The third generation spiking neural networks are originally studied as models to describe the information processing in the biological neural networks, wherein the information is communicated and exchanged via stereotypical action potentials or spikes (Gerstner and Kistler, 2002). Neuroscience studies reveal that the temporal structure and frequency of these spike trains are both important information carriers in the biological neural networks. As will be introduced in section 3.1, the spiking neuron operates asynchronously and integrates the synaptic current from its incoming spike trains. An output spike is generated from the spiking neuron whenever its membrane potential crosses the firing threshold, and this output spike will be propagated to the connected neurons via the axon.

Motivated by the same connectionism principle, SNNs share the same network architectures, either feedforward or recurrent, with the conventional ANNs that use analog neurons. As shown in Figure 1, the early classification decision can be made from the SNN since the generation of the first output spike. However, the quality of the classification decision is typically improved over time with more evidence accumulated. It differs significantly from the synchronous information processing of the conventional ANNs, where the output layer needs to wait until all preceding layers are fully updated. Therefore, despite information is transmitted and processed at a speed that is several orders of magnitude slower in neural substrates than signal processing in modern transistors, biological neural systems can perform complex tasks rapidly. For more overviews about SNNs and their applications, we refer readers to Pfeiffer and Pfeil (2018) and Tavanaei et al. (2019).

As shown in Figure 2, conventional ASR systems uses acoustic and linguistic information preserved in three distinct components to convert speech signals to the corresponding text: (1) an acoustic model for preserving the statistical representations of different speech units, e.g., phones, from speech features, (2) a language model for assigning probabilities to the co-occurring word sequences and (3) a pronunciation lexicon for mapping the phonetic transcriptions to orthography. These resources are jointly used to determine the most likely hypothesis in the decoding stage.

Acoustic modeling can be achieved by using various statistical models such as Gaussian Mixture Models (GMM) for assigning frame-level phone posteriors in conjunction with a Hidden Markov Model (HMM) for duration modeling (Yu and Deng, 2015). More recently, ANN-based approaches have become the standard acoustic models providing state-of-the-art performance across a wide spectrum of ASR tasks (Hinton et al., 2012). Together with numerous ANN architectures explored for acoustic modeling, several end-to-end ANN architectures have been proposed for directly mapping speech features to text with optional use of the other linguistic components (Graves and Jaitly, 2014; Chan et al., 2016; Watanabe et al., 2017).

The probabilistic definition of acoustic modeling becomes more evident via the Bayesian formulation of the speech recognition task. Given a target speech signal that segmented into T overlapped frames, the resulting frame-wise features can be represented as O = [o₁, o₂, …, o_T]. An ASR system assigns the probability P(W|O) to all possible word sequences W = [w₁, w₂, …], and the word sequence W^ with the highest probability is the recognized output, (1)W^=arg max WP(W|O) The probability P(W|O) can be decomposed into two parts by applying the Bayes' rule as below, (2)W^=arg maxWP(O|W)P(W)P(O)

P(O) can be omitted as it does not depend on W. This results in (3)W^=arg max WP(O|W)P(W) which formally defines the theoretical foundation of the conventional ASR systems. P(W) is the prior probability of the word sequence W and this probability is provided by the language model which is trained on a large written corpus of the target language. P(O|W) is the likelihood of the observed feature sequence O given the word sequence W, and this probability is associated with the acoustic model. The acoustic model captures the information about the acoustic component of speech signals, aiming to classify different acoustic units accurately. Traditionally, each phone in the phonetic alphabet is modeled using multiple three-state HMM models for different preceding and following phonetic context (triphone) (Lee, 1990). The emission probability of these HMM states are shared (tied) among different models to reduce the number of model parameters (Hwang and Huang, 1993). The output layer of the ANN-based acoustic model is designed accordingly and trained to assign these frame-level tied triphone HMM state (senone) probabilities (Dahl et al., 2012). The output layer uses the softmax function to normalize the output into a probability distribution. These values are scaled with the prior probabilities of each class, obtained from the training data, to determine the likelihood values. These likelihood values are later combined with the probabilities assigned by the language model during the decoding stage so as to find the most likely hypothesis.

Speech features, used as the inputs to the acoustic model, describe the spectrotemporal dynamics of the speech signal and discriminate among different phones in the target language. Mel-frequency cepstral coefficients(MFCC) (Davis and Mermelstein, 1980) features are commonly used in conjunction with the GMM-HMM acoustic model. The MFCC features are extracted by (1) performing short-time Fourier transform, (2) applying triangular Mel-scaled filter banks to calculate the power at each Mel frequency in log domain (FBANK) and (3) performing a discrete cosine transform to decorrelate the FBANK features. The third step is often skipped and FBANK features are often used when training ANN-based acoustic models since these models can handle correlation among features. In this work, we incorporate deep SNNs for acoustic modeling instead of the conventional ANNs and compare their ASR performance in different ASR scenarios including phone recognition, low-resourced and standard large vocabulary ASR. The ASR performance obtained using popular speech features have been reported to explore the impact of the feature representation space and its dimensionality for SNN-based acoustic models.

SNNs are well-suited for representing and processing spatial-temporal signals, they hence possess great potentials for speech recognition tasks. Tavanaei and Maida (2017a,b) proposed SNN-based feature extractors to extract discriminative features from the raw speech signal using unsupervised spiking-timing-dependent plasticity (STDP) rule. While connecting these SNN-based feature extractors with Support Vector Machine (SVM) or Hidden Markov Model (HMM) classifiers, competitive classification accuracies were demonstrated on the isolated spoken digit recognition task. Wu et al. (2018a,b) introduced a SOM-SNN framework for environmental sound and speech recognition. In this framework, the biological-inspired self-organizing map (SOM) is utilized for feature representation, which maps frame-based acoustic features into a spike-based representation that is both sparse and discriminative. The temporal dynamic of the speech signal is further handled by the SNN classifier. Zhang et al. (2019) presented a fully SNN-based speech recognition framework, wherein the spectral information of consecutive frames are encoded with threshold coding and subsequently classified by the SNN that is trained with a novel membrane potential-driven aggregate-labeling learning algorithm.

Recurrent network of spiking neurons (RSNNs) exhibit greater memory capacity than the aforementioned feedforward frameworks. They can capture long temporal information that are useful for speech recognition tasks. In Zhang et al. (2015), Zhang et al. presented a spiking liquid-state machine (LSM) speech recognition framework which is attractive for low-power very-large-scale-integration (VLSI) implementation. Bellec et al. recently demonstrated state-of-the-art phone recognition accuracy on the TIMIT dataset by adding neuronal adaptation mechanism to the vanilla RSNNs (Bellec et al., 2018). It is the first time that RSNNs approaching the performance of LSTM networks (Greff et al., 2016) on the speech recognition task. These preliminary works on the SNN-based ASR systems are however limited to the phone classification or small vocabulary isolated spoken digit recognition tasks. In this work, we apply deep SNNs to LVCSR tasks and demonstrate competitive accuracies over the ANN-based ASR systems.

As shown in Figure 4, the frame-based features are first extracted and input into the SNN-based acoustic models. Given the short temporal duration of segmented frames and the slow variation of speech signals, these features are typically assumed to be stationary over the short time-period of segmented frames. In this work, we use the integrate-and-fire (IF) neuron model with reset by subtraction scheme (Rueckauer et al., 2017), which can effectively process these stationary frame-based features with minimal computational costs. Although IF neurons do not emulate rich temporal dynamics of biological neurons, they are however ideal for working with the neural representation that employed in this work, where spike timings play an insignificant role.

At each time step t of a discrete-time simulation, with a total number of time steps N_s, the incoming spikes to neuron j at layer l are transduced into synaptic current as follows (4)zjl(t)=∑iwjil-1·θil-1(t)+bjl where θil-1(t) indicates the occurrence of an input spike from afferent neuron i at time step t. In addition, the wjil-1 denotes the synaptic weight that connects presynaptic neuron i from layer l − 1. Here, bjl can be interpreted as a constant injecting current. As shown in Figure 3, neuron j integrates the input current zjl(t) into its membrane potential Vjl(t) as per Equation (5). The Vjl(0) is reset and initialized to zero for every new frame-based feature input. Without loss of generality, a unitary membrane resistance is assumed here. An output spike is generated whenever Vjl(t) crosses the firing threshold ϑ (Equation 6), which we set to a value of 1 for all the experiments by assuming that all synaptic weights are normalized with respect to the ϑ. (5)Vjl(t)=Vjl(t-1)+zjl(t)-ϑ·θjl(t-1) (6)θjl(t)=Θ(Vjl(t)-ϑ) with Θ(x)={1,if x≥00,otherwise According to Equations (4) and (5), the free aggregated membrane potential of neuron j (no firing) in layer l can be expressed as (7)Vjl,f=∑iwjil-1·cil-1+bjl·Ns where cil-1 is the input spike count from pre-synaptic neuron i at layer l − 1 as per Equation (8). (8)cil-1=∑t=1Nsθil-1(t). The Vjl,f summarizes the aggregate membrane potential contributions of the incoming spikes from pre-synaptic neurons while ignoring their temporal structures. As will be explained in the tandem learning framework section, this intermediate quantity links the SNN layers to the coupled ANN layers for parameter optimization.

SNNs process information transmitted via spike trains, therefore, special mechanisms are required to encode the continuous-valued feature vectors into spike trains and decode the classification results from the activity of output neurons. To this end, we adopt the spiking neural encoding scheme that proposed in Wu et al. (2019c). This encoding scheme first transforms frame-based input feature vector X⁰ (e.g., MFCC or FBANK features), where X0=[x10,x20,···,xn0]T, through a weighted layer of rectified linear unit (ReLU) neurons as follows (9)Vj0,f(0)≡aj0=ρ(∑iwji0·xi0+bj0) where wji0 is the strength of the synaptic connection between the input xi0 and ReLU neuron j. The bj0 is the corresponding bias term of the neuron j, and ρ(·) denotes the ReLU activation function. The free aggregate membrane potential Vj0,f(0) is defined to be equal to the activation value aj0 of the ReLU neuron j. We distribute this quantity over the encoding time window N_s and represent it via spike trains as per Equations (10) and (11). (10)θj0(t)=Θ(Vj0,f(t-1)-ϑ) (11)Vj0,f(t)=Vj0,f(t-1)-ϑ·θj0(t) Altogether, the spike train s⁰ and spike count c⁰ that output from the neural encoding layer can be represented as follows (12)s0={θ0(1),…,θ0(Ns)} (13)c0=∑t=1Nsθ0(t) This encoding layer performs weighted transformation inside an end-to-end learning framework. It transforms the original input representation to match the size of the encoding time window N_s and represents the transformed information via spike counts. This encoding scheme is beneficial for rapid inference since the input information can be effectively encoded within a short encoding window. Start from this neural encoding layer, as shown in Figure 4, we input the spike count c^l and s^l to subsequent ANN and SNN layers for tandem learning.

To ensure smooth learning with high precision error gradients derived at the output layer, we use the free aggregate membrane potential of output spiking neurons for neural decoding. Considering that the dimensionality of input feature vectors and output classes are much smaller than that of hidden layers, the computation required will be limited when deploying these two layers onto the edge devices.

Here, we present a recently proposed SNN learning rule, under the tandem neural network configuration, that exploits a connection between the activation value of ANN neurons and the spike count of IF neurons. As the input features are effectively encoded as spike counts, the temporal structure of the spike trains carries negligible information. The effective non-linear transformation of SNN layers therefore can be summarized as (14)cjl=f(sl-1;wjl-1,bjl) where f() denotes the transformation performed by spiking neurons. However, due to the state-dependent nature of spike generation, it is not viable to determine an analytical expression from s^l−1 to cjl directly. Therefore, we simplify the spike generation process by assuming the resulting synaptic currents from s^l−1 are evenly distributed over the encoding time window. As such, the interspike interval can be determined as follows (15)ISIjl=ρ(ϑVjl,f/Ns)=ρ(ϑ(∑iwjil-1cil-1+bjl·Ns)/Ns) Hence, the approximated “spike count” ajl can be derived according to (16)ajl=NsISIjl=1ϑ·ρ(∑iwjil-1cil-1+bjl·Ns) Given a unitary firing threshold ϑ, ajl can be effectively determined from an ANN layer of ReLU neurons by setting the spike count cil-1 as the input and the aggregated constant injecting current bjl·Ns as the bias term. This simplification of spike generation process allows the spike-train level error gradients to be approximated from the ANN layer. Wu et al. (2019c) have revealed that the cosine distances between the approximated ‘spike count’ a^l and the actual SNN output spike count c^l are exceedingly small in a high dimensional space, suggesting high quality error gradients can be approximated from the coupled ANN layers.

Based on this formulation, we constructed tandem neural networks as shown in Figure 4. During the activation forward propagation, the SNN layers are used to determine the exact spike representation which then propagate the aggregate spike counts and spike trains to the subsequent ANN and SNN layers, respectively. This interlaced layer structure ensures the information that forward propagated to the coupled ANN and SNN layers are synchronized. It worth noting that the ANN is just an auxiliary structure to facilitates the training of SNN, while only SNN is used during inference. The details of this tandem learning rule are provided in the Algorithm 1.

To train the deep SNN-based acoustic models, which is the main contribution of this work, several popular speech features have been extracted from the training recordings as described in section 2.2. Before being fed into the SNNs, these input speech features are contextualized by splicing multiple frames so as to exploit more temporal context information. Before training the SNN-based acoustic model, alignments of the speech features with the target senone labels are obtained using a conventional GMM-HMM-based ASR system similar to that described in Dahl et al. (2012). These frame-level alignments enable the training of the deep SNN acoustic model with the tandem learning approach. During the training, the deep SNN learns to map input speech features to posterior probabilities of senones (cf. section 2.2) by passing the input speech frames through multiple layers of spiking neurons.

During the inference phase, the acoustic scores provided by the trained SNN model are combined with the information stored in the language model and pronunciation lexicon. It is a common practice to use the weighted finite state transducers (WFST) (Mohri et al., 2002) as a unified representation of different ASR resources for creating the search graph containing possible hypotheses. The main motivation for using the WFST-based decoding is: (1) the straightforward composition of different ASR resources for constructing a mapping from HMM states to word sequences and (2) the existence of efficient search algorithms operating on WFST that speed up the decoding process. As a result of the search process, the most likely hypotheses are found and stored in the form of a lattice. The ASR output is chosen based on the weighted sum of the acoustic and language model scores belonging to hypotheses in the lattice. For further details of the WFST-based decoding approach used in this work, we refer the reader to Povey et al. (2012). In the following sections, we describe the ASR experiments conducted to evaluate the recognition performance of the proposed SNN-based acoustic modeling in several recognition scenarios.

We report the PER on the development and test sets of TIMIT corpus in Table 1, with numbers in bold being the best performance given by the speaker-independent features. ASR performances of other state-of-the-art systems using various ANN and SNN architectures are given in the upper panel for reference purposes. As the results shown in Table 1, the proposed SNN-based acoustic models are applicable to different speech features and provide comparable or slightly worse ASR performance than the ANNs with the same network structure. In particular, the ANN system trained with the standard 13-dimensional FBANK feature achieves the best PER of 16.9% (18.5%) on the development (test) set. The equivalent SNN system using the same feature achieves slightly worse PER of 17.3% (18.7%) on the development (test) set. Although the state-of-the-art ASR systems (Ravanelli et al., 2018) give approximately 1% lower PER than the proposed SNN-based phone recognition system, it is largely credit to the longer time context explored by the recurrent Li-GRU model.

It worth mentioning that phone recognition is still a challenging task for spiking neural networks. To the best of our knowledge, only one recent work with recurrent spiking neural networks (Bellec et al., 2019) demonstrates some promising test results on this corpus with a PER of 26.4%. In contrast, our system has achieved significantly lower PER compared to this preliminary study of SNN-based acoustic modeling. However, these results are not directly comparable since the proposed system incorporates both an acoustic and a language model during decoding unlike the system described in Bellec et al. (2019).

The experimental results on the TIMIT phone recognition task can be considered as an initial indicator of the compelling prospects of the SNN-based acoustic modeling. Given that the phone recognition task on TIMIT corpus is simplistic compared to the modern LVCSR tasks, we further compare the ANN and SNN performance on newer corpora designed for LVCSR experiments.

In this section, we apply the SNN-based ASR systems to the low-resourced ASR scenario. As summarized in Table 2, the word recognition results on the FAME corpus are reported separately for monolingual Frisian (fy), monolingual Dutch (nl) and code-switched (cs) utterances. The overall performance (all) is also included in the rightmost column. Given that 8.5 h Frisian and 3 h of Dutch speech is used during the training phase, we can compare the ASR performance on different subsets, i.e., fy, nl and cs, to identify the variations in the ASR performance for different levels of low-resourcedness. We omit the results on the development set as they follow a similar pattern to the results on the test set.

In this scenario, the SNN acoustic models consistently provide lower WERs than the ANN models for all speech features. Systems with the FBANK features provide lower WERs than those using MFCC features, which is in line with our observations on the TIMIT corpus. The best performance on the test set is obtained using SNN models trained on 40-dimensional FBANK features with an overall WER of 36.9%. In contrast, the ANN model provides a WER of 39.0% for the same setting, which is relatively 5.4% worse than the SNN model. Moreover, the SNN-based acoustic models achieve a relative improvement of 4.7%, 5.2% and 8.2% on the fy, nl and cs subsets of the test set, respectively. These steady improvements in the recognition accuracies highlight the effectiveness of the SNN-based acoustic modeling in scenarios with limited training data compared to the conventional ANN models. The improved ASR performance with SNNs, in the low-resourced setting, may credit to the noisy weight updates derived by the tandem learning framework. It has been recognized that introducing noises into the training stage improves the generalization capability of ANN-based ASR systems (Yin et al., 2015). As a result, the noisy training of the tandem learning is expected to improve the recognition performance in low-resourced scenarios. Further investigation on the impact of this noisy training procedure remains as future work.

In the final set of ASR experiments, we train acoustic models using the official 100 and 360-h training subsets of the Librispeech corpus to compare the recognition performance of ANN and SNN models in a standard LVCSR scenario. As the results given in the middle panel of Table 3, for 100 h of training data, the ANN systems perform marginally better than the corresponding SNN systems across all different speech features. The absolute WER differences range from 0.1% to 0.6%. These marginal performance degradations of the SNN models is likely due to the reduced representation power of using discrete spike counts. Nevertheless, these results are promising even when comparing to the state-of-the-art ASR systems using more complex ANN architectures as provided in the upper panel of Table 3.

It worth noting that both ANN and SNN systems can take benefit of an increased amount of training data. When increasing the training data from 100 to 360 h, the WERs of the best SNN models reduced from 10.0% (10.3%) to 9.2% (9.4%) for the development (test) sets, respectively. To the best of our knowledge, it is the first time that SNN-based acoustic models have achieved comparable results over the ANN models for LVCSR tasks. These results suggest that SNNs are potentially good candidates for acoustic modeling.

In addition to the promising modeling capability, the SNN-based ASR systems can achieve unprecedented performance gain when implemented on the low-power neuromorphic chips. In this section, we shed light on this prospect by comparing the energy efficiency of ANN- and SNN-based acoustic models. Given that data movements are the most energy-consuming operations for data-driven AI applications, we calculate the average synaptic operations on 5 randomly chosen utterances from the TIMIT corpus and report the ratio of average synaptic operations required per feature classification [SynOps(SNN)/SynOps(ANN)]. To investigate the effect of different feature representations, we repeat our analysis on the 40-dimensional MFCC, FBANK, and FMLLR features as summarized in Table 4 and Figure 5.

Taking advantage of the short encoding time window (N_s = 10), the sparse neuronal activities are observed for all network layers as shown in Figure 5. Among the three features explored in this experiment, it is interesting to note the FMLLR feature achieves the lowest average spike rate. It is likely due to the more discriminative nature of the speaker-dependent feature, while it worth to note that the FMLLR feature is not always available in all ASR scenarios. As provided in Table 4, the SNN implementations taking MFCC, FBANK and FMLLR input features require 1.72, 1.10, and 0.68 times synaptic operations to their ANN counterparts, respectively. Although the average number of synaptic operations required for SNNs that using MFCC and FBANK features are slightly higher than the ANNs, the AC operations performed on SNNs are much cheaper than the MAC operations required for ANNs. One recent study on the Global Foundry 28 nm process has revealed that MAC operations are 14 times more costly than AC operations and requires 21 times more chip area (Rueckauer et al., 2017). This study provides some good indicators for the potential energy and chips area savings that can be received from deploying SNNs onto the emerging neuromorphic chips for inference (Merolla et al., 2014; Davies et al., 2018). While the actual energy savings for SNN-based acoustic models are dependent on the chip architectures and materials used, which is beyond the scope of this work.

The phone and word recognition experiments on the well-known TIMIT and Librispeech benchmarks have demonstrated the promising modeling capacity of SNN acoustic models and their applicability to different input features. These preliminary results have shown that the recognition performance of SNNs is either comparable or slightly worse than the ANNs with the same network architecture on the TIMIT and Librispeech benchmarks. A possible reason for this performance degradation is the reduced representation power of the discrete neural representation (i.e., spike counts) as compared to the continuous floating-point representation of the ANNs (Wu et al., 2019c). This performance gap could potentially be closed by extending the encoding window N_s of SNNs. Moreover, the recognition performance of ANN and SNN models in a low-resourced scenario is also investigated. In this scenario, the SNN acoustic models outperform the conventional ANNs that could be attributed to the noisy training of the tandem learning framework, wherein error gradients of the SNN layers are approximated from the coupled ANN layers.

The neural encoding scheme adopted in this work allows input features to be encoded inside a short encoding time window for rapid processing by SNNs. It is attractive for the time-synchronous ASR tasks that require real-time performance. The preliminary study of the energy efficiency on the TIMIT corpus reveals attractive energy and chip area savings, as compared to the equivalent ANNs, can be achieved when deploying the offline trained SNNs onto neuromorphic chips. The recent study of a keyword spotting task on the Loihi neuromorphic research chip (Blouw et al., 2019) has also demonstrated the compelling energy savings, real-time performance and good scalability of emerging NC architectures over conventional low-power AI chips designed for ANNs.

The active development of open-source software toolkits plays a significant role in the rapid progress of ASR research, instances include the Kaldi (Povey et al., 2011) and ESPnet (Watanabe et al., 2018). In this work, we demonstrate that state-of-the-art SNN acoustic models can be easily developed in PyTorch and integrated into the PyTorch-Kaldi Speech Recognition Toolkit (Ravanelli et al., 2019). This software toolkit integrates the efficiency of Kaldi and the flexibility of PyTorch, therefore, it can support the rapid development of SNN-based ASR systems.

The recurrent neural networks have shown great modeling capability for temporal signals by exploring long temporal context information in the input signals (Graves and Jaitly, 2014). As future work, we will explore the recurrent networks of spiking neurons for large-vocabulary ASR tasks to further improve the recognition performance.

The substantial research efforts are devoted to reducing the computational cost and memory footprint of ANNs during inference, instances include network compression (Han et al., 2015), network quantization (Courbariaux et al., 2016; Zhou et al., 2016) and knowledge distillation (Hinton et al., 2015). While the computational paradigm underlying the efficient biological neural networks is fundamentally different from ANNs and hence fosters enormous potentials for neuromorphic computing architectures. Furthermore, grounded on the same connectionism principle, the information of both ANN and SNN are encoded in the network connectivity and connection strength. Therefore, SNN can also take benefits from these early research works on the network compression and quantization of ANNs to further reduce its memory footprint and computation cost (Deng et al., 2019).

The event-driven silicon cochlea audio sensors (Liu et al., 2014) are designed to mimic the functional mechanism of human cochlea and transform input audio signals into spiking events. Given temporally sparse information is transmitted in the surrounding environment, these sensors have shown greater coding efficiency than conventional microphone sensors (Liu et al., 2019). There are some interesting preliminary ASR studies explore the input spiking events captured by these silicon cochlea sensors (Acharya et al., 2018; Anumula et al., 2018). Additionally, Dominguez-Morales et al. (2018) have proposed a fully SNN-based framework for voice commands recognition, wherein the event-driven silicon cochlea audio sensor is directly interfaced with the SpiNNaker neuromorphic processor through the Address-Event Representation protocol (AER). Notably, a buffering layer is introduced to ensure real-time performance. However, the scale of the ASR tasks explored in these studies is relatively small comparing to modern ASR benchmarks due to the limited availability of event-based ASR corpora. Pan et al. (2020) recently proposed an efficient and perceptually motivated auditory neural encoding scheme to encode the large-scale ASR corpora collected by microphone sensors into spiking events. With this encoding scheme, approximately 50% spiking events can be reduced with negligible interference to the perceptual quality of inputs audio signals. Taking benefits from these earlier research on the neuromorphic auditory front-end, we are expecting to further improve the energy efficiency of SNN-based ASR systems.

The promising initial results demonstrated by the SNN-based large vocabulary ASR systems in this work is the first step toward a myriad opportunities for the integration of state-of-the-art ASR engines into mobile and embedded devices with power restrictions. In the long run, the SNN-based ASR systems are expected to take benefits from ever-growing research on novel neuromoprhic auditory front-end, SNN architectures, neuromorphic computing architectures and ultra-low-power non-volatile memory devices to further improve the computing performance.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.