Building and evaluation of a real room impulse response dataset

Abstract

This paper presents BUT ReverbDB -a dataset of real room impulse responses (RIR), background noises and re-transmitted speech data. The retransmitted data includes LibriSpeech test-clean, 2000 HUB5 English evaluation and part of 2010 NIST Speaker Recognition Evaluation datasets. We provide a detailed description of RIR collection (hardware, software, post-processing) that can serve as a "cook-book" for similar efforts. We also validate BUT ReverbDB in two sets of automatic speech recognition (ASR) experiments and draw conclusions for augmenting ASR training data with real and artificially generated RIRs. We show that a limited number of real RIRs, carefully selected to match the target environment, provide results comparable to a large number of artificially generated RIRs, and that both sets can be combined to achieve the best ASR results. The dataset is distributed for free under a non-restrictive license and it currently contains data from 8 rooms, which is growing. The distribution package also contains a Kaldi-based recipe for augmenting publicly available AMI closetalk meeting data and test the results on an AMI single distant microphone set, allowing it to reproduce our experiments.

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

“…Therefore, these methods will not work well for low-frequency components of the sound, which inevitably introduces significant simulation error at low frequencies under 500Hz compared to accurate wave-based solvers [23]. Far-field ASR experiments that aim to compare the effectiveness of using simulated IRs against real IRs confirm that real IRs are superior in training better ASR systems [9]. The main drawback of geometric acoustic simulation is the absence of low-frequency wave effects such as diffraction [24] and room resonance [25], of which sound diffraction is a less noticeable phenomenon.…”

“…Based on this representation, we calculate and extract sub-band EQs for a set of recorded IRs, collected from the BUT Reverb Database (ReverbDB) [9]. The BUT ReverbDB contains 1891 IRs and 9114 background noises (both with some repetitions), recorded in 9 different real-world environments.…”

“…Recorded IRs, simulated IRs, and compensated IRs are pre-split into disjoint subsets containing 773 and 194 IRs for generating farfield training and development sets, while another 242 recorded IRs are reserved for creating the test set. We randomly sample recorded noises from BUT ReverbDB for all augmented datasets and do not extensively experiment with the noise addition strategy, as that is not the focus of this work and has been studied in [6,9]. The final augmented dataset composition and properties are shown in Table 1.…”

“…This makes it challenging to gather or create large enough far-field ASR training set that works for all situations. Therefore, using data augmentation to enrich existing speech data and approximate far-field speech becomes the most viable solution [6,7,8,9].…”

Speech Enhancement for Noise-Robust Speech Synthesis Using Wasserstein GAN

“…The original utterances are sampled at 48kHz, which we down-sample to 16kHz for faster processing. We used noise signals from the BUTReverbDB database [19] to contaminate clean speech utterances. The noise files consist of recordings from silent office and conference rooms.…”