Low-Frequency Compensated Synthetic Impulse Responses For Improved Far-Field Speech Recognition

Abstract

We propose a method for generating low-frequency compensated synthetic impulse responses that improve the performance of farfield speech recognition systems trained on artificially augmented datasets. We design linear-phase filters that adapt the simulated impulse responses to equalization distributions corresponding to realworld captured impulse responses. Our filtered synthetic impulse responses are then used to augment clean speech data from Lib-riSpeech dataset [1]. We evaluate the performance of our method on the real-world LibriSpeech test set. In practice, our low-frequency compensated synthetic dataset can reduce the word-error-rate by up to 8.8% for far-field speech recognition.

TS-RIR: Translated Synthetic Room Impulse Responses for Speech Augmentation

“…• We show that, on a modified Kaldi LibriSpeech far-field ASR benchmark [2], far-field speech augmented using our improved RIRs outperforms the far-field speech augmented using unmodified RIRs by up to 19.9%.…”

“…On the other hand, we observe a boost or diminishing effect in the frequency response at different frequency bands in real RIRs due to wave modes created by room resonance. Some methods tend to compensate for the missing room response in synthetic RIRs using a sub-band room equalization approach [2].…”

“…We evaluate the performance of improved synthetic RIRs on a far-field speech dataset augmented by convolving the LibriSpeech clean speech dataset [1] with RIRs and adding background noise. We show that far-field speech augmented using our improved synthetic RIRs reduces the word error rate by up to 19.9% in Kaldi far-field automatic speech recognition benchmark [2].…”

An Overview of Methods for Generating, Augmenting and Evaluating Room Impulse Response Using Artificial Neural Networks

“…The authors of the IR-GAN algorithm also evaluate the generated RIRs by performing an ASR test, using the Kaldi LibriSpeech acoustic model, which is based on the (Tang et al, 2020b). The paper states that the proposed algorithm can reduce WER by almost 9% compared to the GAS method.…”

A study on more realistic room simulation for far-field keyword spotting

“…However, the collection of a sufficiently-varied dataset with multiple positions per room is a tedious task and needs to be repeated for different microphone geometries. There have been numerous studies on the differences between using simulated and measured RIRs, with several attempts to close the performance gap: point-source noises [8], directional sources [12], and artificially mimicking low-frequency wave effects [13].…”