Uniphore's submission to Fearless Steps Challenge Phase-2

TL;DR

This paper presents supervised CNN-based systems for speech activity detection and speaker identification in the Fearless Steps Challenge Phase-2, achieving competitive results through a shared architecture and innovative voting-based identification.

Contribution

It introduces a unified CNN architecture for both SAD and SID tasks and proposes a two-level voting method for speaker identification.

Findings

Speech detection score of 5.96% on dev set

Speaker retrieval accuracy of over 82%

Shared CNN architecture for SAD and SID tasks

Abstract

We propose supervised systems for speech activity detection (SAD) and speaker identification (SID) tasks in Fearless Steps Challenge Phase-2. The proposed systems for both the tasks share a common convolutional neural network (CNN) architecture. Mel spectrogram is used as features. For speech activity detection, the spectrogram is divided into smaller overlapping chunks. The network is trained to recognize the chunks. The network architecture and the training steps used for the SID task are similar to that of the SAD task, except that longer spectrogram chunks are used. We propose a two-level identification method for SID task. First, for each chunk, a set of speakers is hypothesized based on the neural network posterior probabilities. Finally, the speaker identity of the utterance is identified using the chunk-level hypotheses by applying a voting rule. On SAD task, a detection cost function score of 5.96%, and 5.33% are obtained on dev and eval sets, respectively. A top 5 retrieval accuracy of 82.07% and 82.42% are obtained on the dev and eval sets for SID task. A brief analysis is made on the results to provide insights into the miss-classified cases in both the tasks.