UT-Scope: Towards LVCSR under Lombard effect induced by varying types and levels of noisy background

Abstract

Adverse environments impact the performance of automatic speech recognition systems in two ways -directly by introducing acoustic mismatch between the speech signal and acoustic models, and indirectly by affecting the way speakers communicate to maintain intelligible communication over noise (Lombard effect). Currently, an increasing number of studies have analyzed Lombard effect with respect to speech production and perception, yet limited attention has been paid to its impact on speech systems, especially within a larger vocabulary context. This study presents a large vocabulary speech material captured in the recently acquired portion of UT-Scope database, produced in several types and levels of simulated background noise (highway, crowd, pink). The impact of noisy background variations on speech parameters is studied together with the effects on automatic speech recognition. Front-end cepstral normalization utilizing a modified RASTA filter is proposed and shown to improve recognition performance in a side-byside evaluation with several common and state-of-the-art normalization algorithms.

Robustness in Speech, Speaker, and Language Recognition: “You’ve Got to Know Your Limitations”

“…Figure 3 demonstrates performance of a neutral speech-trained hidden Markov model (HMM) ASR system when tested on TIMITlike [32] utterances produced by speakers that were exposed to three levels of a highway (HWY), large crowd (CRD/LCR), and pink noise (PNK) played back through headhpones (70, 80, and 90 dB SPL for HWY and CRD; 65, 75, 85 dB SPL for PNK). A close-talk microphone channel providing high SNR recordings was used in the ASR experiment on 31 US-born subjects' (25 females, 6 males) drawn from the UT-Scope Lombard Effect set [9] (see [33] for more details on the ASR experiment). It can be seen that the word error rate (WER) grows rapidly from the baseline no-noise Neutral condition once the speakers are exposed to increasing noise levels -while the recorded speech signal retains a high SNR.…”

A comparison of front-end compensation strategies for robust LVCSR under room reverberation and increased vocal effort

“…In our recent study, RASTALP -a modified RASTA filter that approximates the low-pass portion of the original RASTA by a smoothing low pass filter [9] and the high-pass portion by CMN or other segmentbased normalizations [9,18] was introduced. Compared to the original high order band-pass RASTA filter, RASTALP is a filter of significantly lower (2 nd ) order, which helps reduce transient effects typical for RASTA filtering.…”

“…Several studies have considered individual effects of room reverberation [2,3,[5][6][7] and increased vocal effort [8,9] on ASR, and reported compensation strategies for their alleviation. However, to the best of our knowledge, this study is the first to consider the individual as well as the combined effects of reverberation and increased vocal effort on ASR.…”

“…It is shown that post-processing the MHEC with cepstral gain normalization (CGN) [12] combined with modified low-pass RASTA filtering [9], which has been recently introduced for robust ASR under noisy Lombard effect conditions, results in considerable improvement in performance under reverberant modal and high vocal effort speech.…”

An investigation on back-end for speaker recognition in multi-session enrollment

“…We utilized two different acoustic features: (a) Mel frequency Cepstral coefficients (MFCC), MFCCs are normalized by quantile cepstral normalization (QCN) [17] and low-pass RASTA filtering [18], and (b) Rectangular Filter-bank Cepstral Coefficients (RFCC) [17], which are processed through feature warping [19]. All features are 39-dimensional (12 cepstral coefficients+C 0 +∆+∆∆).…”