Phase-Aware Signal Processing for Automatic Speech Recognition

Abstract

Conventional automatic speech recognition (ASR) often neglects the spectral phase information in its front-end and feature extraction stages. The aim of this paper is to show the impact that enhancement of the noisy spectral phase has on ASR accuracy when dealing with speech signals corrupted with additive noise. Apart from proof-of-concept experiments using clean spectral phase, we also present a phase enhancement method as a phase-aware front-end and modified group delay as a phaseaware feature extractor, and the combination thereof. In experiments, we demonstrate the improved performance for each individual component and their combination, compared to the conventional phase-unaware Mel Frequency Cepstral Coefficients (MFCCs)-based ASR. We observe that the estimated phase information used in the front-end or feature extraction component improves the ASR word accuracy rate (WAR) by 20.98 % absolute for noise corrupted speech (averaged over SNRs ranging from 0 to 20 dB).

The Conversation: Deep Audio-Visual Speech Enhancement

“…Speech enhancement methods have traditionally only dealt with filtering the spectral magnitudes, however many approaches have been recently been proposed for jointly enhancing the magnitude and phase spectra [1,8,9,10,11,12,13]. The prevalent method for estimating phase spectra from given magnitudes in speech synthesis is the one proposed by Griffin and Lim [14].…”

Considerações Sobre o Uso do Sinal de Fase em Sistemas de Reconhecimento Automático de Fala

“…Além disso, pode-se ainda utilizar as estratégias discutidas em [8], [11] e [12] para o tratamento das raízes próximas a circunferência de raio unitário. Dentre elas, destacam-se as que utilizam bancos de filtros em MF.…”

“…Em [10], cepstros complexos são usados como atributos para a aplicação em conversores de texto para fala. Já em [5] e [8], atributos criados a partir das derivadas do espectro de fase no domínio do tempo e da frequência, representados pela frequência instantânea e pelo atraso de grupo (group delay -GD), respectivamente, são utilizados em aplicações de realce do sinal de fala e em sistemas de ASR.…”

“…Na literatura daárea, trabalhos que adotam o GD como atributos discriminativos relatam resultados satisfatórios em sistemas de ASR [6], [8], [11]- [13]. Em [13], a robustez ao ruído dos atributos de GDé discutida.…”

On the Use of Multi-lingual Approach for a Cloud-based Transcription System for the ‘Ilonggoish’ Dialect

“…Automatic speech recognition (ASR) aims to map an audio signal, containing speech, into a text transcription containing a sequence of words. Basically, the goal is to match the transcription as close as possible to the audio message, with no particular understanding of the meaning or scope of what was spoken [2].…”