Content uploaded by Haibin Wu
Author content
All content in this area was uploaded by Haibin Wu on Sep 09, 2024
Content may be subject to copyright.
PARTIALLY FAKE AUDIO DETECTION BY SELF-ATTENTION-BASED
FAKE SPAN DISCOVERY
Haibin Wu1, Heng-Cheng Kuo2, Naijun Zheng4, Kuo-Hsuan Hung2
Hung-Yi Lee1, Yu Tsao2, Hsin-Min Wang3, Helen Meng4
1Graduate Institute of Communication Engineering, National Taiwan University
2Research Center for Information Technology Innovation, Academia Sinica, Taiwan
3Institute of Information Science, Academia Sinica, Taiwan
4Human-Computer Communications Laboratory, The Chinese University of Hong Kong
ABSTRACT
The past few years have witnessed the significant advances of speech
synthesis and voice conversion technologies. However, such tech-
nologies can undermine the robustness of broadly implemented bio-
metric identification models and can be harnessed by in-the-wild at-
tackers for illegal uses. The ASVspoof challenge mainly focuses
on synthesized audios by advanced speech synthesis and voice con-
version models, and replay attacks. Recently, the first Audio Deep
Synthesis Detection challenge (ADD 2022) extends the attack sce-
narios into more aspects. Also ADD 2022 is the first challenge to
propose the partially fake audio detection task. Such brand new
attacks are dangerous and how to tackle such attacks remains an
open question. Thus, we propose a novel framework by introduc-
ing the question-answering (fake span discovery) strategy with the
self-attention mechanism to detect partially fake audios. The pro-
posed fake span detection module tasks the anti-spoofing model to
predict the start and end positions of the fake clip within the par-
tially fake audio, address the model’s attention into discovering the
fake spans rather than other shortcuts with less generalization, and
finally equips the model with the discrimination capacity between
real and partially fake audios. Our submission ranked second in the
partially fake audio detection track of ADD 2022.
Index Terms—Anti-spoofing, partially fake audio detection,
audio deep synthesis detection challenge
1. INTRODUCTION
The past few years have witnessed significant advances in speech
synthesis and voice conversion technologies, and recently emerged
adversarial attacks, such that even humans may not be capable to dis-
tinguish the real users’ speech from the synthesised speech [1–23].
Such technologies can undermine the robustness of broadly imple-
mented biometric identification models, e.g. automatic speaker ver-
ification (ASV) models, and can be harnessed by in-the-wild attack-
ers for criminal usage. For instance, an attacker can generate fake
audios to manipulate the voiceprint-based security entrance system
to accept the attacker falsely, and get access to normally protected in-
formation and valuables. Additionally, an imposter can call the bank
center, fool the biometric identification system to accept him/her as
a registered user, and transfer money to the imposter’s account. Con-
sidering the severe harm caused by synthesized fake audio, it is crit-
ical to devise methods to tackle such threats.
The ASVspoof challenge [1–4], a community-led challenge,
arouses the attention from both the industry and the academia
to tackle the spoofing audios in both physical access and logical
access. In logical access, attacks are mainly from synthesized au-
dios by advanced speech synthesis and voice conversion models,
while in physical access, replayed audios are adopted as attacks.
The challenge attracts various international teams, and various
high-performance anti-spoofing models have been proposed to ad-
dress the two kinds of attacks mentioned above. The adversarial
attacks for ASV and anti-spoofing models have been well investi-
gated [6–9, 14, 15]. To solve further challenging attack situations
in realistic applications, the first Audio Deep Synthesis Detection
challenge (ADD 2022) [5] extends the attack scenarios to fake au-
dio detection. They consider the fake audios perturbed by diverse
background noise, and attacks from the latest speech synthesis and
voice conversion models. Additionally, the organizers propose par-
tially fake audio detection track, where the attacks are composed of
hiding small fake clips into real speech. Partially fake audios are
dangerous, and ADD 2022 is the first challenge attempting to tackle
this type of brand new attacks, which is an open question, and is the
focus of this paper
During generation of partially fake audio, only small clips of
synthetic speech are inserted, and thus the fake audio contains a
large proportion of genuine user’s audio. Through experimentation,
we find it is hard to distinguish the fake and real audios by directly
implementing the previous state-of-the-art spoofing countermeasure
models, such as Light Convolutional Neural Network (LCNN) [24]
and Squeeze-and-Excitation Network (SENet) [25]. To allow the
model discover the small anomalous clip in real speech, we design
a proxy task to make the model answer “where are the start and end
points” of such anomalous clips. During training, the anti-spoofing
model not only has to predict the fake or real label for each utter-
ance, but also to find the start and end positions of the fake clips
within the utterance. Identifying the time segments of the fake clips
is similar to extraction-based question-answering, which determines
the answer span in a document. Also, to further improve the capacity
of the anti-spoofing model to tackle the “question-answering” task,
we introduce the self-attention [26] strategy. The experimental re-
sults illustrate the effective discrimination capacity of our proposed
method between real and partially fake audios.
Our main contributions are two-fold:
• We proposed a brand new framework inspired by the extrac-
tion question-answering strategy for locating the fake regions
in the fake, overall input audio, in order to improve the per-
formance for partially fake audio detection.
arXiv:2202.06684v2 [eess.AS] 15 Feb 2022
Fig. 1. The proposed framework. X, Z, H, A are the acoustic features, hidden features, bottleneck features and the output for Question-answer
layer, respectively. fand gare the SENet feature extractor and self-attention layer, corresponding to (1)-(8) and (9) in Table 1, respectively.
QA and AF are the question-answering (fake span discovery) and anti-spoofing layers with loss calculation procedures respectively.
• We further equipped the fake span discovery strategy with the
self-attention mechanism to get a better detection capacity.
Also, our submission ranked second in the partially fake audio de-
tection track of ADD 2022
This paper is organized as follows: Section 2 introduces the
proposed method, namely self-attention-based question-answering
framework for partially fake audio detection Section 3 presents ex-
perimental setups, followed by section 4 reporting on experimental
results and analysis. Section 5 presents the conclusion.
2. METHODOLOGY
In this section, we will introduce the anti-spoofing method equipped
with the proposed question-answering strategy and self-attention
mechanism. We firstly present the details of the proposed frame-
work. And then we will clarify the rationale of the proposed method.
2.1. Proposed anti-spoofing model
We adopt the base model SENet [25], which is a variant of ResNet
[27] equipped with squeeze-and-excitation networks [28], and we
perform some modifications to that model. The modified model
architecture is shown in Table 1. Let X= [x1, x2, ...xT]denote
the Tframes of input acoustic features. The extracted hidden fea-
tures by the SENet feature extractor are denoted as f(X) = Z=
[z1, z2, ...zT], where fis the (1)-(8) layers in Table 1. The bottle-
neck features are denoted as g(Z) = H= [h1, h2, ...hT], where g
is the self-attention layer, the layer (9) in Table 1, and ht∈Rn. The
self-attention layer is one layer of transformer [26]. The question-
answering layer (a) is one fully-connected layer with the input di-
mension as n, the dimension for ht, and with output dimension as 2.
The 2 dimensions represent how likely will htbe the start or end po-
sition of the fake clip. Given Has the input, the question-answering
layer will output A= [a1, a2, ...aT], where at∈R2. The question-
answering loss Lqa is denoted as:
Lqa =−(log exp(a1
s)
PT
t=1 exp(a1
t)+log exp(a2
e)
PT
t=1 exp(a2
t)),(1)
where sand eare the start and end positions for the fake clip, a1
t
and a2
tare the values for first and second dimensions of atat the tth
frame. We will not incorporate the Lqa for training with real utter-
ances. For the pooling layer (b), there are three pooling strategies
in this paper, average pooling, self-attentive pooling (SAP) [29] and
attentive statistics pooling (ASP) [30]. Based on the bottleneck fea-
tures H, the pooling layer (b) followed by the prediction layer (c)
will output S= [s0, s1], indicating whether the utterance is fake or
real. The anti-spoofing loss Laf is denoted as:
Laf =−log exp(sl)
P1
j=0 exp(sj),(2)
where l∈ {0,1}is the target label. The final loss is
L=Lqa +Laf .(3)
2.2. Rationale
In the partially fake audio detection track, there is only a small pro-
portion of fake audio frames in the overall piece of input speech. Pre-
vious state-of-the-art anti-spoofing models [24, 25] tackle the prob-
lem of identifying whether a whole audio utterance is real or fake.
Hence, previous strategies are not designed to identify anomalous
regions within one utterance. Thus the previous models intuitively
attain the ability to distinguish between utterances but there is no
guarantee that such models can discover the abnormal regions within
a single utterance. To evaluate the performance of the previous state-
of-the-art anti-spoofing models, we direct train binary classification
Table 1. Proposed anti-spoofing model.
layer Type Filter / Stride Output shape
(1) Conv 7×7/1×2 16 ×501 ×40
(2) BatchNorm − −
(3) ReLU − −
(4) MaxPool 3×3/1×2 16 ×501 ×20
(5) SEResNet Module×3−16 ×501 ×20
(6) SEResNet Module×4−32 ×501 ×10
(7) SEResNet Module×6−64 ×501 ×5
(8) SEResNet Module×3−128 ×501 ×3
(9) Self-attention −501 ×384
(a) Question-answering −501 ×2
(b) Pooling −384
(c) Prediction −2
anti-spoofing models for the partially fake audio detection task with
reference to previous papers. We discover that these well-trained
models do not have the discriminative ability for the adaptation set
provided by the organizers of ADD 2022. A reasonable explanation
is that the models may have learned some shortcuts to differenti-
ate the audios with real and fake labels in the training set, but what
the models have learned can not generalize to the adaptation set. In
other words, the models cannot discover the fake regions for fake
audio detection.
Thus, to regularize the model to learn to distinguish between
the real and partially fake audios, we propose a proxy task to let
the model discover the abnormal parts within a piece of partially
fake audio. The proposed anti-spoofing model has to predict not
only whether the input utterance is real or fake, but also output the
start and end of each anomalous region. We name this proxy task
as question-answering, or fake span discovery proxy task, in which
the model has to answer “where is the fake clip” in a piece of par-
tially fake audio. The extraction-based question-answering models
in natural language processing (NLP) often take a question and a
passage as input, build representations for the passage and the ques-
tion respectively, match the question and passage embeddings, and
output the start and end positions within the passage as the answer.
We adopt the analogy of extraction-based question-answering here.
The passage is the partially fake utterance, and the answer span is
the time of the fake clip. By the question-answering proxy task,
the model can learn to find the fake clips within an utterance, thus
benefiting the model to distinguish between the audios with or with-
out fake clips. Moreover, the self-attention module followed by the
question-answering task addresses the model to attend on the fake re-
gions, and helps reduce the question-answering loss, resulting in bet-
ter discrimination capacity between real and partially fake audios.
3. EXPERIMENTAL SETUP
3.1. Data preparation
3.1.1. Dataset construction
The training set and dev set, which are based on Mandarin pub-
licly available corpus AISHELL-3 [31], provided by the organizers
of ADD 2022, cannot be directly adopted to tackle the problem of
partially fake audio detection track, as the whole utterance sample in
them is either real or fake. During the training phase, for construct-
ing fake audios, we generate the partially fake audio by inserting a
clip of audio into the real audios. The inserted clips are derived from
three sources: 1). fake audios in the training and dev set provided
by ADD 2022. 2). Real audios other than the victim audio in the
training and dev set. 3). audios re-synthesised by the traditional
vocoders, including Griffin-Lim [32] and WORLD [33], based on
the real audios in the training and dev set. It is hard to train text-
to-speech (TTS) or voice conversion (VC) models based on the lim-
ited real data provided by the organizer, so we choose the traditional
vocoders, namely Griffin-Lim and WORLD, to increase the diver-
sity of fake audios. As for the validation set, we adopt the adaptation
set consisting of partially fake audios synthesised by ADD2022 for
selecting the models. We report the equal error rate (EER) for the
testing set released by the organizer, as EER is the main evaluation
metric for the partially fake audio detection track.
Table 2. The EERs with (w/) or without (w/o) self-attention.
FFT window size w/o attention w/ attention
384 23.6% 14.3%
768 22.0% 17.9%
3.1.2. Input representations
Mel-spectrograms, which are based on short-time Fourier transform
(STFT) where the window size of fast Fourier transform (FFT) is
varied from 384 to 768, the hop size is 128, and the number of out-
put bins is 80, are used as input features for most of our experiments
and are denoted by MSTFT in following sections. Besides spec-
tral features, some extra experiments are operated on cepstral and
NN-based features to increase diversity for achieving a better per-
formance in the stage of fusion. The FFT window size, hop size, and
number of output bins are fixed to 384, 128, and 80 respectively for
Mel-frequency cepstral coefficients (MFCC), linear frequency cep-
stral coefficients (LFCC), and SincNet [34], as we find the FFT win-
dow size of 384 performs well as shown in Table 3.
3.1.3. Data augmentation
We perform on-the-fly data augmentation by adding noise from MU-
SAN dataset [35], performing room impulse response (RIR) simula-
tion [36] and applying codec algorithms (a-law and µ-law) [37].
3.2. Implementation details
The backbone model is shown in Table 1. Three kinds of attention,
average pooling (Avg), attentive statistics pooling (ASP) and self-
attentive pooling (SAP) are adopted for experiments. All the models
are optimized by Adam with the learning rate of 0.001 and weight
decay as 1e−4.
4. EXPERIMENTAL RESULTS AND ANALYSIS
First of all, we illustrate that the question-answering (QA) strategy
drastically decreases the EERs. We conduct experiments with and
without the QA strategy. The experimental results show that the
trained models without the QA strategy attain the EERs of around
40%, which indicates that such models can not distinguish the par-
tially fake audios from the genuine audios. Due to the poor perfor-
mance of models without the QA strategy on the adaptation set, we
Table 3. The EERs using MSTFT features. w/o or w/ mean with or without. w/ or w/o re-synthesis correspond to using the re-synthesised
audios by Griffin-Lim and WORLD or not.
feature FFT window size pooling method w/o augmentation w/ augmentation
w/o re-synthesis w/ re-synthesis w/o re-synthesis w/ re-synthesis
MSTFT
384 Avg 14.3% 19.9% 11.9% 14.2%
512 Avg 13.2% 20.5% 13.0% 14.8%
640 Avg 18.5% 19.9% 18.9% 13.3%
768 Avg 17.9% 16.8% 14.8% 12.6%
MSTFT
384 SAP 16.9% 17.5% 15.6% 12.6%
512 SAP 17.0% 18.0% 13.9% 12.5%
640 SAP 12.1% 15.3% 15.3% 11.1%
768 SAP 15.2% 17.8% 11.7% 14.8%
MSTFT
384 ASP 17.3% 15.9% 14.9% 11.9%
512 ASP 14.9% 15.8% 12.9% 11.1%
640 ASP 17.5% 15.9% 15.8% 11.2%
768 ASP 14.8% 17.9% 14.5% 22.1%
Table 4. The EERs for three different features
feature MFCC LFCC SincNet
EER 12.5% 11.1% 16.1%
decide not to submit the results on testing sets of such models to the
leaderboard to save the submission times.
Next, we verify the effectiveness of the self-attention layer by
Table 2. As the input and output feature dimensions after the self-
attention layer are the same, the model without the self-attention
layer can be constructed by directly removing (9) in Table 1. In
the following experiments, the performances on the testing set will
be directly displayed. We show the EERs under two settings of FFT
window size due to space limitation, and the other settings are with
the same trend. Table 2 shows that the improvements are signifi-
cant in two settings with different window sizes. The EERs decrease
9.3% and 4.1% absolute after adding self-attention for the FFT win-
dow size of 384 and 768 respectively, which illustrates the significant
improvements by introducing the self-attention layer.
Therefore, the model with self-attention will be adopted for the
following experiments, unless specified otherwise. In the main ex-
periments as shown in Table 3, the input representations are MSTFTs
with hop size of 128, output bins as 80, and FFT window size rang-
ing from 384-768. Table 3 exhausts the experimental settings under
four different window sizes, three pooling strategies, whether to use
the data augmentation and whether to use the re-synthesised fake au-
dios by Griffin-Lim and WORLD. We have the following observa-
tions. First, EERs are improved with the help of data augmentation
in most of the setups. Secondly, enlarging the training set by the re-
synthesised data usually benefits the EERs when data augmentation
is conducted. Lastly, the SAP and ASP pooling significantly im-
prove the EERs when both data re-synthesis and augmentation are
applied. We also can observe that the best EER for a single model is
11.1% shown in Table 3.
In order to increase diversity of models for achieving a better
performance in the stage of model fusion, we further take MFCC,
LFCC and SincNet as input features to train the models. We cannot
exhaust all the settings due to limited computing resources, thus we
refer to Table 3 to select the setting to conduct the experiments. We
fix the FFT window size as 384, apply only ASP pooling, adopt data
augmentation and the re-synthesised data. We observe from Table 4
that the LFCC feature gets EER as 11.1%, reaching the best sin-
gle model performance in our experimental settings. For the further
work, we plan to explore the potential of different front-end features
to get better performance.
For the fusion method, we tried average fusion, weighted aver-
age fusion, min fusion and max fusion. The best submission, which
is fused by the average scores of the top 5 models, achieves the best
7.9% EER and ranks second in partially fake audio detection track.
5. CONCLUSION
Inspired by extraction-based question answering, this paper proposes
a self-attention-based, fake span discovery strategy. The proposed
strategy tasks the anti-spoofing model to predict the start and end
position of the fake clip within the partially fake audio, address the
model’s attention into discovering the fake spans rather than other
patterns with less generalization, and finally equips the model with
the discriminate capacity between real and partially fake audios. Our
final submitted model gave 7.9% EER, and ranked 2nd in the par-
tially fake audio detection track of ADD 2022. Such a strategy can
be model-agnostic and feature-agnostic. Our future work will ex-
plore the potential of the proposed strategy by adopting other back-
bone anti-spoofing models and front-end features.
6. ACKNOWLEDGEMENT
This research is partially funded by the Centre for Perceptual and
Interactive Intelligence, an InnoCentre of The Chinese University of
Hong Kong This work was done while Haibin Wu was a visiting stu-
dent at Human-computer Communications Laboratory, The Chinese
University of Hong Kong.
7. REFERENCES
[1] Z. Wu, T. Kinnunen, et al., “Asvspoof 2015: the first auto-
matic speaker verification spoofing and countermeasures chal-
lenge,” in Sixteenth Annual Conference of the International
Speech Communication Association, 2015.
[2] T. Kinnunen, M. Sahidullah, et al., “The asvspoof 2017 chal-
lenge: Assessing the limits of replay spoofing attack detec-
tion,” 2017.
[3] M. Todisco, X. Wang, et al., “Asvspoof 2019: Future hori-
zons in spoofed and fake audio detection,” arXiv preprint
arXiv:1904.05441, 2019.
[4] J. Yamagishi, X. Wang, et al., “Asvspoof 2021: accelerat-
ing progress in spoofed and deepfake speech detection,” arXiv
preprint arXiv:2109.00537, 2021.
[5] J. Yi, R. Fu, J. Tao, S. Nie, H. Ma, C. Wang, T. Wang, Z. Tian,
Y. Bai, S. Liang, S. Wang, S. Zhang, X. Yan, L. Xu, and H. Li,
“Add 2022: the first audio deep synthesis detection challenge,”
in ICASSP. IEEE, 2022.
[6] H. Wu, A. T. Liu, and H.-y. Lee, “Defense for black-box
attacks on anti-spoofing models by self-supervised learning,”
arXiv preprint arXiv:2006.03214, 2020.
[7] H. Wu, S. Liu, H. Meng, and H.-y. Lee, “Defense against
adversarial attacks on spoofing countermeasures of asv,” in
ICASSP. IEEE, 2020, pp. 6564–6568.
[8] Z. Peng, X. Li, and T. Lee, “Pairing weak with strong: Twin
models for defending against adversarial attack on speaker ver-
ification,” in Proc. INTERSPEECH, 2021.
[9] H. Wu et al., “Adversarial defense for automatic speaker ver-
ification by cascaded self-supervised learning models,” in
ICASSP. IEEE, 2021, pp. 6718–6722.
[10] S. Liu, H. Wu, H.-y. Lee, and H. Meng, “Adversarial attacks on
spoofing countermeasures of automatic speaker verification,”
in ASRU. IEEE, 2019, pp. 312–319.
[11] X. Li et al., “Replay and synthetic speech detection with
res2net architecture,” in ICASSP. IEEE, 2021, pp. 6354–6358.
[12] X. Li, X. Wu, H. Lu, X. Liu, and H. Meng, “Channel-wise
gated res2net: Towards robust detection of synthetic speech
attacks,” arXiv preprint arXiv:2107.08803, 2021.
[13] H. Wu, Y. Zhang, Z. Wu, D. Wang, and H.-y. Lee, “Voting for
the right answer: Adversarial defense for speaker verification,”
arXiv preprint arXiv:2106.07868, 2021.
[14] H. Wu et al., “Improving the adversarial robustness for speaker
verification by self-supervised learning,” arXiv preprint
arXiv:2106.00273, 2021.
[15] H. Wu et al., “Spotting adversarial samples for speaker verifi-
cation by neural vocoders,” arXiv preprint arXiv:2107.00309,
2021.
[16] Z. Wu, N. Evans, T. Kinnunen, J. Yamagishi, F. Alegre, and
H. Li, “Spoofing and countermeasures for speaker verification:
A survey,” speech communication, vol. 66, pp. 130–153, 2015.
[17] Z. Wu, S. Gao, E. S. Cling, and H. Li, “A study on replay attack
and anti-spoofing for text-dependent speaker verification,” in
APSIPA. IEEE, 2014, pp. 1–5.
[18] M. R. Kamble et al., “Advances in anti-spoofing: from the
perspective of asvspoof challenges,” APSIPA Transactions on
Signal and Information Processing, vol. 9, 2020.
[19] R. K. Das, J. Yang, and H. Li, “Assessing the scope of gener-
alized countermeasures for anti-spoofing,” in ICASSP. IEEE,
2020, pp. 6589–6593.
[20] W. Chenglong, Y. Jiangyan, et al., “Global and temporal-
frequency attention based network in audio deepfake detec-
tion,” Journal of Computer Research and Development, 2021.
[21] H. Ma, J. Yi, J. Tao, Y. Bai, Z. Tian, and C. Wang, “Continual
learning for fake audio detection.,” in INTERSPEECH, 2021,
pp. 886–890.
[22] J. Yi, Y. Bai, J. Tao, Z. Tian, C. Wang, T. Wang, and R. Fu,
“Half-truth: A partially fake audio detection dataset,” in IN-
TERSPEECH, 2021, pp. 1654–1658.
[23] X. Wang and J. Yamagishi, “A comparative study on recent
neural spoofing countermeasures for synthetic speech detec-
tion,” arXiv preprint arXiv:2103.11326, 2021.
[24] G. Lavrentyeva, S. Novoselov, E. Malykh, A. Kozlov, O. Kuda-
shev, and V. Shchemelinin, “Audio replay attack detection with
deep learning frameworks.,” in Interspeech, 2017, pp. 82–86.
[25] C.-I. Lai, N. Chen, J. Villalba, and N. Dehak, “Assert: Anti-
spoofing with squeeze-excitation and residual networks,” arXiv
preprint arXiv:1904.01120, 2019.
[26] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones,
A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all
you need,” in Advances in neural information processing sys-
tems, 2017, pp. 5998–6008.
[27] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning
for image recognition,” in CVPR, 2016, pp. 770–778.
[28] J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation net-
works,” in Proceedings of the IEEE conference on computer
vision and pattern recognition, 2018, pp. 7132–7141.
[29] G. Bhattacharya, M. J. Alam, and P. Kenny, “Deep speaker
embeddings for short-duration speaker verification,” in Inter-
speech, 2017, pp. 1517–1521.
[30] K. Okabe, T. Koshinaka, and K. Shinoda, “Attentive statistics
pooling for deep speaker embedding,” in Interspeech, 2018,
pp. 2252–2256.
[31] Y. Shi, H. Bu, X. Xu, S. Zhang, and M. Li, “Aishell-3: A multi-
speaker mandarin tts corpus and the baselines,” arXiv preprint
arXiv:2010.11567, 2020.
[32] D. Griffin and J. Lim, “Signal estimation from modified
short-time fourier transform,” IEEE Transactions on acous-
tics, speech, and signal processing, pp. 236–243, 1984.
[33] M. Morise, F. Yokomori, and K. Ozawa, “World: a vocoder-
based high-quality speech synthesis system for real-time appli-
cations,” IEICE TRANSACTIONS on Information and Systems,
vol. 99, no. 7, pp. 1877–1884, 2016.
[34] M. Ravanelli and Y. Bengio, “Speaker recognition from raw
waveform with sincnet,” in 2018 IEEE Spoken Language Tech-
nology Workshop (SLT). IEEE, 2018, pp. 1021–1028.
[35] D. Snyder, G. Chen, and D. Povey, “Musan: A music, speech,
and noise corpus,” arXiv preprint arXiv:1510.08484, 2015.
[36] T. Ko, V. Peddinti, et al., “A study on data augmentation of
reverberant speech for robust speech recognition,” in ICASSP,
2017, pp. 5220–5224.
[37] C. Recommendation, “Pulse code modulation (pcm) of voice
frequencies,” in ITU. 1988.
