Automatic speech recognition of cochlear implant-like spectrally reduced speech

Cong Thanh Do (1, 2) , Mantha Vijay Kumar (1) , Dominique Pastor (3, 1) , André Goalic (1, 2)

1 SC - Département Signal et Communications
2 Lab-STICC - Laboratoire des sciences et techniques de l'information, de la communication et de la connaissance (UMR 3192)
3 Lab-STICC_TB_CID_TOMS

In this paper, we investigate the behavior of automatic speech recognition (ASR) system with cochlear implant-like spectrally reduced speech (SRS). The SRSs have been synthesized by using amplitude modulation (AM) and frequency modulation(FM) components extracted by using the frequency amplitude modulation encoding (FAME) strategy. In fact, three sets of SRS were synthesized, one by using only AM components and the two others by using both AM and FM components, but with or without subtraction of rapidly varying FM components. Therefore, along with investigating the behavior of ASR using SRS, we investigate the effect of subtracting rapidly varying FM components of SRS in ASR. Recognition results show that even though slowly varying FM components are useful to support human speech recognition, they bring no significant improvement in ASR when the SRS spectral resolution is sufficiently good. The TIDIGITS speech database and HTK speech recognition software were used for the experimentations.

Domaines: Traitement du signal et de l'image [eess.SP]

Cong Thanh Do, Mantha Vijay Kumar, Dominique Pastor, André Goalic. Automatic speech recognition of cochlear implant-like spectrally reduced speech. NCC 2009 : 15th National Conference on Communication, Jan 2009, Guwahati, India. ⟨hal-00725205⟩