Transfer Learning for Speech Intelligibility Improvement in Noisy Environments

Biswas, Ritujoy; Nathwani, Karan; Abrol, Vinayak

doi:10.21437/interspeech.2021-150

Cited by 4 publications

(1 citation statement)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Further, it is found that quality [9], [10] and intelligibility [11], [12] affect ASR performance. Therefore, a drop in the performance of the general purpose automatic speech recognition (gp-asr) due to incorrect estimation of RIR is due to deterioration of speech quality [13], [14], [15] and speech intelligibility [16], [17]. It is therefore, plausible to assume that enhancing the performance of gp-asr will result in a better user-experience, which depends on accurate estimation of RIR.…”

Section: Introductionmentioning

confidence: 99%

Learning Optimal Microphone Location for Enhanced ASR Performance using Limited Data

Biswas¹,

Nathwani²,

Kopparapu³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

<p>The placement of a microphone at its correct position is crucial for automatic speech recognition (ASR) in a single-microphone distant speech recognition setup. Understandably, it is a practical problem that the performance degrades due to accidental displacement of the microphone from the correct position through routine use and maintenance. In this paper, an approach is formulated which makes it possible to determine the optimal position of the microphone, given the geometry of the room, the location of the speaker, and the inexact location of the microphone. The metric used to measure the correctness of the microphone position is the average character error rate (cer) of the ASR system at that position. The proposed approach uses limited available data which is usually the case when dealing with microphone position error analysis while considering room acoustics. A neural network classifier is trained using limited data to associate the location of the microphone with the corresponding class of cer values. The obtained prediction is used to trace the microphone back to the optimal position.</p>

show abstract