Attentive Adversarial Learning for Domain-Invariant Training

TL;DR

This paper introduces an attentive adversarial training method that enhances domain-invariant acoustic features for speech recognition by focusing on more domain-sensitive components, leading to improved WER performance.

Contribution

The paper proposes an attention mechanism within ADIT to automatically weight features, improving domain-invariance and discriminability without altering the core DNN architecture.

Findings

Achieves 13.6% relative WER reduction over multi-conditional model

Achieves 9.3% relative WER reduction over baseline ADIT

Attention mechanism enhances focus on domain-sensitive phonetic features

Abstract

Adversarial domain-invariant training (ADIT) proves to be effective in suppressing the effects of domain variability in acoustic modeling and has led to improved performance in automatic speech recognition (ASR). In ADIT, an auxiliary domain classifier takes in equally-weighted deep features from a deep neural network (DNN) acoustic model and is trained to improve their domain-invariance by optimizing an adversarial loss function. In this work, we propose an attentive ADIT (AADIT) in which we advance the domain classifier with an attention mechanism to automatically weight the input deep features according to their importance in domain classification. With this attentive re-weighting, AADIT can focus on the domain normalization of phonetic components that are more susceptible to domain variability and generates deep features with improved domain-invariance and senone-discriminativity over ADIT. Most importantly, the attention block serves only as an external component to the DNN acoustic model and is not involved in ASR, so AADIT can be used to improve the acoustic modeling with any DNN architectures. More generally, the same methodology can improve any adversarial learning system with an auxiliary discriminator. Evaluated on CHiME-3 dataset, the AADIT achieves 13.6% and 9.3% relative WER improvements, respectively, over a multi-conditional model and a strong ADIT baseline.

Tables

Table 1. Table 1 : The ASR WER (%) of AADIT-S with addative attention of LSTM acoustic models for different sizes of attention window ( L + R + 1 𝐿 𝑅 1 L+R+1 ) on real development set of CHiME-3.

$L+R+1$	15	21	25	31
WER	17.89	17.63	17.89	18.07

Table 2. Table 2 : The ASR WER (%) of AADIT-S with addative attention of LSTM acoustic models for different dimensions of keys and values ( r a superscript 𝑟 𝑎 r^{a} ) on real development set of CHiME-3.

$r^a$	256	512	1024	2048
WER	18.53	17.63	18.01	18.19

Table 3. Table 3 : The ASR WERs (%) of multi-conditional (MC) LSTM acoustic models, ADIT, single-head AADIT, single-head AADIT with positional encoding (PE) and multi-head (MH) AADIT on real development set of CHiME-3. Both the additive (AD) and dot-product (DP) attentions are used for each AADIT system.

System	Attention Type	Domain
		Speaker	Environment
MC	-	19.23	19.23
ADIT	-	18.40	18.31
AADIT	AD	17.63	16.82
	DP	17.67	16.61
AADIT + PE	AD	17.57	16.68
	DP	17.37	16.94
MH AADIT	AD	17.33	17.10
	DP	17.25	16.97

Equations

$$M_y({\mathbf{f}}_t)=M_y(M_f({\mathbf{x}}_t))=p(s|{\mathbf{x}}_t;{\theta }_f,{\theta }_y).$$

$$M_d(M_f({\mathbf{x}}_t))$$

$${\mathcal{L}}_{\text{domain}}({\theta }_f,{\theta }_d)=-\frac{1}{T}\sum _{t=1}^T\log p(d_t|{\mathbf{f}}_t;{\theta }_d)$$

$$=-\frac{1}{T}\sum _{t=1}^T\sum _{u\in \text{mathbbm}U}\text{mathbbm}1[u=d_t]\log M_d(M_f({\mathbf{x}}_t)),$$

$${\mathcal{L}}_{\text{senone}}({\theta }_f,{\theta }_y)$$

$$=-\frac{1}{T}\sum _{t=1}^T\sum _{s\in \text{mathbbm}S}\text{mathbbm}1[s=y_t]\log M_y(M_f({\mathbf{x}}_t)).$$

$${\mathbf{k}}_t$$

$${\mathbf{q}}_t$$

$$a_{t,\tau }=\frac{\exp (e_{t,\tau })}{{\sum }_{{\tau }^{\prime }=t-L}^{t+R}\exp (e_{t,{\tau }^{\prime }})},$$

$$e_{t,\tau }=\frac{{\mathbf{k}}_{\tau }^{\top }{\mathbf{q}}_t}{\sqrt{r_a}},$$

$$e_{t,\tau }={\mathbf{g}}^{\top }\tanh ({\mathbf{k}}_{\tau }+{\mathbf{q}}_t+\mathbf{b}),$$

$${\mathbf{c}}_t=\sum _{\tau =t-L}^{t+R}{a}_{t,\tau }{\mathbf{f}}_{\tau }.$$

$$p(u|{\mathbf{x}}_t;{\theta }_f,{\theta }_a,{\theta }_d)$$

$$=M_d(M_a(M_f({\mathbf{x}}_t))).$$

$${\mathcal{L}}_{\text{domain}}({\theta }_f,{\theta }_a,{\theta }_d)=-\frac{1}{T}\sum _{t=1}^T\log p(d_t|{\mathbf{x}}_t;{\theta }_f,{\theta }_a,{\theta }_d)$$

$$=-\frac{1}{T}\sum _{t=1}^T\sum _{u\in \text{mathbbm}U}\text{mathbbm}1[u=d_t]\log M_d(M_a(M_f({\mathbf{x}}_t))).$$

$$({\hat{\theta }}_f,{\hat{\theta }}_y)$$

$$({\hat{\theta }}_a,{\hat{\theta }}_d)$$

$${\theta }_f\leftarrow {\theta }_f-\mu \left[\frac{\partial {\mathcal{L}}_{\text{senone}}}{\partial {\theta }_f}-\lambda \frac{\partial {\mathcal{L}}_{\text{domain}}}{\partial {\theta }_f}\right],$$

$${\theta }_a\leftarrow {\theta }_a-\mu \frac{\partial {\mathcal{L}}_{\text{domain}}}{\partial {\theta }_a},$$

$${\theta }_d\leftarrow {\theta }_d-\mu \frac{\partial {\mathcal{L}}_{\text{domain}}}{\partial {\theta }_d},$$

$${\theta }_y\leftarrow {\theta }_y-\mu \frac{\partial {\mathcal{L}}_{\text{senone}}}{\partial {\theta }_y},$$

Full text

Index Terms—  adversarial learning, attention, domain-invariant training, neural network, automatic speech recognition

1 Introduction

The deep neural network (DNN) based acoustic models have been widely used in automatic speech recognition (ASR) and have achieved extraordinary performance improvement [1, 2, 3]. However, the performance of a multi-conditional acoustic model trained with speech data from a variety of environments, speakers, microphone channels, etc. is still affected by the spectral variations in each speech unit caused by the inter-domain variability [4]. Recently, adversarial learning [5] has effectively improved the noise robustness of the DNN acoustic model for ASR [6, 7, 8] or that of the deep embeddings for speaker verification [9] by using gradient reversal layer network [10] or domain separation network [11]. Similar idea has also been applied to reduce the effect of inter-speaker [12, 13], inter-language [14] and inter-dialect [15] variability in acoustic modeling that is trained with speech from multiple speakers, multiple dialects or multiple languages. We name all the above approaches adversarial domain-invariant training (ADIT) by referring to each speaker, environment, language, etc. that contributes to the condition-variability of the training data a domain.

To perform ADIT, an additional DNN domain classifier is introduced and is jointly trained with the multi-conditional DNN acoustic model to simultaneously optimize the primary task of minimizing the senone classification loss and the secondary task of mini-maximizing the domain classification loss. Through this adversarial multi-task learning procedure, a shared feature extractor is learned as the bottom layers of the DNN acoustic model that maps the input speech frames from different domains into domain-invariant and senone-discriminative deep hidden features, so that further senone classification is based on representations with the domain factor already normalized out. After ADIT, only the DNN acoustic model is used to generate word transcription for test data from unseen domain through one-pass online decoding [12]. The domain classifier is not used in ASR. In this work, we focus on improving the domain classifier in ADIT to generate deep features with increased domain-invariance without changing the DNN acoustic model.

In ADIT, the sequence of deep features generated by the feature extractor are weighted equally before taken as the input to the domain classifier. In fact, the deep features corresponding to different phonetic components are affected nonuniformly by domain variability and show different domain-discriminativity to the domain classifier. To improve ADIT, we introduce an attention mechanism to allow the domain classifier to attend to different positions in time of deep feature sequence with nonuniform weights. The weights are automatically and dynamically determined by the attention mechanism according to the importance of the deep features in domain classification. We call this method attentive adversarial DIT (AADIT). With AADIT, the domain classifier induces attentive reversal gradients that emphasize on the domain normalization of more domain-discriminative deep features, improving the domain invariance of the acoustic model and thus the ASR performance over ADIT.

Self-attention [16] is a new technique to improve encoder-decoder end-to-end models [17, 18, 19]. It is used in [20] to enhance the performance of the generative adversarial network [5] by generating images based on cues from all feature locations. To the best of our knowledge, we introduce, for the first time, the attention mechanism only as an auxiliary component to the external of a DNN acoustic model to reduce the domain variability and improve ASR performance. Note that, similar to the domain classifier, the auxiliary attention block does not participate in the ASR decoding, so the proposed AADIT framework can be widely applied to acoustic model with any DNN architectures. AADIT can also improve the robustness in knowledge transferring for T/S learning [21, 22] as in [8]. More generally, the same methodology can be used to enhance the capability of the discriminators in any generative adversarial network [5] or gradient reversal layer network [10] for improved domain [10, 23] and speaker [24] adaptation, speech enhancement [25, 26, 27], speech synthesis [28, 29], voice conversion [30], speaker verification [9], image generation [5, 31] and translation [32], etc.

We perform AADIT to suppress speaker and environment variabilities of the DNN acoustic model and thus to improve ASR. We explore two types of local attention mechanisms: additive attention and dot-product attention for AADIT and investigate the effect of attention window size, key/query dimension, positional encoding and multi-head attention on the ASR performance. Evaluated on CHiME-3 dataset, AADIT of DNN acoustic model achieves 13.6% and 9.3% relative word error rate (WER) improvements over the multi-conditional model and ADIT, respectively.

2 Adversarial Domain-Invariant Training

ADIT aims at reducing the variances of hidden and output unit distributions of the DNN acoustic model that are caused by the inherent inter-domain variability in the speech signal. To achieve domain-robustness, one solution is to learn a domain-invariant and senone-discriminative deep hidden feature in the DNN acoustic model through adversarial multi-task learning and make senone posterior predictions based on the learned deep feature. In order to do so, we need a sequence of speech frames $\mathbf{X}=\{\mathbf{x}_{1},\ldots,\mathbf{x}_{T}\},\mathbf{x}_{t}\in\mathbbm{R}^{r_{x}},t=1,\ldots,T$, a sequence of senone labels $\mathbf{Y}=\{y_{1},\ldots,y_{T}\}$, $y_{t}\in\mathbbm{R}$ aligned with $\mathbf{X}$ and a sequence of domain labels $\mathbf{D}=\{d_{1},\ldots,d_{T}\},d_{t}\in\mathbbm{R}$ aligned with $\mathbf{X}$. We view the first few layers of the acoustic model as a feature extractor network $M_{f}$ with parameters $\theta_{f}$ that maps input speech frames $X$ from different domains to intermediate deep hidden features $\mathbf{F}=\{\mathbf{f}_{1},\ldots,\mathbf{f}_{T}\},\mathbf{f}_{t}\in\mathbbm{R}^{r_{f}}$ and the upper layers of the acoustic model as a senone classifier $M_{y}$ with parameters $\theta_{y}$ that maps the deep features $\mathbf{F}$ to the senone posteriors $p(s|\mathbf{f}_{t};\theta_{y}),s\in\mathbbm{S}$ as follows:

[TABLE]

We further introduce a domain classifier network $M_{d}$ which maps the deep features $\mathbf{F}$ to the domain posteriors $p(u|\mathbf{f}_{t};\mathbf{\theta}_{d})$, $u\in\mathbbm{U}$ as follows:

[TABLE]

where $u$ is one domain in the set of all domains $\mathbbm{U}$. To make the deep features $\mathbf{F}$ domain-invariant, the distributions of $\mathbf{F}$ from different domains should be as close to each other as possible. Therefore, we jointly train $M_{f}$ and $M_{d}$ with an adversarial objective, in which $\mathbf{\theta}_{f}$ is adjusted to maximize the frame-level domain classification loss $\mathcal{L}_{\text{domain}}$ while $\mathbf{\theta}_{d}$ is adjusted to minimize $\mathcal{L}_{\text{domain}}$ below:

[TABLE]

where $\mathbbm{1}[\cdot]$ is the indicator function which equals to 1 if the condition in the squared bracket is satisfied and 0 otherwise. This minimax competition will first increase the discriminativity of $M_{d}$ and the domain-invariance of the deep features generated by $M_{f}$, and will eventually converge to the point where $M_{f}$ generates extremely confusing deep features that $M_{d}$ is unable to distinguish.

At the same time, we want to make $\mathbf{F}$ senone-discriminative by minimizing the cross-entropy senone classification loss between the predicted senone posteriors and the senone labels below:

[TABLE]

In ADIT, the acoustic model network and the condition classifier network are trained to jointly optimize the primary task of senone classification and the secondary task of domain classification with an adversarial objective function.

3 Attentive Adversarial Domain-Invariant Training

In ADIT, the mini-maximization of the domain classification loss (Eq. (3)) plays an important role in normalizing the intermediate deep feature $\mathbf{F}$ against different domains. However, the domain classification loss is still computed from a sequence of equally-weighted deep features. In fact, not all deep features are equally affected by domain variability and provide identical domain-discriminative information to the domain classifier. For example, deep features extracted from voiced frames are more domain-discriminative than those from the silence; deep features aligned with vowels are in general more susceptible to domain variability than those with consonants. To address this problem, we introduce an attention machanism to dynamically and automatically adjust the weights for the deep features in order to put more emphasis on the domain normalization of more domain-discriminiative deep features and therefore enhance the overall domain-invariance of the deep features. The acoustic model with such a domain-invariant deep feature is expected to achieve improved ASR performance over ADIT.

In the proposed AADIT, we use soft local (time-restricted) [33, 34] self-attention [16] because it is more suitable for ASR where the input sequence consists of a relatively large number of speech frames. The local attention we adopt selectively focuses on a small window of context centered at the current time and can jointly attend different points in time with different weights. Specifically, for each deep feature $\mathbf{f}_{t}$ at time $t$, the keys are the projections of deep features in an $r_{a}$ dimensional space within the attention window of size $L+R+1$, i.e., $\mathbf{K}_{t}=\{\mathbf{k}_{t-L},\ldots,\mathbf{k}_{t},\ldots,\mathbf{k}_{t+R}\}$, and the query $\mathbf{q}_{t}$ is the projection of $\mathbf{f}_{t}$ in the $r_{a}$ dimensional space, i.e.,

[TABLE]

where $\mathbf{W}_{k}$ is a $r_{a}$ by $r_{f}$ key projection matrix, $\mathbf{W}_{q}$ is a $r_{a}$ by $r_{f}$ query projection matrix and $L$ and $R$ are the length of left and right context, respectively in the attention window. The attention probability $\mathbf{a}_{t}$ of each current frame $\mathbf{f}_{t}$ against all the context deep features in the attention window, i.e., $\mathbf{V}_{t}=\{\mathbf{f}_{t-L},\ldots,\mathbf{f}_{t},\ldots,\mathbf{f}_{t+R}\}$, is computed by normalizing the similarity scores $e_{t,\tau}\in\mathbbm{R}$ between the query $\mathbf{q}_{t}$ and each key $\mathbf{k}_{\tau}$ in the window $\mathbf{K}_{t}$, i.e.,

[TABLE]

where $\tau=t-L,\ldots,t,\ldots,t+R$ and $a_{t,\tau}\in\mathbbm{R}$ is the $[\tau-(t-L)]^{\text{th}}$ dimension of the attention probability vector $\mathbf{a}_{t}\in\mathbbm{R}^{L+R+1}$. The similarity scores $e_{t,\tau}$ can be computed in two different ways according to the type of attention mechanism applied:

• •

Dot-product attention

[TABLE]

• •

Additive attention

[TABLE]

where $\mathbf{g}\in\mathbbm{R}^{r_{a}}$ is a column vector, $\mathbf{b}\in\mathbbm{R}^{r_{a}}$ is the bias column vector.

Therefore, a context vector $\mathbf{c}_{t}$ is formed at each time $t$ as a weighted sum of the deep features in the attention window $\mathbf{V}_{t}$ with the attention vector $\mathbf{a}_{t}$ serving as the combination weights, i.e.,

[TABLE]

As shown in Fig. 1, we view the entire attention process described in Eq.(5) to Eq.(10) as a single attention function $M_{a}(\cdot)$ with parameters $\mathbf{\theta}_{a}=\{\mathbf{W}_{k},\mathbf{W}_{q},\mathbf{g},\mathbf{b}\}$ which takes in the sequence of deep features $\mathbf{F}$ as the input and outputs the sequence of context vectors $\mathbf{C}=\{\mathbf{c}_{1},\ldots,\mathbf{c}_{T}\},\mathbf{c}_{t}\in\mathbbm{R}^{r_{a}}$, i.e., $\mathbf{c}_{t}=M_{a}(\mathbf{f}_{t})$.

The domain classifier $M_{d}$ then takes in the context vector $\mathbf{c}_{t}$ as the input to predict the frame-level domain posteriors for $u\in\mathbbm{U}$.

[TABLE]

To make the deep features $\mathbf{F}$ domain-invariant, $M_{f}$, $M_{a}$ and $M_{d}$ are jointly trained with an adversarial objective, in which $\mathbf{\theta}_{f}$ is adjusted to maximize the frame-level domain classification loss $\mathcal{L}_{\text{domain}}(\mathbf{\theta}_{f},\mathbf{\theta}_{a},\mathbf{\theta}_{d})$ while $\mathbf{\theta}_{a}$ and $\mathbf{\theta}_{d}$ are adjusted to minimize $\mathcal{L}_{\text{domain}}(\mathbf{\theta}_{f},\mathbf{\theta}_{a},\mathbf{\theta}_{d})$ below:

[TABLE]

In AADIT, the acoustic model network, the condition classifier network and attention function are trained to jointly optimize the primary task of senone classification and the secondary task of domain classification with an adversarial objective function as follows

[TABLE]

where $\lambda$ controls the trade-off between $\mathcal{L}_{\text{senone}}$ and $\mathcal{L}_{\text{domain}}$, and $\hat{\mathbf{\theta}}_{y},\hat{\mathbf{\theta}}_{f},\hat{\mathbf{\theta}}_{a}$ and $\hat{\mathbf{\theta}}_{d}$ are the optimized parameters.

The parameters are updated as follows via back propagation with stochastic gradient descent:

[TABLE]

where $\mu$ is the learning rate. Note that the negative coefficient $-\lambda$ in Eq. (15) induces attentive reversal gradient that maximizes $\mathcal{L}_{\text{domain}}$ in Eq. (3) to make the deep feature domain-invariant. For easy implementation, a gradient reversal layer is introduced as in [10], which acts as an identity transform in the forward propagation and multiplies the gradient by $-\lambda$ during the backward propagation.

Note that only the optimized DNN acoustic model consisting of $M_{f}$ and $M_{y}$ on the left side of Fig. 1 is used for ASR on test data. The attention block $M_{a}$ and domain classifer $M_{d}$ (on the right) are discarded after AADIT.

We further extend the keys, queries and values with a one-hot encoding of the relative positions versus the current time in an attention window as in [34] and compute the attention vectors based on the extended representations. We also introduce a multi-head attention as in [16] by projecting the deep features $H$ times to get $H$ keys and queries in $H$ different spaces. Note that the dimension of projection space for each attention head is one $H^{\text{th}}$ of that in single-head attention to keep the number of parameters unchanged.

4 Experiments

We perform AADIT of a multi-conditional acoustic model to suppress the speaker variability (AADIT-S) and environment variability (AADIT-E) for robust ASR.

4.1 Baseline System

As the baseline system, we first train a long short-term memory (LSTM)- hidden Markov model (HMM) acoustic model [35, 36, 37] using multi-conditional training data of CHiME-3 [38]. The CHiME-3 dataset incorporates Wall Street Journal (WSJ) corpus sentences spoken under four challenging noisy environments, i.e, on buses, in cafes, in pedestrian areas, at street junctions and one clean environment, i.e., in booth, recorded using a 6-channel tablet. The real far-field noisy speech from the 5th microphone channel in CHiME-3 development data set is used for testing. A standard WSJ 5K word 3-gram language model is used for decoding.

We train the baseline LSTM acoustic model with 9137 clean and 9137 noisy training utterances of CHiME-3 dataset by using cross-entropy criterion. The 29-dimensional log Mel filterbank features together with 1st and 2nd order delta features (totally 87-dimensional) for both the clean and noisy utterances are extracted as in [39]. The features are fed as the input to the LSTM after global mean and variance normalization. The LSTM has four 1024-dimensional hidden layers. Each hidden layer is followed by a 512-dimensional projection layer. The output layer of the LSTM has 3012 output units corresponding to 3012 senone labels. The multi-style LSTM acoustic model achieves 19.23% WER on the noisy test data.111Note that our experimental setup does not achieve the state-of-the-art performance on CHiME-3 (e.g., we did not perform beamforming, sequence training or use recurrent neural network language model for decoding.) since our goal is to simply verify the improved capability of AADIT in reducing inter-domain variability over ADIT.

4.2 Adversarial Domain-Invariant Training

We further perform ADIT to reduce the speaker variability (ADIT-S) and environment variability (ADIT-E) of the baseline multi-conditional LSTM with 9137 noisy training utterances in CHiME-3. The $M_{f}$ is initialized with the first $P$ layers of the LSTM and $M_{y}$ is initialized with the rest $(7-P)$ hidden layers plus the output layer. $P$ indicates the position of the deep hidden feature in the acoustic model. For ADIT-S, the speaker classifier $M_{d}$ is a feedforward DNN with 3 hidden layers and 512 hidden units for each layer. The output layer of $M_{d}$ has 87 units predicting the posteriors of 87 speakers in the training set. For ADIT-E, $M_{d}$ is an environment classifier with the same architecture as in ADIT-S except for the 5 output units predicting 5 environments in CHiME-3. $M_{f}$, $M_{y}$ and $M_{d}$ are jointly trained with an adversarial multi-task objective as described in Section 2. $P$ and $\lambda$ are fixed at $4$ and $0.5$ in our experiments. The ADIT-S and ADIT-E LSTM acoustic models achieve 18.40% and 18.31% WER on the real test data, respectively, which are 4.3% and 4.8% relative improvements over the multi-conditional baseline.

4.3 Attentive Adversarial Domain-Invariant Training

We further perform AADIT-S and AADIT-E with the same training data as ADIT. $M_{f}$, $M_{y}$ and $M_{d}$ have exactly the same architectures as the ones for ADIT in Section 4.2 expect that $M_{d}$ has only one hidden layer to keep the number of parameters similar as that of ADIT. $P$ and $\lambda$ are also fixed at $4$ and $0.5$ for all the experiments.

We first investigate the impact of attention window size $L+R+1$ on the ASR WER via AADIT-S with additive attention in Table 1. In this work, we only use symmetric attention window with $L=R$. We fix the key and query dimension $r_{a}$ at $512$. The WER begins to decrease when window size is larger than 21, so we choose $L=R=10$ for the following experiments. Then we explore the effect of different key and query dimensions $r^{a}$ on the ASR performance through AADIT-S with addative attention in Table 2. The WER first decreases until reaching the minimum at $r^{a}=512$ and then increases as $r^{a}$ grows larger. Therefore, we fix $r^{a}$ at $512$ for the following experiments. Note that, by setting $r^{a}=512$, the number of learnable parameters are kept roughly the same as in ADIT.

Further, we perform AADIT-S and AADIT-E with both additive and dot-product attentions and summarize the WER results for different type of domains in Table 3. We see that AADIT-S achieves 17.63% WER with additive attention which is 8.3% and 4.2% relatively improved over baseline multi-conditional LSTM and ADIT-S, respectively. AADIT-E performs significantly better than AADIT-S with a WER of 16.61% when using dot-product attention, which is 13.6% and 9.3% relatively improved over baseline multi-conditional model and ADIT-E, respectively. Dot-product attention performs similar to additive attention for AADIT.

We also investigate the effect of positional encoding on AADIT. In Table 3, positional encoding does not consistently improve the AADIT, so we do not use it for the following experiments. We further perform AADIT with multi-head additive and dot-product attentions. The number of heads is fixed at $8$ and the key/query dimension for each head is $512/8=64$. We observe that the multi-head AADIT-S only slightly improves the WER of single-head one, and multi-head AADIT-E does not further improve the WER. Considering the significantly better WER with less computational cost, we suggest using single-head AADIT-E for robust ASR.

5 Conclusions

We advance the domain classifier of ADIT with an attention mechanism to re-weight the deep features in a DNN acoustic model according to their importance in domain classification. With AADIT, the deep features more susceptible to domain variability are normalized with more emphasis and therefore, the overall domain-invariance of the acoustic model is greatly enhanced. The attention mechanism only serves as an auxiliary component to the external of the acoustic model that does not participate in ASR and thus can improve the DNN acoustic model with any architectures.

Evaluated on CHiME-3 dataset, the single-head AADIT achieves 13.6% and 9.3% relative WER gains over a multi-conditional LSTM acoustic model and a strong ADIT baseline, respectively. AADIT-E performs significantly better than AADIT-S. The additive and dot-product attentions achieve similar ASR performance. WERs of AADIT do not improve significantly with additional positional encoding and multi-head self-attention.