[Paper Review] Unsupervised Learning of Audio Segment Representations using Sequence-to-sequence Recurrent Neural Networks

Yu-An Chung, Chao-Chung Wu+2 more|arXiv (Cornell University)|Mar 3, 2016

Music and Audio ProcessingComputer Science24 references33 citations

TL;DR

This paper proposes Audio Word2Vec, an unsupervised method that learns fixed-dimensional vector representations for variable-length audio segments using a sequence-to-sequence autoencoder with LSTM units. By jointly training an encoder and decoder to minimize reconstruction error, the model captures sequential phonetic structures, achieving superior performance in query-by-example Spoken Term Detection compared to Dynamic Time Warping, with significantly reduced computational cost.

ABSTRACT

The vector representations of fixed dimensionality for words (in text) offered by Word2Vec have been shown to be very useful in many application scenarios, in particular due to the semantic information they carry. This paper proposes a parallel version, the Audio Word2Vec. It offers the vector representations of fixed dimensionality for variable-length audio segments. These vector representations are shown to describe the sequential phonetic structures of the audio segments to a good degree, with very attractive real world applications such as query-by-example Spoken Term Detection (STD). In this STD application, the proposed approach significantly outperformed the conventional Dynamic Time Warping (DTW) based approaches at significantly lower computation requirements. We propose unsupervised learning of Audio Word2Vec from audio data without human annotation using Sequence-to-sequence Audoencoder (SA). SA consists of two RNNs equipped with Long Short-Term Memory (LSTM) units: the first RNN (encoder) maps the input audio sequence into a vector representation of fixed dimensionality, and the second RNN (decoder) maps the representation back to the input audio sequence. The two RNNs are jointly trained by minimizing the reconstruction error. Denoising Sequence-to-sequence Autoencoder (DSA) is furthered proposed offering more robust learning.

Motivation & Objective

• To develop an unsupervised method for learning fixed-dimensional vector representations of variable-length audio segments.
• To enable semantic and phonetic representation learning in audio without human annotations.
• To improve query-by-example Spoken Term Detection (STD) performance over conventional Dynamic Time Warping (DTW) methods.
• To reduce computational requirements in audio retrieval tasks through learned audio embeddings.

Proposed method

• A sequence-to-sequence autoencoder (SA) is used, consisting of an encoder and decoder, both implemented with Long Short-Term Memory (LSTM) units.
• The encoder maps an input audio sequence into a fixed-dimensional vector representation.
• The decoder reconstructs the original audio sequence from the learned vector representation.
• The model is trained end-to-end by minimizing the reconstruction error between the input and output sequences.
• A denoising variant, Denoising Sequence-to-Sequence Autoencoder (DSA), is introduced to improve robustness by corrupting input sequences during training.
• The learned audio embeddings capture sequential phonetic structures, enabling effective downstream audio retrieval tasks.

Experimental results

Research questions

• RQ1 Can unsupervised learning of audio segment representations be effectively achieved without human annotations?
• RQ2 How well can sequence-to-sequence autoencoders with LSTMs capture phonetic and sequential structure in variable-length audio segments?
• RQ3 Can the learned audio embeddings outperform traditional DTW-based methods in query-by-example Spoken Term Detection?
• RQ4 Does the proposed method reduce computational cost while maintaining or improving retrieval accuracy?

Key findings

• The proposed Audio Word2Vec method significantly outperformed conventional Dynamic Time Warping (DTW) approaches in query-by-example Spoken Term Detection.
• The method achieved higher retrieval accuracy with substantially lower computational requirements than DTW-based systems.
• The sequence-to-sequence autoencoder with LSTMs successfully learned meaningful, fixed-dimensional representations of variable-length audio segments.
• The denoising variant (DSA) improved robustness, suggesting enhanced generalization under noisy or corrupted input conditions.
• The learned audio embeddings effectively captured sequential phonetic structures, enabling semantic and phonetic similarity modeling.
• The unsupervised training paradigm successfully extracted useful audio representations without requiring human-annotated transcriptions.