and the classical Dynamic Time Warping (DTW) approach.