All content in this area was uploaded by Titouan Parcollet on Mar 18, 2020
Content may be subject to copyright.
All content in this area was uploaded by Titouan Parcollet on Nov 01, 2019
Content may be subject to copyright.
E2E-SINCNET: TOWARD FULLY END-TO-END SPEECH RECOGNITION
Titouan Parcollet?‡† Mohamed Morchid?Georges Linar`
‡University of Oxford, UK
Modern end-to-end (E2E) Automatic Speech Recognition (ASR)
systems rely on Deep Neural Networks (DNN) that are mostly
trained on handcrafted and pre-computed acoustic features such as
Mel-ﬁlter-banks or Mel-frequency cepstral coefﬁcients. Nonethe-
less, and despite worse performances, E2E ASR models processing
raw waveforms are an active research ﬁeld due to the lossless nature
of the input signal. In this paper, we propose the E2E-SincNet, a
novel fully E2E ASR model that goes from the raw waveform to
the text transcripts by merging two recent and powerful paradigms:
SincNet and the joint CTC-attention training scheme. The conducted
experiments on two different speech recognition tasks show that our
approach outperforms previously investigated E2E systems relying
either on the raw waveform or pre-computed acoustic features, with
a reported top-of-the-line Word Error Rate (WER) of 4.7% on the
Wall Street Journal (WSJ) dataset.
Index Terms—End-to-end speech recognition, SincNet.
ASR systems are either hybrid DNN-HMM or end-to-end (E2E).
The former set of ASR models provides state-of-the-art perfor-
mances on numerous speech-related real-world tasks [1, 2] but
involves multiple sub-blocks trained separately, and often requires
separate and a strong human expertise. E2E systems, on the other
hand, propose to directly transcribe a sequence of acoustic input fea-
tures [3, 4] with a single model usually composed of different Neural
Networks (NN) trained jointly in an end-to-end manner. In particu-
lar, a major challenge is to automatically generate an alignment from
the raw signal that often contains several thousands of data point per
second, to the text, only consisting of a single character or concept
in the same time scale.
Recently, E2E approaches started to outperform traditional
DNN-HMM baselines on common speech recognition tasks with
the introduction of more efﬁcient sequence training objectives
[5, 6], more powerful architectures [7, 8] and attention mecha-
nisms [9, 10, 11, 12]. Despite being named “E2E”, the latter models
still require pre-processed acoustic features such as Mel-ﬁlter-banks,
alleviating a pure E2E pipeline based on the raw audio signal.
Processing raw waveforms in the speciﬁc context of ASR is an
active challenge [13, 14, 15, 7]. Most of these works rely on modi-
ﬁed Convolutional Neural Networks (CNNs) to operate over the sig-
nal. As an example, in , the authors propose to combine a log
non-linearity with a CNN architecture that exactly matches an output
dimension equivalent to standard Mel-ﬁlter-banks features, forcing
the input layer to learn the latter signal transformation. Nevertheless,
and as demonstrated in , CNNs are not efﬁcient at learning com-
mon acoustic features due to the lack of constraint on the numerous
trainable parameters. Consequently, the authors proposed SincNet, a
speciﬁc convolutional layer that integrates common acoustic ﬁlters,
such as band-pass ﬁlters, to replace the convolutional kernel weights
drastically reducing the number of parameters. Furthermore, it is
demonstrated that the learned ﬁlters have a much better frequency re-
sponse than those learned with traditional CNNs, resulting in better
performances in a speaker recognition task. Then, SincNet has been
combined with a straightforward fully-connected DNN in the con-
text of a DNN-HMM ASR system also outperforming CNNs trained
with both pre-computed acoustic features and raw waveforms .
Unfortunately, there is no available model combining both the
efﬁcacy of SincNet to operate over raw signals, and the latest train-
ing scheme for E2E systems. Therefore, we propose to bridge
this gap by investigating and releasing1a fully E2E model, named
E2E-SincNet, combining SincNet with the joint CTC-attention
training scheme  and resulting in a customizable, efﬁcient and
interpretable E2E ASR system. Contributions of the paper are
1. Enhance the original SincNet to ﬁt bi-directional recurrent
neural networks (RNN).
2. Merge the later model with the joint CTC-attention method
 to create E2E-SincNet1based on the well-known ESPnet
toolkit  (Section 2).
3. Evaluate the model alongside with other baseline models on
the WSJ and TIMIT speech recognition tasks (Section 3).
The conducted experiments show that E2E-SincNet obtains su-
perior and state-of-the-art (SOTA) performances to both traditional
E2E models operating on raw waveform with CNNs, and SOTA E2E
architectures relying on pre-computed acoustic features.
2. END-TO-END SPEECH RECOGNITION
This section introduces the necessary building blocks to conceive a
fully E2E automatic speech recognition system. First, latent acoustic
features are extracted from the raw waveform signal with a speciﬁc
kernelized CNN, also known as SincNet  (Section 2.1). The
latter model is then merged with a joint CTC-attention  training
procedure (Section 2.2.1), based on an encoder-decoder architecture
 (Section 2.2.2).
2.1. Processing raw waveforms with SincNet
Traditional parametric CNNs operate over the raw waveform by per-
forming multiple time-domain convolutions between the input signal
and a certain ﬁnite impulse response  as:
1Code is available at: https://github.com/TParcollet/E2E-SincNet
y[n] = x[n]×f[n] =
with x[n]a part of the speech signal, f[n]a ﬁlter of length L, and
y[n] the output ﬁnally ﬁltered. In this case, all the elements of f
are learnable parameters. SincNet proposes to replace fwith a pre-
deﬁned function gthat only depends on much fewer parameters to
describe its behavior. In , the authors implemented gas a ﬁlter-
bank composed of rectangular bandpass ﬁlters. Such function can
be written in the time domain as:
g[n, f1, f2] = 2f2sinc(2πf2n)−2f1sinc(2πf1n),(2)
with f1and f2the two learnable parameters that describe the low
and high cutoff frequencies of the bandpass ﬁlters, and sinc(x) =
x. Such parameters are randomly initialized in the interval
2], with fsequal to the input signal frequency sampling. It is
also important to notice that gis smoothed based on the Hamming
Other deﬁnitions of the ﬁlter ghave been proposed including
triangular, Gammatone, and Gaussian ﬁlters  demonstrating su-
perior performances over Eq. 2 due to better ﬁlter responses to the
signal. As a matter of fact, SincNet allows an important ﬂexibility to
efﬁciently enhance traditional acoustic-based CNNs with prior and
well-investigated knowledge. Finally, SincNet ﬁlters are facilitating
the interpretability of the model by being easily extracted and ap-
plied over any signal for further investigations of the transformations
[16, 22, 21].
Unfortunately, SincNet has only been investigated with a mere
fully-connected DNN based on a hybdrid DNN-HMM setup [17,
21]. We propose to connect SincNet to a recurrent encoder-decoder
structure trained in a complete E2E manner following the joint CTC-
attention procedure .
2.2. Joint CTC-attention models
2.2.1. Connectionist Temporal Classiﬁcation
In E2E ASR systems, the task of sequence-to-sequence mapping
from an input acoustic signal X= [x1, ..., xn]to a sequence of
symbols T= [t1, ..., tm]is complex due to: 1) Xand Tcould be in
arbitrary length; 2) The alignment between Xand Tis unknown in
most cases; 3) Tis usually shorter than Xin terms of symbols.
To alleviate these problems, connectionist temporal classiﬁca-
tion (CTC) has been proposed . First, a softmax is applied at
each timestep, or frame, providing a probability of emitting each
symbol Xat that timestep. This probability results in a symbol se-
quences representation P(O|X), with O= [o1, ..., on]in the latent
space O. A blank symbol 0−0is introduced as an extra label to allow
the classiﬁer to deal with the unknown alignment. Then, Ois trans-
formed to the ﬁnal output sequence with a many-to-one function z(.)
deﬁned as follows:
z(o1, o2,−, o3,−)
z(o1, o2, o3, o3,−)
z(o1,−, o2, o3, o3)
= (o1, o2, o3).(3)
Consequently, the output sequence is a summation over the
probability of all possible alignments between Xand Tafter apply-
ing the function z(O). Accordingly to  the parameters of the
models are learned based on the cross entropy loss function:
During the inference, a best path decoding algorithm is performed.
Therefore, the latent sequence with the highest probability is ob-
tained by performing the argmax of the sof tmax output at each
timestep. The ﬁnal sequence is obtained by applying the function
z(.)to the latent sequence.
2.2.2. Attention-based encoder-decoder
Conversely to CTC, encoder-decoder models  do not suffer
from a forced many-to-one mapping. Indeed, the input signal
X= [x1, ..., xn]is entirely consumed by a ﬁrst encoder neural
network (e.g. a recurrent neural network), before being fed to a sec-
ond one that is free to emit any number of outputs T= [t1, ..., tm]
starting from the information contained in the last latent space of
the encoder. Major bottlenecks are therefore related to the ability
of the encoder to map correctly an entire sequence to an arbitrary
latent space, and to the decoder that is not aware of the sequential
order of the input signal. To alleviate these issues, attention-based
encoder-decoder have been proposed .
From a high-level perspective, an attention-based decoder is able
to look over the complete set of the hidden states generated by the
encoder, making it feasible to “choose” the relevant information in
the time-domain . More precisely, an attention-based encoder-
decoder consists of two RNNs. The encoder part remains mostly
unaltered and maps an input sequence of arbitrary length n,X=
[x1, ..., xn]to nhidden vectors h,(h1, ..., hn). Then, the attention-
decoder generates moutput distributions O= [o1, ..., om]corre-
sponding to mtimesteps, by attending to both the nencoded hand
the previously generated token ot−1. Two special tokens, denot-
ing the start-of-sentence and end-of-sentence are added to integrate
boundaries. The loss function is nearly identical to CTC, except that
a condition on the previous ground-truth token (otruth
t−1) is added :
Following , our proposed E2E model relies on a location-
based attention method . The attention weight vector at time
step tnamed atidentiﬁes the focus location in the entire encoded
hidden sequence hat time step twith a context vector ct:
at,ihi, at,i =exp(γet,i)
γis the sharpening factor  and et,k relates the importance or
energy of the kannotation vector for predicting the ioutput token.
In the case of a location-based attention, et,k is computed as:
et,k =wTtanh(W st−1+vhk+U(F∗at−1) + b, (7)
with w,W,V,F,Yand bare different trainable parameters, st−1is
the hidden state of the decoder at the previous time step, and ∗is the
convolution product. It is important to note that the implementation
of evaries accordingly to the type of attention mechanism employed
. Then, the decoder generates an output token otfrom an input
vector based on both the context vector ctand the previous state
Fig. 1. Illustration of the proposed E2E-SincNet. The batch dimension is omitted for readability. The raw signal is ﬁrst encoded with a
SincNet layer, followed by multiple 1D convolutions and a bidirectional RNN. Then a CTC and an attention-based decoders emit a sequence
of text symbols and are trained jointly.
st−1alongside with updating the current state stfollowing RNNs
equations with st−1,ctand ot. Unfortunately, ASR systems solely
relying on an attention mechanism are highly perturbed by noisy
data that generate wrong alignments . Furthermore, it has been
shown that it is difﬁcult to train such models from scratch on wide
input sequences [25, 9].
2.2.3. Joint CTC-attention
To overcome the limitations of both CTC training and attention-
based encoder-decoder models and to beneﬁt from their strengths,
 introduced the joint CTC-attention paradigm. The key idea of
the latter method relies on the introduction of the CTC loss as an
auxiliary task to the attention-based encoder-decoder training. More
precisely, both losses are combined and controlled with a ﬁxed hy-
perparameter λ(0≤λ≤1) as:
Ljoint = (1 −λ)Lenc,dec +λLCT C .(8)
SincNet has only been combined with mere feed-forward NNs [16,
17], while the joint CTC-attention approach has only been applied
to pre-computed acoustic features such as MFCCs and Mel-ﬁlter-
banks [5, 18]. We propose to combine SincNet to the latter training
procedure in a efﬁcient, interpretable and fully E2E ASR approach
The E2E architecture is composed of three components: 1) An
encoder that operates over the raw audio signal with a ﬁrst SincNet
layer followed by None-dimensional convolutional layers. The la-
tent features are then consumed by a traditional bidirectional RNN.
2) A simple CTC decoder that produces a token for each time step
encoded. 3) An attention-based decoder that looks out over the en-
tire encoded hidden sequence to output the right symbol. The model
is trained following the joint CTC-attention loss function (Eq. 8).
In this Section, E2E-SincNet is compared to other state-of-the-art
end-to-end ASR systems with two different speech recognition tasks.
First, datasets alongside with pre-computed and raw acoustic fea-
tures are detailed (Section 3.1). Then, baselines and proposed mod-
els architectures are described (Section 3.2). Finally, we report and
discuss the results in Section 3.3.
3.1. Speech recognition datasets and acoustic features
E2E-SincNet is evaluated in two different tasks of phoneme recog-
nition with the TIMIT dataset, and word recognition with the Wall
Street Journal corpus.
3.1.1. The TIMIT phoneme recognition task
The TIMIT  dataset is composed of a standard 462-speaker
training dataset, a 50-speakers development dataset and a core test
dataset of 192 sentences for a total of 5hours of clean speech. Dur-
ing the experiments, the SA records of the training set are removed
and the development set is used for early stopping. The accuracy is
reported in terms of Phoneme Error Rate (PER). TIMIT is consid-
ered as a challenging task for E2E systems due to its very limited
amount of available training data (less than 5hours).
3.1.2. The Wall Street Journal speech recognition task
Only the full “train-si284” dataset is considered as a training set (81
hours), due to the fact that the models have already been evaluated
on the smaller TIMIT dataset. The usual “test-eval92” is used at test-
ing time, while “test-dev93” is considered as a validation dataset.
The accuracy is reported in terms of Word Error Rate (WER).
3.1.3. Acoustic features
In the original SincNet proposal , chunks of raw signal are cre-
ated every 400ms with a 10ms overlapping. Instead, we propose to
split the waveform of each speech sentence into blocks of 25ms. In-
deed,  introduce a SincNet followed by a DNN that requires both
right and left contexts to be trained properly. Our approach relies
on a combination of SincNet with a RNN allowing the latter con-
text to be captured within the recurrent connections, making it fea-
sible to drastically reduce both the input dimension and the VRAM
consumption at training time (i.e. by a factor of 5). Then, other
E2E systems usually process either pre-computed acoustic features.
Therefore, 23 and 80 Mel-ﬁlter-banks are extracted for the TIMIT
and WSJ datasets respectively, based on windows of size 25ms with
3.2. Models architectures
Two different E2E ASR models operating on the raw waveform
and relying on the encoder-decoder approach with the joint CTC-
attention training scheme are introduced (see Figure 1).
E2E-SincNet. The encoder is made of a speciﬁc SincNet layer
and 3one-dimensional convolutional layers with 256 −128 −128
ﬁlters, followed by a bidirectional LSTM composed with 4or 6lay-
ers of size 512 for the TIMIT and WSJ tasks respectively. A one-
dimensional maxpooling of length 3is applied after the convolu-
tional and SincNet layers to reduce the signal dimension. In ,
the authors introduced a SincNet layer composed of 128 ﬁlters of
size 251. We propose to increase the number of ﬁlters to 512 and
to decrease their size to 129 to enhance the local resolution of the
ﬁlters, better ﬁtting to the task of speech recognition. Finally, the
decoder relies on a simple attention layer of size 512 combined with
the CTC loss (Section 2.2.3).
E2E-CNN. This architecture is proposed to highlight the impact
of the SincNet layer in E2E-SincNet. More precisely, E2E-CNN is
identical to E2E-SincNet but with a traditional convolutional layer
with 512 ﬁlters to replace the SincNet one.
Models are trained based on the Adadelta optimizer with vanilla
hyperparameters  for 20 and 15 epochs during the TIMIT and
WSJ tasks respectively. The joint CTC-attention loss control hyper-
parameter λ(Eq. 8) is set to 0.5for the TIMIT experiments and
decreased to 0.2with WSJ. No dropout is applied and the results
observed on the test dataset are reported with respect to the best per-
formances obtained on the validation dataset.
3.3. Results and discussions
Table 1 reports the results obtained by our approaches compared to
a more traditional E2E model operating on Mel-ﬁlter-banks on the
TIMIT dataset. First, it is worth underlining that the E2E-SincNet
obtains the best performances with a PER of 19.3% on the test
dataset, compared to 20.5% for the baseline and 21.1% for the non-
SincNet alternative representing a relative improvement of 1.2%
and 1.3% respectively. Unfortunately, TIMIT is a very challenging
task for E2E systems due to the small amount of available training
data (less than 5hours), resulting in worse performances in com-
parison to hybrid DNN-HMM ASR systems . Therefore, it is of
crucial interest to scale the E2E-SincNet model to a larger dataset to
validate its suitability to real-world tasks.
Table 2 reports the performances obtained by various SOTA E2E
models on the WSJ dataset by integrating a 3-gram recurrent lan-
guage model (RNNLM) . “Jasper” uses a transformerXL lan-
guage model. First, the proposed E2E-SincNet obtains a top-of-line
Table 1. Results obtained with different E2E ASR systems on the
TIMIT phoneme recognition tasks. “Fea.” details the type of input
features employed, and “Valid.” denotes the validation dataset.
Results are expressed in Phoneme Error Rate (i.e. lower is better).
Models Fea. Valid. % Test %
E2E-CNN RAW 18.9 21.1
ESPnet (VGG)  FBANK 17.9 20.5
E2E-SincNet RAW 17.3 19.3
WER of 4.5% on the “test eval92” dataset, outperforming all the
baselines. Indeed, a previous best score of 5.9% was reported in
, highlighting a relative improvement of 1.2%.
Table 2. Results obtained with different E2E ASR systems on the
WSJ dataset. “Fea.” details the type of input features employed,
“Valid.” denotes the validation dataset, “-ASG” is the auto segmen-
tation criterion (i.e a variation of CTC) and “-Att.” is attention only.
Results are expressed in Word Error Rate (i.e. lower is better).
Models Fea. Valid. Test
BiGRU-Att.  FBANK - 9.3
Wav2Text  FBANK 12.9 8.8
Jasper  FBANK 9.3 6.9
E2E-CNN RAW 9.8 6.5
ESPnet (VGG)  FBANK 9.7 6.4
CNN-GLU-ASG  RAW 8.3 6.1
SelfAttention-CTC  FBANK 8.9 5.9
E2E-SincNet RAW 7.8 4.7
The E2E-SincNet outperforms the E2E-CNN with a relative
gain of 1.8% on both TIMIT and WSJ task. This demonstrates
the efﬁcacy of the SincNet layer to learn an expressive ﬁltered
signal enabling a better and lossless latent representation of the
raw waveform. It is interesting to note that “transformer“ models
have recently obtained better performances on multiple ASR tasks
. Nonetheless, transformers are a speciﬁc architecture that dif-
fer signiﬁcantly from the presented models and are therefore not
considered in our benchmarks.
Summary. In this paper, we introduced E2E SincNet, a fully end-
to-end automatic speech recognition system able to process the raw
waveform based on an adaptation of the recent SincNet with the
powerful joint CTC-attention training paradigm. The conducted
experiments on two different speech recognition-related tasks have
demonstrated the superiority of our approach over various other E2E
systems based on both pre-computed acoustic features and the raw
waveform, achieving one of the best result observed so far with an
E2E ASR model on the Wall Street Journal dataset.
Future work. SincNet currently suffers from various issues. First, it
is important to investigate other ﬁlters to efﬁciently operate over the
raw signal. Then, an alternative to the maxpooling must be explored
to alleviate the risk of aliasing in the ﬁltered signal.
Acknowledgments. This work was supported by the AISSPER
project through the French National Research Agency (ANR) under
Contract AAPG 2019 ANR-19-CE23-0004-01 and by the Engineer-
ing and Physical Sciences Research Council (EPSRC) under Grant:
 Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukas Bur-
get, Ondrej Glembek, Nagendra Goel, Mirko Hannemann, Petr
Motlicek, Yanmin Qian, Petr Schwarz, Jan Silovsky, Georg
Stemmer, and Karel Vesely, “The kaldi speech recognition
toolkit,” in IEEE 2011 Workshop on Automatic Speech Recog-
nition and Understanding. Dec. 2011, IEEE Signal Processing
Society, IEEE Catalog No.: CFP11SRW-USB.
 M. Ravanelli, T. Parcollet, and Y. Bengio, “The pytorch-kaldi
speech recognition toolkit,” in In Proc. of ICASSP, 2019.
 Alex Graves and Navdeep Jaitly, “Towards end-to-end speech
recognition with recurrent neural networks,” in International
Conference on Machine Learning, 2014, pp. 1764–1772.
 Ying Zhang, Mohammad Pezeshki, Phil´
Saizheng Zhang, Cesar Laurent Yoshua Bengio, and Aaron
Courville, “Towards end-to-end speech recognition with
deep convolutional neural networks,” arXiv preprint
 Suyoun Kim, Takaaki Hori, and Shinji Watanabe, “Joint ctc-
attention based end-to-end speech recognition using multi-task
learning,” in 2017 IEEE international conference on acous-
tics, speech and signal processing (ICASSP). IEEE, 2017, pp.
 Yiming Wang, Tongfei Chen, Hainan Xu, Shuoyang Ding,
Hang Lv, Yiwen Shao, Nanyun Peng, Lei Xie, Shinji Watan-
abe, and Sanjeev Khudanpur, “Espresso: A fast end-
to-end neural speech recognition toolkit,” arXiv preprint
 Neil Zeghidour, Nicolas Usunier, Gabriel Synnaeve, Ronan
Collobert, and Emmanuel Dupoux, “End-to-end speech recog-
nition from the raw waveform,” in Interspeech 2018, 2018.
 Jason Li, Vitaly Lavrukhin, Boris Ginsburg, Ryan Leary,
Oleksii Kuchaiev, Jonathan M Cohen, Huyen Nguyen, and
Ravi Teja Gadde, “Jasper: An end-to-end convolutional neural
acoustic model,” arXiv preprint arXiv:1904.03288, 2019.
 Dzmitry Bahdanau, Jan Chorowski, Dmitriy Serdyuk, Phile-
mon Brakel, and Yoshua Bengio, “End-to-end attention-based
large vocabulary speech recognition,” in 2016 IEEE interna-
tional conference on acoustics, speech and signal processing
(ICASSP). IEEE, 2016, pp. 4945–4949.
 Takaaki Hori, Shinji Watanabe, Yu Zhang, and William
Chan, “Advances in joint ctc-attention based end-to-end speech
recognition with a deep cnn encoder and rnn-lm,” arXiv
preprint arXiv:1706.02737, 2017.
 Linhao Dong, Shuang Xu, and Bo Xu, “Speech-transformer: a
no-recurrence sequence-to-sequence model for speech recog-
nition,” in 2018 IEEE International Conference on Acous-
tics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp.
 Julian Salazar, Katrin Kirchhoff, and Zhiheng Huang, “Self-
attention networks for connectionist temporal classiﬁcation in
speech recognition,” in ICASSP 2019-2019 IEEE Interna-
tional Conference on Acoustics, Speech and Signal Processing
(ICASSP). IEEE, 2019, pp. 7115–7119.
 Dimitri Palaz, Ronan Collobert, and Mathew Magimai Doss,
“End-to-end phoneme sequence recognition using convolu-
tional neural networks,” arXiv preprint arXiv:1312.2137,
uske, Pavel Golik, Ralf Schl¨
uter, and Hermann Ney,
“Acoustic modeling with deep neural networks using raw time
signal for lvcsr,” in Fifteenth annual conference of the interna-
tional speech communication association, 2014.
 Yedid Hoshen, Ron J Weiss, and Kevin W Wilson, “Speech
acoustic modeling from raw multichannel waveforms,” in 2015
IEEE International Conference on Acoustics, Speech and Sig-
nal Processing (ICASSP). IEEE, 2015, pp. 4624–4628.
 Mirco Ravanelli and Yoshua Bengio, “Speaker recogni-
tion from raw waveform with sincnet,” arXiv preprint
 Mirco Ravanelli and Yoshua Bengio, “Speech and speaker
recognition from raw waveform with sincnet,” arXiv preprint
 Shinji Watanabe, Takaaki Hori, Shigeki Karita, Tomoki
Hayashi, Jiro Nishitoba, Yuya Unno, Nelson Enrique Yalta So-
plin, Jahn Heymann, Matthew Wiesner, Nanxin Chen, et al.,
“Espnet: End-to-end speech processing toolkit,” arXiv preprint
 Lawrence R Rabiner and Ronald W Schafer, Theory and ap-
plications of digital speech processing, vol. 64, Pearson Upper
Saddle River, NJ, 2011.
 Sanjit Kumar Mitra and Yonghong Kuo, Digital signal pro-
cessing: a computer-based approach, vol. 2, McGraw-Hill
New York, 2006.
 Erfan Loweimi, Peter Bell, and Steve Renals, “On learning
interpretable cnns with parametric modulated kernel-based ﬁl-
ters,” Proc. Interspeech 2019, pp. 3480–3484, 2019.
 Mirco Ravanelli and Yoshua Bengio, “Interpretable convolu-
tional ﬁlters with sincnet,” arXiv preprint arXiv:1811.09725,
 Alex Graves, Santiago Fern´
andez, Faustino Gomez, and J¨
Schmidhuber, “Connectionist temporal classiﬁcation: la-
belling unsegmented sequence data with recurrent neural net-
works,” in Proceedings of the 23rd international conference
on Machine learning. ACM, 2006, pp. 369–376.
 Minh-Thang Luong, Hieu Pham, and Christopher D Manning,
“Effective approaches to attention-based neural machine trans-
lation,” arXiv preprint arXiv:1508.04025, 2015.
 Jan K Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk,
Kyunghyun Cho, and Yoshua Bengio, “Attention-based mod-
els for speech recognition,” in Advances in neural information
processing systems, 2015, pp. 577–585.
 John S Garofolo, Lori F Lamel, William M Fisher, Jonathan G
Fiscus, and David S Pallett, “Darpa timit acoustic-phonetic
continous speech corpus cd-rom. nist speech disc 1-1.1,” NASA
STI/Recon technical report n, vol. 93, 1993.
 Matthew D Zeiler, “Adadelta: an adaptive learning rate
method,” arXiv preprint arXiv:1212.5701, 2012.
 Andros Tjandra, Sakriani Sakti, and Satoshi Nakamura,
“Attention-based wav2text with feature transfer learning,” in
2017 IEEE Automatic Speech Recognition and Understanding
Workshop (ASRU). IEEE, 2017, pp. 309–315.
 Shigeki Karita, Nelson Enrique Yalta Soplin, Shinji Watanabe,
Marc Delcroix, Atsunori Ogawa, and Tomohiro Nakatani, “Im-
proving transformer-based end-to-end speech recognition with
connectionist temporal classiﬁcation and language model inte-
gration,” in INTERSPEECH 2019, 2019.