Spectral feature mapping with mimic loss for robust speech recognition

Bagchi, Deblin; Plantinga, Peter; Stiff, Adam; Fosler‐Lussier, Eric

doi:10.48550/arxiv.1803.09816

Cited by 1 publication

(2 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…( 1). Compared to IRM (L1), IRM (CGAN) has an additional adversarial loss term with λ = 0.01 as in (Bagchi et al, 2018;Pascual et al, 2017). A parameter exploring policy gradients (Sehnke et al, 2010) based black-box optimization, which is similar to the one used in (Zhang et al, 2018), is also compared.…”

Section: Objective Evaluation With Different Loss Functionsmentioning

confidence: 99%

“…The deep models were then optimized by minimizing the distance between generated speech and clean speech. However, the distance (objective function) is usually based on simple L p loss (where p = 1 or 2), which does not reflect human auditory perception or ASR accuracy (Bagchi et al, 2018) well. In fact, several researches have indicated that an enhanced speech with a smaller L p distance, does not guarantee a higher quality or intelligibility score (Fu et al, 2018b;Koizumi et al, 2018).…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

MetricGAN: Generative Adversarial Networks based Black-box Metric Scores Optimization for Speech Enhancement

Fu¹,

Liao²,

Tsao³

et al. 2019

Preprint

View full text Add to dashboard Cite

Adversarial loss in a conditional generative adversarial network (GAN) is not designed to directly optimize evaluation metrics of a target task, and thus, may not always guide the generator in a GAN to generate data with improved metric scores. To overcome this issue, we propose a novel MetricGAN approach with an aim to optimize the generator with respect to one or multiple evaluation metrics. Moreover, based on Metric-GAN, the metric scores of the generated data can also be arbitrarily specified by users. We tested the proposed MetricGAN on a speech enhancement task, which is particularly suitable to verify the proposed approach because there are multiple metrics measuring different aspects of speech signals. Moreover, these metrics are generally complex and could not be fully optimized by L p or conventional adversarial losses.

show abstract

Section: Objective Evaluation With Different Loss Functionsmentioning

confidence: 99%