Interpretable Deep Learning Model for the Detection and Reconstruction of Dysarthric Speech

TL;DR

This paper introduces an interpretable deep learning model that detects and reconstructs dysarthric speech by leveraging a low-dimensional latent space, improving detection accuracy and speech fluency.

Contribution

It presents a novel encoder-decoder architecture that factorizes speech into an interpretable latent space, enhancing dysarthria detection and speech reconstruction capabilities.

Findings

Latent space conveys interpretable dysarthria characteristics.

Model generates speech with improved fluency.

Higher detection accuracy for dysarthria.

Abstract

This paper proposed a novel approach for the detection and reconstruction of dysarthric speech. The encoder-decoder model factorizes speech into a low-dimensional latent space and encoding of the input text. We showed that the latent space conveys interpretable characteristics of dysarthria, such as intelligibility and fluency of speech. MUSHRA perceptual test demonstrated that the adaptation of the latent space let the model generate speech of improved fluency. The multi-task supervised approach for predicting both the probability of dysarthric speech and the mel-spectrogram helps improve the detection of dysarthria with higher accuracy. This is thanks to a low-dimensional latent space of the auto-encoder as opposed to directly predicting dysarthria from a highly dimensional mel-spectrogram.

Tables

Table 1. Table 1: Configuration of the neural network blocks.

Neural block	Config
Audio encoder
2x CNN	20 channels, 5x5 kernel, RELU, VALID
GRU	20 hidden states, 1 layer
Dense	20 units, tanh
Dysarthric space	2 units, linear
Text encoder
3x CNN	40 channels, 5x5 kernel, RELU, SAME
GRU	27 hidden states, 1 layer
Audio decoder
Dense bottleneck	96 units, RELU
GRU query	29 hidden states, 1 layer
GRU decoder	128 hidden states, 1 layer
Linear projection	frames_num x melsp bins units, linear

Table 2. Table 2: Accuracy of dysarthria detection including 95% CI. Classifier task - target mel-spectrogram (ML) is not observed during training. Multitask - both targets ML and dysarthric labels are observed

System	Accuracy	Precision	Recall
Word level
Multitask	0.853 (0.849 - 0.857)	0.831	0.911
Classifier task	0.820 (0.815 - 0.824)	0.818	0.855
Gillespie et al.[11]	0.753 (na)	0.823	0.728
Speaker level
Multitask	0.929 (0.790-0.984)	1.000	0.867
Classifier task	0.929 (0.790-0.984)	0.933	0.933
Gillespie et al.[11]	0.929 (na)	na	na

Equations

$$\sum _{i=1}^m\alpha log(p(d_i|X_i,\theta ))+(1-\alpha )log(p(Y_i|X_i,T_i,\theta ))$$

Full text

We present a novel deep learning model for the detection and reconstruction of dysarthric speech. We train the model with a multi-task learning technique to jointly solve dysarthria detection and speech reconstruction tasks. The model key feature is a low-dimensional latent space that is meant to encode the properties of dysarthric speech. It is commonly believed that neural networks are “black boxes” that solve problems but do not provide interpretable outputs. On the contrary, we show that this latent space successfully encodes interpretable characteristics of dysarthria, is effective at detecting dysarthria, and that manipulation of the latent space allows the model to reconstruct healthy speech from dysarthric speech. This work can help patients and speech pathologists to improve their understanding of the condition, lead to more accurate diagnoses and aid in reconstructing healthy speech for afflicted patients.

Index Terms: dysarthria detection, speech recognition, speech synthesis, interpretable deep learning models

1 Introduction

Dysarthria is a motor speech disorder manifesting itself by a weakness of muscles controlled by the brain and nervous system that are used in the process of speech production, such as lips, jaw and throat [1]. Patients with dysarthria produce harsh and breathy speech with abnormal prosodic patterns, such as very low speech rate or flat intonation, which makes their speech unnatural and difficult to comprehend. Damage to the nervous system is the main cause of dysarthria [1]. It can happen as an effect of multiple possible neurological disorders such as cerebral palsy, brain stroke, dementia or brain cyst [2, 3].

Early onset detection of dysarthria may improve the quality of life for people affected by these neurological disorders. According to Alzheimer’s Research UK2015 [4], 1 out of 3 people in the UK born in 2015 will develop dementia in their life. Manual detection of dysarthria conducted in clinical conditions by speech pathologists is costly, time-consuming and can lead to an incorrect diagnosis [5, 6]. With an automated analysis of speech, we can detect an early onset of dysarthria and recommend further health checks with a clinician even when a human speech pathologist is not available. Speech reconstruction may help with better identification of the symptoms and enable patients with severe dysarthria to communicate with other people.

Section 2 presents related work. In Section 3 we describe the proposed model for detection and reconstruction of dysarthria. In Section 4 we demonstrate the performance of the model with experiments on detection, interpretability, and reconstruction of healthy speech from dysarthric speech. We conclude with our remarks.

2 Related work

2.1 Dysarthria detection

Deep neural networks can automatically detect dysarthric patterns without any prior expert knowledge [7, 8]. Unfortunately, these models are difficult to interpret because they are usually composed of multiple layers producing multidimensional outputs with an arbitrary meaning and representation. Contrarily, statistical models based on a fixed vector of handcrafted prosodic and spectral features such as jitter, shimmer, Noise to Harmonic Ratio (NHR) or Mel-Frequency Cepstral Coefficients (MFCC) offer good interpretability but require experts to manually design predictor features [9, 10, 11, 12].

The work of Tu Ming et al. on interpretable objective evaluation of dysarthria [13] is the closest we found to our proposal. The main difference is that our model not only provides interpretable characteristics of dysarthria but also reconstructs healthy speech. Their model is based on feed-forward deep neural networks with a latent layer representing four dimensions of dysarthria: nasality, vocal quality, articulatory precision, and prosody. The final output of the network represents general dysarthria severity on a scale from 1 to 7. The input to this model is described by a 1201-dimensional vector of spectral and cepstral features that capture various aspects of dysarthric speech such as rhythm, glottal movement or formants. As opposed to this work, we use only mel-spectrograms to present the input speech to the model. Similarly to our approach, Vasquez-Correa et al. [8] uses a mel-spectrogram representation for dysarthria detection. However, they use 160 ms long time windows at the transition points between voiced and unvoiced speech segments, in contrast to using a full mel-spectrogram in our approach.

2.2 Speech reconstruction

There are three different approaches to the reconstruction of dysarthric speech: voice banking, voice adaptation and voice reconstruction [5]. Voice banking is a simple idea of collecting a patient’s speech samples before their speech becomes unintelligible and using it to build a personalized Text-To-Speech (TTS) voice. It requires about 1800 utterances for a basic unit-selection TTS technology [14] and more than 5K utterances for building a Neural TTS voice [15]. Voice adaptation requires as little as 7 minutes of recordings. In this approach, we start with a TTS model of an average speaker and adapt its acoustic and articulatory parameters to the target speaker [16].

Both voice banking and voice adaptation techniques rely on the availability of recordings for a healthy speaker. The voice reconstruction technique overcomes this shortcoming. This technique aims at restoring damaged speech by tuning parameters representing the glottal source and the vocal tract filter [17, 18]. In our model, we take a similar approach. However, instead of making assumptions on what parameters should be restored, we let the model automatically learn the best dimensions of the latent space that are responsible for dysarthric speech. Reconstruction of healthy speech by manipulating the latent space of a dysarthric speech is a promising direction, however, so far we only managed to successfully apply this technique in a single-speaker setup.

Variational Auto-Encoder (VAE) [19] is a probabilistic latent space model that has recently become popular for the reconstruction of various signals such as text [20, 21] and speech [22, 23].

3 Proposed model

The model consists of two output networks, jointly trained, with a shared encoder as shown in Figure 1. The audio and text encoders produce a low-dimensional dysarthric latent space and a sequential encoding of the input text. The audio decoder reconstructs input mel-spectrogram from a dysarthric latent space and encoded text. Logistic classification model predicts the probability of dysarthric speech from the dysarthric latent space. In Table 1 we present the details of various neural blocks used in the model.

Let us define a matrix $X:[n_{mels},n_{f}]$ representing a mel-spectrogram (frame length=50ms and frame shift=12.5ms), where $n_{mels}=128$ is the number of mel-frequency bands and $n_{f}$ is the number of frames. Let us define a matrix $T:[n_{c},n_{t}]$ representing a one-hot encoded input text, where $n_{c}$ is the number of unique characters in the alphabet and $n_{t}$ is the number of characters in the input text. The mel-spectrogram $X$ is encoded into 2-dimensional dysarthria latent space $\mathbf{l}=\{l_{1},l_{2}\}$ and then used as a conditioning variable for estimating the probability of dysarthria $d\backsim p(d|X,\theta)$ and reconstructing the mel-spectrogram $Y\backsim p(Y|X,T,\theta)$. Limiting the latent space to 2 dimensions makes the model more resilient to overfitting. The theta is a vector of trainable parameters of the model.

Let us define a training set of $m$ tuples of $((X,T),y)$, where $y\in\{0,1\}$ is the label for normal/dysarthric speech and $m$ is the number of speech mel-spectrograms for dysarthric and normal speakers. We optimize a joint cost of the predicted probability of dysarthria and mel-spectrogram reconstruction defined as a weighted function:

[TABLE]

where $log(p(d_{i}|X_{i},\theta))$ is the cross-entropy between the predicted and actual labels of dysarthria, and $log(p(Y_{i}|X_{i},T_{i},\theta))$ is the log-likelihood of a Gaussian distribution for the predicted mel-spectrogram with a unit variance, a.k.a L2 loss. We used backpropagation and mini-batch stochastic gradient descent with a learning rate of 0.03 and a batch size of 50. The whole model is initialized with Xavier’s method [24] using the magnitude value of 2.24. Hyper-parameters of the model presented in Table 1 were tuned with a grid search optimization. We used MxNet framework for implementing the model [25].

3.1 Mel-spectrogram and text encoders

For the spectrogram encoder, we use a Recurrent Convolutional Neural Network model (RCNN) [26]. The convolutional layers, each followed by a max-pooling layer, extract local and time-invariant patterns of the glottal source and the vocal tract. The GRU layer models temporal patterns of dysarthric speech [27]. The last state of the GRU layer is processed by two dense layers. Dropout [28] with probability of 0.5 is applied to the output of the activations for both CNN layers, GRU layer, and the dense layer.

Text encoder encodes the input text using one-hot encoding, followed by three CNN layers and one GRU layer. Outputs of both audio and text encoders are concatenated via matrix broadcasting, producing a matrix $E:[n_{c}+n_{l},n_{t}]$, where $n_{l}$ is dimensionality of the dysarthria latent space.

3.2 Spectrogram decoder and dysarthria detector

For decoding a mel-spectrogram, similarly to Wang et al. [29], we use a Recurrent Neural Network (RNN) model with attention. The dot-product attention mechanism [30] plays a crucial role. It informs to which elements of the encoder output the decoder should pay attention at every decoder step. The RNN network that produces a query vector for the attention, takes as input $r$ predicted mel-spectrogram frames from the previous time-step. The output of the RNN decoder is projected via a linear dense layer into $r$ number of mel-spectrogram frames. Similarly to Wang et al. [29], we found that it is important to preprocess the mel-spectrogram with a dense layer and dropout regularization to improve the overall generalization of the model.

The dysarthria detector is created from a 2-dimensional dense layer. It uses a tanh activation followed by a softmax function that represents the probability of dysarthric speech.

4 Experiments

4.1 Dysarthric speech database

There is no well-established benchmark in the literature to compare different models for detecting dysarthria. Aside from the most popular dysarthric corpora, UA-Speech [31] and TORGO [32], there are multiple speech databases created for the purpose of a specific study, for example, corpora of 57 dysarthric speakers [12] and Enderby Frenchay Assessment dataset [6]. Many corpora, including TORGO and HomeService [33], are available under non-commercial license.

In our experiments we use the UA-Speech database from the University of Illinois [31]. It contains 11 male and 4 female dysarthric speakers of different dysarthria severity levels and 13 control speakers. 455 isolated words are recorded for each speaker with 1 to 3 repetitions. Every word is recorded through a 7-channel microphone array, producing a separate wav file of 16 kHz sampling rate for every channel. It contains 9.4 hours of speech for dysarthric speakers and 4.85 hours for control speakers. UA-Speech corpus comes with intelligibility scores that are obtained from a transcription task performed by 5 naive listeners.

To control variabilities in recording conditions, we normalized mel-spectrograms for every recorded word independently with a z-score normalization. We considered removing the initial period of silence at the beginning of recorded words but we decided against it. We found that for dysarthric speakers of high speech intelligibility, the average length of the initial silence period that lasts 0.569sec +- 0.04674 (99% CI) is comparable with healthy speakers with the length of 0.532sec +- 0.055. Because we can predict unvoiced periods with merely 85% of accuracy [34], removing the periods of silence for dysarthric speakers with poor intelligibility is very inaccurate.

4.2 Automatic detection of dysarthria

To define the training and test sets, we use a Leave-One-Subject-Out (LOSO) cross-validation scheme. For each training, we include all speakers but one that is left out to measure the prediction accuracy on unseen examples. The accuracy, precision and recall metrics are computed at a speaker level (the average dysarthria probability of all the words produced by the speaker is compared to a target speaker dysarthria label $\in\{0,1\}$), and a word level (comparing target dysarthria label with predicted dysarthria probability for all words independently).

As a baseline, we use the Gillespie’s et al. model that is based on Support Vector Machine classifier [11]. It uses 1595 low-level predictor features processed with a global z-score normalization. It reports a 75.3 and 92.9 accuracy in the dysarthria detection task at the word and speaker levels respectively, following LOSO cross-validation. However, Gillespie uses 336 words from the UA-Speech corpus with 12 words per speaker, whereas we use all 455 words across all speakers.

In our first model, only dysarthric labels are observed and we achieved an accuracy on the word and speaker levels of 82% and 93% respectively. By training the multi-task model, in which both targets, i.e. mel-spectrogram and dysarthric labels, are observed, the accuracy on the word level increased by 3 percents to the value of 85.3% (Table 2). We found that the UA-Speech database includes multiple recorded words for healthy speakers that contain intelligibility errors, different words than asked or background speech of other people. These issues affect the accuracy of detecting dysarthric speech.

Krishna reports a 97.5% accuracy on UA-Corpus [7]. However, after email clarification with the author, we found that they estimated the accuracy taking into account only the speakers with a medium level of dysarthria. Narendra et al. achieved 93.06% utterance level accuracy on the TORGO dysarthric speech database [35]. As opposed to the related work, our model does not need any expert knowledge to design hand-crafted features and it can learn automatically using a low-dimensional latent space that encodes characteristics of dysarthria.

4.3 Interpretable modeling of dysarthric patterns

We analyze the correlation between the dysarthric latent space and the intelligibility of speakers. We look at 550 audio samples of a single ’Command’ word across the 15 dysarthric speakers and 13 healthy speakers.

In an unsupervised training (Figure 2), target labels of dysarthric/normal speech are not presented to the model. Dysarthric speakers are well separated from normal speakers and the dimension 2 of the latent space is negatively correlated with the intelligibility scores (Pearson correlation of -0.84, two-sided p-value < 0.001). In a supervised variant (Figure 3), we train the model jointly with both reconstructed mel-spectrogram and the target dysarthria labels observed. Both dimensions of the latent space are highly correlated with the intelligibility scores (dimension 1 with correlation of -0.76 and dimension 2 with correlation of 0.70, both with p-value < 0.001).

The sign of the correlation has no particular meaning. Retraining the model multiple times results in both positive and negative correlations between the latent space and the intelligibility of speech. A high correlation between dysarthric latent space and intelligibility scores suggests that by moving along the dimensions of the latent space, we should be able to reconstruct speech of dysarthric speakers and improve its intelligibility. We explore this in the next experiment.

4.4 Reconstruction of dysarthric speech

First we trained a supervised multi-speaker model with all dysarthric and control speakers but we achieved poor reconstruction results with almost unintelligible speech. We think this is due to a high variability of dysarthric speech across all speakers, including various articulation, prosody and fluency problems. To better understand the potential for speech reconstruction, we narrowed the experiment down to two speakers, male speaker M05 and a corresponding control speaker. We have chosen M05 subject because their speech varies across different levels of fluency and we wanted to observe this pattern when manipulating the latent space. For example, when pronouncing the word ’backspace’, M05 uttered consonants ’b’ and ’s’ multiple times, resulting in ’ba ba cs space’.

We analyzed a single category of 19 computer command words, such as ’command’ or ’backspace’. For every word uttered by M05, we generated 5 different versions of speech, fixing dimension 2 of the latent space to the value of -0.1, and using the values of [-0.5, 0, 0.5, 1, 1.5] for dimension 1. Audio samples of reconstructed speech were obtained by converting predicted mel-spectrograms to waveforms using the Griffin-Lim algorithm [36].

We conducted MUSHRA perceptual test [37]. Every listener was presented with 6 versions of a given word at the same time, 5 reconstructions and one version of recorded speech. We asked listeners to evaluate the fluency of speech on a scale from 0 to 100. We used 10 US based listeners from the Amazon mTurk platform, in total providing us with 1140 evaluated speech samples.

As shown in Figure 4, by moving along dimension 1 of the latent space, we can improve the fluency of speech, generating speech with levels of fluency not observed in the training data. In the pairwise two-sided Wilcoxon signed-rank, all pairs of ranks are different from each other with p-value < 0.001, except of {orig, d1=1.0}, {d1=-0.5, d1=0.0}, {d1=-0.5, d1=0.5}. Examples of original and reconstructed mel-spectrograms are shown in Figure 5.

We found that manipulation of the latent space changes both the fluency of speech and the timbre of voice and it is possible that dysarthria is so tied up with speaker identify making it fruitless to disentangle them. We replaced a deterministic dysarthric latent space with a Gaussian variable and trained the model with an additional Kullback-Leibler loss [19, 38] but we did not manage to separate the timbre of voice from dysarthria. Training the model with an additional discriminative cost to ensure that every dimension of the latent space is directly associated with a particular speech factor can potentially help with this problem [20].

5 Conclusions

This paper proposed a novel approach for the detection and reconstruction of dysarthric speech. The encoder-decoder model factorizes speech into a low-dimensional latent space and encoding of the input text. We showed that the latent space conveys interpretable characteristics of dysarthria, such as intelligibility and fluency of speech. MUSHRA perceptual test demonstrated that the adaptation of the latent space let the model generate speech of improved fluency. The multi-task supervised approach for predicting both the probability of dysarthric speech and the mel-spectrogram helps improve the detection of dysarthria with higher accuracy. This is thanks to a low-dimensional latent space of the auto-encoder as opposed to directly predicting dysarthria from a highly dimensional mel-spectrogram.

6 Acknowledgements

We would like to thank A. Nadolski, J. Droppo, J. Rohnke and V. Klimkov for insightful discussions on this work.