Advancing Dysarthric Speech-to-Text Recognition with LATTE: A Low-Latency Acoustic Modeling Approach for Real-Time Communication

Abstract

Dysarthria, a motor speech disorder characterized by slurred and often unintelligible speech, presents substantial challenges for effective communication. Conventional automatic speech recognition systems frequently underperform on dysarthric speech, particularly in severe cases. To address this gap, we introduce low-latency acoustic transcription and textual encoding (LATTE), an advanced framework designed for real-time dysarthric speech recognition. LATTE integrates preprocessing, acoustic processing, and transcription mapping into a unified pipeline, with its core powered by a hybrid architecture that combines convolutional layers for acoustic feature extraction with bidirectional temporal layers for modeling temporal dependencies. Evaluated on the UA-Speech dataset, LATTE achieves a word error rate of 12.5%, phoneme error rate of 8.3%, and a character error rate of 1%. By enabling accurate, low-latency transcription of impaired speech, LATTE provides a robust foundation for enhancing communication and accessibility in both digital applications and real-time interactive environments.

Keywords

acoustic features dysarthric speech speech impairment visual encoding

Introduction

Dysarthria is a complex motor speech disorder resulting from impaired neuromuscular control, characterized by difficulties in articulation, phonation, and prosody. This condition manifests as slurred, slow, or monotone speech, significantly compromising intelligibility and expressiveness.^1–6 Dysarthria often arises from neurological conditions such as stroke, Parkinson’s disease, and amyotrophic lateral sclerosis, leading to varied severity levels that can profoundly affect social interactions, emotional well-being, and overall quality of life. Individuals with dysarthria frequently encounter barriers to effective communication, resulting in frustration and diminished self-esteem, thereby impacting their participation in educational and occupational.⁷ Conventional automatic speech recognition (ASR) systems, including widely used technologies such as Google speech-to-text, excel at processing fluent, normal speech but struggle to accurately recognize dysarthric utterances. The irregularities associated with dysarthric speech such as imprecise phoneme production, variable speech tempo, and inconsistent vocal quality complicate the conversion of speech to text. Effective recognition is essential not only for facilitating direct communication but also for enabling various applications that require transforming dysarthric speech into comprehensible text, thus enhancing interaction with digital devices and services.⁸

Despite advancements in speech technology, there remains a considerable gap in developing ASR systems tailored for dysarthric speech. Previous research has primarily focused on conventional acoustic modeling approaches, which fail to accommodate the unique phonetic and acoustic challenges presented by dysarthria. Addressing these complexities is crucial for creating specialized ASR solutions capable of accurately decoding dysarthric utterances. Recent progress in neural networks, deep learning, and natural language processing has significantly advanced medical science, with successful applications in disease diagnosis, medical imaging, and neurological disorder detection.^9–13 Inspired by these achievements,^14–19 this study leverages state-of-the-art deep learning architectures to address the pressing challenge of dysarthric speech recognition, aiming to bring similar breakthroughs to the domain of impaired speech processing.

This study introduces an advanced low-latency acoustic transcription and textual encoding (LATTE) framework for dysarthric speech recognition. The framework integrates convolutional layers to extract salient spatial representations from complex acoustic patterns and employs bidirectional temporal layers to capture dependencies in both forward and backward directions. By combining these processes within a streamlined recognition pipeline, the LATTE framework effectively addresses the irregular rhythm, variable speech rates, and unpredictable pauses associated with dysarthria, while ensuring computational efficiency suitable for real-time applications. Ultimately, this research seeks to enhance the usability of speech recognition technologies for individuals affected by dysarthria, contributing to improved communication accessibility, inclusivity, and overall quality of life.

While prior studies have investigated transformer-based architectures and self-supervised frameworks for impaired or dysarthric speech, they often suffer from high latency and computational overhead, making them unsuitable for real-time deployment. Moreover, limited exploration has been directed toward low-latency architectures that balance speed, efficiency, and accuracy for spontaneous dysarthric speech. The proposed LATTE framework advances beyond prior work by introducing a lightweight, latency-optimized hybrid architecture that achieves faster inference with minimal accuracy degradation. Unlike earlier transformer or conventional models that emphasize accuracy at the cost of processing time, LATTE incorporates streamlined convolutional front-end layers and efficient temporal encoders, achieving real-time responsiveness without sacrificing recognition precision. This design uniquely positions LATTE as a bridge between clinical-grade speech recognition accuracy and practical, assistive technology usability, broadening its relevance across both health care and human–computer interaction domains.

The remainder of this article is organized as follows: The “Related Work” section reviews the related work in dysarthric speech recognition. The “Mathematical Representation of LATTE Framework” section presents the mathematical modeling of the proposed framework. The “Dataset and Participant Characteristics” section describes the dataset and participant characteristics. The “Methodology: LATTE Framework” section details the proposed LATTE framework. The “Evaluation and Results” section presents the evaluation criteria and the results. The “Ethical Considerations and Societal Impact” section discusses ethical considerations and social impact. The “Error Analysis” section describes an error analysis, and the “Conclusion and future works” section details the conclusion and the future directions.

Related Work

Dysarthric speech recognition has gained significant attention in recent years due to the growing need for accessible and efficient communication technologies for individuals with speech impairments. The inherent challenges of dysarthric speech such as high variability, reduced intelligibility, and limited corpus availability have motivated the adoption of advanced machine learning and deep learning architectures. Recent advancements in low-latency ASR and cross-lingual speech-to-text modeling have also contributed to improving real-time and multilingual speech systems, offering valuable insights applicable to impaired speech recognition tasks. For instance, authors in and²⁰ proposed a low-latency ASR framework based on dynamic streaming transformers optimized for real-time inference, while²¹ introduced a cross-lingual self-supervised speech-to-text model enabling robust multilingual adaptation. These developments highlight the broader research direction toward efficient, adaptive, and inclusive ASR systems. Consequently, hybrid deep architectures, transformer-based models, multimodal fusion techniques, and self-supervised learning (SSL) paradigms have been increasingly explored to enhance both recognition accuracy and computational efficiency. This section reviews state-of-the-art contributions and highlights methodological advancements across these dimensions, concluding with a structured analysis table that consolidates the key techniques and outcomes reported in the literature.

Geng et al.²² explored data augmentation techniques such as vocal tract length perturbation, tempo, and speed perturbation, applied to both normal and disordered speech. Using the UA-Speech corpus, their approach achieved a word error rate (WER) of 26.37%. Similarly, Celin et al.²³ proposed a two-level augmentation strategy using a virtual linear microphone array-based synthesis followed by multiresolution feature extraction. Their hybrid DNNHMM– system trained on UA-Speech and Tamil dysarthric speech corpora reported a WER of 35.75%.

Chandrakala et al.²⁴ introduced a histogram of states-based approach within a Deep Neural Network-Hidden Markov Model (DNN–HMM) framework, enabling compact and discriminative embeddings of dysarthric utterances. Their method reported WERs of 53.12%, 73.74%, and 43.15% across subsets of the UA-Speech corpus. Yu et al.²⁵ proposed the MAV-HuBERT (Mean Absolute Value–Hidden-Unit BERT) framework, which integrates both audio and visual modalities, incorporating a convolutional neural network (CNN)-based facial feature encoding in the first stage and AV-HuBERT pretraining in the second stage. This approach addressed model overfitting issues and achieved a WER of 63.98% on UA-Speech. Similarly, Xiong et al.²⁶ incorporated articulatory-based representations alongside acoustic features using LSTM-RNNs, reporting a WER of 48.36%.

Recent transformer-based approaches have further advanced the field. Mahum et al.²⁷ employed a Swin transformer for spectrogram-based spatial feature extraction, reaching a WER of 16.4%. He et al.²⁸ developed a dual-stream transformer integrating convolutional encoders and TDNN, achieving 17.3%. Yue et al.²⁹ proposed a multimodal transformer that combined audio and visual features, reporting 18.2%. Peng et al.³⁰ designed a phoneme-aware conformer, improving temporal alignment with a WER of 14.9%. Shahamiri et al.³¹ incorporated context-awareness into their transformer, yielding a WER of 15.1%. Hu et al.³² investigated multilingual Wav2Vec2.0 pretraining but faced generalization challenges, reporting 17.6%. Mehmood et al.³³ utilized an SSL–CTC framework, with a WER of 19.3%. Finally, Shujie et al.³⁴ integrated SSL into TDNN–conformer fusion with Wav2Vec2.0-based rescoring, achieving an improved WER of 18.17% on UA-Speech.

Table 1 summarizes these studies, highlighting the techniques, corpora, and reported recognition performance.

Table 1.

Comparative analysis of related work in dysarthric speech recognition

Study (year)	Technique	Corpus	WER (%)
Xiong et al. (2018)²⁶	Acoustic–articulatory fusion using LSTM–RNN	UA-Speech	48.36
Celin et al. (2020)²³	Virtual microphone array + hybrid DNN–HMM	UA-Speech, Tamil	35.75
Chandrakala et al. (2021)²⁴	HoS embeddings within DNN–HMM	UA-Speech	43.15–73.74
Eng et al. (2022)	VTLP + tempo/speed perturbation	UA-Speech	26.37
Yu et al. (2023)²⁵	MAV-HuBERT + CNN visual encoder	UA-Speech	63.98
Shujie et al. (2023)³⁴	TDNN–Conformer fusion + W2V2 rescoring	UA-Speech	18.17
Hu et al. (2024)³²	Multilingual W2V2	Multiple	17.60
Mahum et al. (2025)²⁷	Swin Transformer	UA-Speech	16.40
He et al. (2025)²⁸	Dual-stream Transformer	UA-Speech	17.30
Yue et al. (2025)²⁹	Multimodal Transformer	UA-Speech	18.20
Peng et al. (2025)³⁰	Phoneme-aware Conformer	UA-Speech	14.90
Shahamiri et al. (2025)³¹	Context-aware Transformer	UA-Speech	15.10
Mehmood et al. (2025)³³	SSL–CTC framework	UA-Speech	19.30
Proposed LATTE (2025)	CNN + BiLSTM temporal modeling	UA-Speech	12.50

BiLSTM, bidirectional long short-term memory; CNN, convolutional neural network; LATTE, low-latency acoustic transcription and textual encoding; SSL, self-supervised learning; VTLP, vocal tract length perturbation; WER, word error rate; W2V2, Wav2Vec 2.0.

Bold value highlights the result of the proposed framework.

Mathematical Representation of LATTE Framework

In this study, the task of dysarthric speech recognition is formulated as the mapping of impaired speech signals into their most probable textual encodings under a low-latency constraint. The proposed framework, termed LATTE, is designed to integrate acoustic modeling, temporal mapping, and recognition into a unified hybrid structure. Given an input acoustic sequence

X = {x_{1}, x_{2}, ..., x_{T}}, x_{t} \in R^{d},

where T denotes the total number of time frames, and d is the feature dimension (e.g., Mel-frequency cepstral coefficient [MFCC] or spectrogram features), the objective of LATTE is to generate the most probable recognized word sequence

Y = {y_{1}, y_{2}, ..., y_{N}}, y_{n} \in V,

where N is the output sequence length, and V denotes the vocabulary set.

Formally, dysarthric speech recognition in LATTE can be expressed as a conditional probability maximization:

\overset{\land}{Y} = \arg \max_{Y} P (Y | X; θ)

(1)

where θ represents the learnable parameters of the LATTE framework.

The framework combines convolutional layers for acoustic feature extraction with bidirectional long short-term memory (BiLSTM) layers for temporal mapping and contextual modeling. The CNN component enhances robustness by extracting salient local spectral–temporal features from dysarthric speech, reducing distortions due to articulation impairments. The BiLSTM component then captures forward and backward temporal dependencies, which are crucial for recognizing speech with irregular rhythm, variable pacing, and inconsistent prosody. The CNN feature extraction process is defined as:

F = f_{CNN} (X; θ_{c})

(2)

where F represents the extracted feature maps, and θ _c are the parameters of the CNN layers. The BiLSTM-based temporal encoding operates on F as:

\vec{h} t = LSTMf (F t, \vec{h} t - 1; θ_{f}),

(3)

\overset{\leftarrow}{h} t = LSTMb (F t, \overset{\leftarrow}{h} t + 1; θ_{b}),

(4)

h t = [\vec{h} t; \overset{\leftarrow}{h} t],

(5)

where h_t denotes the concatenated hidden representation at time t, while θ _f and θ _b are the parameters of the forward and backward LSTM layers, respectively. Finally, the recognition output is computed through a softmax classification layer:

P (y_{t} | h_{t}) = Softmax (W h_{t} + b),

(6)

where W and b denote the transformation weights and bias terms of the output layer, ensuring the mapping from encoded features to textual symbols. The overall training objective of LATTE is to minimize the negative log-likelihood loss:

L (θ) = \sum_{(X, Y) \in D} \log P (Y | X; θ),

(7)

where D is the dysarthric speech dataset. This objective ensures that LATTE learns an efficient low-latency mapping from impaired acoustic patterns to textual encodings, thereby achieving robust and accurate dysarthric speech recognition.

Dataset and Participant Characteristics

The availability of publicly accessible speech corpora containing dysarthric speech samples is limited. Notable examples include Nemours,³⁵ TORGO,³⁶ and UA-Speech.³⁷ Additionally, Google has developed the Euphonia dataset,³⁸ which includes dysarthric samples; however, this dataset was not accessible during the course of this study, and our request for access was not fulfilled. Among the existing datasets, UA-Speech is the largest, encompassing a higher number of dysarthric participants and being the most frequently utilized corpus in dysarthric speech recognition research.³⁹ Consequently, it serves as the foundation for this study.

The UA-Speech corpus is a widely recognized dataset specifically designed for research in dysarthric speech recognition. This dataset is invaluable for developing and evaluating models aimed at improving speech-to-text transcription for individuals with dysarthria. It encompasses a diverse range of speech samples from a variety of speakers, including different ages, genders, and levels of dysarthria, thereby providing a broad representation of dysarthric speech patterns. UA-Speech was developed by the University of Illinois researchers and features speech samples obtained from 19 dysarthric individuals with speech intelligibility levels ranging from 2% to 95%. The intelligibility levels of dysarthric speakers range from very low (0%–25%) intelligibility to low (25%–50%), mild (50%–75%), and high (75%–100%) intelligibility. Speech intelligibility can be defined as to what extent speech is comprehensible by a typical listener and is one of the mechanisms to define the severity of dysarthria. The corpus overall provides speech samples collected from 28 speakers, including 15 dysarthric speakers and 13 healthy control speakers—the speech samples of the other four dysarthric participants are not publicly available (speakers M02, M03, F01, M06). The dataset features both spontaneous and read speech recordings, which are instrumental in capturing the inherent variability and distinctive characteristics of dysarthric speech. Each speech sample is meticulously transcribed with phonetic and orthographic annotations, ensuring precise alignment between spoken utterances and their textual counterparts. Such detailed transcription is vital for training effective and accurate speech recognition models. Furthermore, the recordings are conducted under controlled conditions to minimize background noise and recording artifacts, thus enhancing the overall data quality and reliability. With thousands of utterances available, the UA-Speech corpus provides a robust foundation for training, validating, and testing various speech recognition systems. Its extensive utilization in diverse research studies focused on speech synthesis, recognition, and pathology underscores its significance as a fundamental resource for advancing dysarthric speech recognition technologies, ultimately facilitating improved communication for individuals affected by dysarthria. The dataset details are summarized in Table 2.

Table 2.

Summary of the UA-Speech corpus and participant characteristics

ID	Participant	Gender	Age	Intelligibility score	Level (%)
1	M04	M	>18	2
2	M03	F	51	6	Very low
3	M12	M	19	7
4	M01	M	>18	17
5	M07	M	58	28
6	M502	F	30	29	Low
7	M16	M	40	43
8	M05	M	21	58
9	M11	M	48	62	Mild
10	M04	F	18	62
11	M09	M	18	86
12	M14	M	44	90
13	M10	M	21	93	High
14	M08	M	28	95
15	M05	F	22	95
16–27	Control (UA)	4 F/7 M	N/A	N/A	Healthy
28	LJ Speech	M	N/A	N/A	Healthy

F, female; M, male; N/A, not available.

Methodology: LATTE Framework

In this section, we introduce the proposed LATTE framework for dysarthric speech recognition. The section discusses the overall pipeline of the framework, including input processing and transcription mapping of the dysarthric corpus, the acoustic preprocessing and feature extraction procedures, and finally the architecture of the proposed framework. Together, these components outline how LATTE systematically processes speech data, aligns it with textual representations, and supports efficient and reliable recognition.

Preprocessing and transcription mapping for dysarthric corpus

This study utilizes the UA-Speech corpus, a benchmark dataset for dysarthric speech recognition. Preprocessing plays a pivotal role in organizing and standardizing the corpus, thereby ensuring accurate feature extraction and reliable model training. Key parameters such as MAX_LEN, VOCAB_SIZE, and NUM_MFCC were defined to control input sequence length, vocabulary coverage, and the dimensionality of MFCCs. Directory paths for audio files, Master Label Files (MLFs), lexicons, and speaker wordlists were systematically specified to streamline resource access.

Metadata was extracted from speaker and word-recording Excel sheets, enabling a precise mapping of each utterance to its corresponding audio file. The MLFs were parsed to align utterance-level transcriptions with audio samples, and tokens were normalized to uppercase for consistency. The final output was a structured dictionary, where audio filenames served as keys and transcribed sequences as values, ensuring robust pairing between speech and text.

The overall process is illustrated in Figure 1, while the detailed dataset mapping is provided in Table 3. This condensed pipeline establishes a reliable foundation for addressing the challenges of impaired speech recognition.

FIG. 1.

Condensed preprocessing pipeline for dysarthric speech corpus.

Table 3.

Systematic mapping of speech utterances to audio files and transcriptions

Speaker ID	Audio filename	Transcription
M01	M01_WORD1.wav	WORD1
M01	M01_WORD2.wav	WORD2
F02	F02_WORD1.wav	WORD1
F02	F02_WORD2.wav	WORD2

Acoustic preprocessing and feature extraction

Feature extraction transforms raw dysarthric speech into discriminative representations suitable for recognition. In this study, we employ MFCCs, which capture perceptually relevant spectral properties of speech.

The process begins by segmenting the signal into overlapping frames, followed by applying the discrete Fourier transform to obtain frequency spectra. These spectra are filtered on the Mel scale to emphasize human auditory sensitivity, and the logarithm of the filterbank energies is decorrelated using the discrete cosine transform (DCT). The i-th MFCC coefficient is computed as:

MFCCi = \sum k = 1^{K} \log (X_{k}) \cdot \cos \frac{i (k - 0.5) π}{K},

(8)

where X _k is the Mel-filtered spectral magnitude, and K is the number of Mel bands. This yields frame-level feature vectors compactly representing spectral envelopes.

For each audio sample Xn with sampling rate sr, the MFCC matrix is given by:

F_{n} = MFCC (x_{n}, s r),

(9)

with rows corresponding to frames and columns to cepstral coefficients. These features are paired with transcriptions y _n , yielding the dataset:

D = {(F_{n}, y_{n}) | n = 1, 2, ..., N} .

(10)

This structured representation enables the model to learn mappings between impaired acoustic patterns and their linguistic transcriptions. Figure 2 illustrates the MFCC extraction pipeline. The raw dysarthric speech signal is first segmented using framing and windowing, followed by Fast Fourier Transform to convert each frame into its frequency-domain representation. The Mel filter bank then emphasizes perceptually meaningful frequency bands, and logarithmic compression stabilizes amplitude variations. Finally, the DCT decorrelates the compressed filterbank energies to produce compact MFCC feature vectors used as model inputs.

FIG. 2.

Layered MFCC extraction pipeline for dysarthric speech. MFCC, Mel-frequency cepstral coefficient.

To prepare the dataset, transcriptions were first loaded into a dictionary mapping each audio identifier to its text. The tokenizer converted text into integer sequences with a vocabulary size defined by VOCAB_SIZE, and sequences were padded for uniformity. Input features X were similarly padded, and the dataset was split into training and testing subsets (80:20).

Since model outputs operate on fixed time steps, transcription labels were downsampled to match the output resolution. Given a sequence y and target length L, downsampling is defined as:

y_{downsampled} = y [: : \frac{| y |}{L}] [: L] .

(11)

This alignment ensures temporal consistency between labels and predicted outputs, stabilizing model training and evaluation. In our case, all label sequences were standardized to 25 steps. Table 4 summarizes the preprocessing parameters applied.

Table 4.

Summary of preprocessing parameters

Parameter	Value
Input sequence length (MAX_LEN)	Fixed, padded
Vocabulary size (VOCAB_SIZE)	Derived from tokenizer
MFCC coefficients (NUM_MFCC)	13 per frame
Output time steps (L)	25
Train–test split	80:20

LATTE architecture design

This section describes the LATTE architecture utilized for dysarthric speech recognition. The model integrates CNN layers and BiLSTM layers to effectively process and classify feature-extracted speech data. The architecture is designed to leverage the strengths of both CNNs and LSTMs, capturing both local features and temporal dependencies from the input speech sequences.

The input to the CNN–BiLSTM model is a sequence of feature vectors derived from MFCCs. Each input sequence X is represented as:

X = [x_{1}, x_{2}, ..., x_{T}]

(12)

where x _t is the feature vector at time step t, and T represents the total number of time steps in the sequence.

In the CNN component, the model employs one-dimensional convolutional layers to extract local patterns from the input sequences. The convolution operation is defined as:

F k = σ \sum i = 1^{d} w_{ki} X + b_{k}

(13)

where w_ki denotes the weight of the k-th filter, b_k is the bias term for the k-th filter, d is the kernel size, and σ (x) = max(0, x) is the ReLU activation function. The model includes two convolutional layers: the first with 64 filters and a kernel size of 3, and the second with 128 filters and a kernel size of 3. Dropout with a rate of 0.2 is applied after each convolutional block to reduce overfitting. Following the CNN layers, the model incorporates two BiLSTM layers. LSTMs are effective in learning long-term dependencies by maintaining internal state vectors. The BiLSTM processes the input sequence in both forward and backward directions, enhancing the model’s ability to capture contextual information. The LSTM output at time step t is given by:

h t = LSTM (F t, h t - 1, c t - 1)

(14)

The output from the BiLSTM is concatenated as:

H_{t} = h_{\vec{h_{t}}}_{, \overset{\leftarrow}{h_{t}}} i

(15)

where ht represents the hidden state at time step t, F _t is the input from the CNN layers, and ht − 1 and c_{t −1} are the hidden and cell states from the previous time step, respectively.

The subsequent layer is a TimeDistributed Dense layer with 64 units and ReLU activation. This layer is applied to each time step to map the LSTM outputs into a higher-dimensional space while preserving the temporal structure.

The final output is produced by a TimeDistributed Dense layer with softmax activation. This layer outputs a probability distribution over the vocabulary size. The softmax function is defined as:

P (y_{t} = c | H t) = \frac{e^{W_{c} H_{t}}}{- j = 1^{C} e^{W_{j}}^{H_{t}}}

(16)

where W _c is the weight matrix for class c, C is the total number of classes (vocabulary size), and P (y_t = c | H) is the probability of class c at time step t. The model uses the sparse categorical cross-entropy loss function, defined as:

L = \frac{\sum^{T}}{t = 1} \log P (y_{t} | H_{t})

(17)

where L is the loss, and P (y_t | H_t) is the predicted probability for the true class at time step t. For optimization, the Adam optimizer is employed, which dynamically adjusts the learning rate during training. The model is compiled with the sparse categorical cross-entropy loss function and accuracy as the performance metric. This hybrid model, by combining CNN for feature extraction and LSTM for sequence modeling, effectively captures both spectral and temporal characteristics of dysarthric speech, providing a robust framework for speech recognition tasks. The model training process involves feeding the model with the downsampled training data and labels, while simultaneously evaluating its performance on the downsampled validation data. The training is conducted over 10 epochs with a batch size of 16. Training a model on speech recognition data involves adjusting the model parameters to minimize the loss function and improve recognition accuracy. During this phase, the model learns to extract relevant features from the input data and map them to the correct transcriptions. By iterating through the data over multiple epochs, the model’s parameters are fine-tuned to enhance its ability to generalize to new, unseen data. The validation data help monitor the model’s performance throughout the training process, ensuring that the model does not overfit to the training data. Regular evaluation on the validation set provides insights into the model’s accuracy and robustness, which are crucial for effective speech recognition.

The architecture of the proposed model LATTE is outlined in Table 5, which presents the structure of the network, including the types of layers used, their output shapes, and the number of trainable parameters. This table provides a clear depiction of the model’s overall complexity and computational footprint. The model begins with an input layer that defines the input dimensions but contains no trainable parameters. Following this, a series of convolutional layers (e.g., Conv1D) are employed to extract local features from the input data. These layers contribute a substantial number of parameters—for instance, the first convolutional layer comprises approximately 7744 trainable parameters. To reduce the feature dimensionality and retain the most salient characteristics, pooling layers such as MaxPooling1D are applied, which do not introduce additional parameters. To mitigate overfitting and enhance generalization, dropout layers are interleaved within the architecture; these layers also contain no parameters. The subsequent bidirectional recurrent layers play a crucial role in capturing long-range temporal dependencies in both forward and backward directions. These layers substantially increase the parameter count, with two bidirectional layers containing approximately 263,168 and 394,240 parameters, respectively. Finally, TimeDistributed layers are utilized to apply dense transformations across each temporal step in the sequence, with parameter counts of around 16,448 and 29,575 for the two such layers.

Table 5.

Model architecture overview

Layer (type)	Output shape	Parameters
InputLayer	(None, 100, 40)	0
Conv1D	(None, 100, 64)	7744
MaxPooling1D	(None, 50, 64)	0
Dropout	(None, 50, 64)	0
Conv1D	(None, 50, 128)	24,704
MaxPooling1D	(None, 25, 128)	0
Dropout	(None, 25, 128)	0
Bidirectional LSTM	(None, 25, 256)	263,168
Dropout	(None, 25, 256)	0
Bidirectional LSTM	(None, 25, 256)	394,240
Dropout	(None, 25, 256)	0
TimeDistributed Dense	(None, 25, 64)	16,448
TimeDistributed Dense	(None, 25, 455)	29,575

This structured architecture enables efficient feature extraction, temporal modeling, and sequential representation learning while maintaining a balance between performance and computational efficiency.

Experimental settings

For reproducibility, the proposed LATTE framework was evaluated on the UA-Speech corpus. Audio was resampled to 16 kHz, and 13-dimensional MFCCs were extracted with 25 ms frames and 10 ms overlap. Transcriptions were tokenized, padded to 25 steps, and split 80:20 for training and testing. The LATTE framework used 2 CNN layers (64 and 128 filters, kernel size 3), 2 BiLSTM layers (128 units each), and TimeDistributed Dense layers. The dropout rate was 0.2. Training used the Adam optimizer (learning rate 0.001), sparse categorical cross-entropy loss, batch size 16, and 10 epochs. Experiments were conducted on a workstation with an Intel Core i7 CPU, NVIDIA RTX 3080 GPU, 32 GB RAM, Python 3.11, and TensorFlow 2.14.

Computational efficiency analysis

The computational efficiency of the proposed LATTE framework is comprehensively evaluated in terms of trainable parameters, time complexity, and memory utilization. The LATTE framework has been meticulously designed to ensure real-time operability, low-latency response, and lightweight processing, which are essential characteristics for dysarthric speech recognition in practical communication systems. LATTE integrates two principal processing modules: convolutional layers for acoustic feature extraction and temporal layers for modeling sequential dependencies across speech segments. This design paradigm not only enhances the efficiency of feature representation but also maintains computational scalability across both training and inference phases. The computational complexity of the convolutional block in LATTE can be expressed as:

O_{conv} = K \times F_{i n} \times F_{out} \times T

where K denotes the kernel size, F_in and F_out represent the number of input and output filters, and T is the temporal sequence length. These layers perform spectral decomposition and spatial abstraction, progressively condensing the input representation passed to subsequent modules. The temporal modeling stage, consisting of bidirectional recurrent processing, is governed by the following complexity:

O_{temp} = 4 \times (n^{2} + n \times m)

where n is the number of hidden units, and m represents the feature dimension per timestep. This accounts for the gating operations inherent in the temporal modeling mechanism. Despite the bidirectional processing, LATTE achieves notable computational efficiency due to the reduced temporal resolution generated by prior convolutional compression. The final dense projection layer has the complexity:

O_{dense} = N_{i n} \times N_{out}

where N_in and N_out represent the input and output dimensions, respectively. This layer transforms the encoded acoustic representations into the target text domain, optimizing the speech-to-text alignment.

The overall computational footprint of LATTE as presented in Table 6. demonstrates an optimal balance between model expressiveness and runtime efficiency. Parameter sharing, dropout regularization, and hierarchical feature compression further minimize redundancy and enhance the system’s capacity for low-resource deployment.

Table 6.

Computational efficiency summary of the low-latency acoustic transcription and textual encoding framework

Component	Complexity/Detail
Convolutional layer 1	0 (3 × 40 × 64 × 100)
Convolutional layer 2	0 (3 × 64 × 128 × 50)
Temporal layer 1	0 (8 × (1282 + 128 × 128))
Temporal layer 2	0 (8 × (1282 + 128 × 128))
Time-distributed dense	0 (64 × 455)
Total trainable parameters	≈ 1.25 M
Training time (per epoch)	≈ 42 seconds
Inference time (per sample)	≈ 0.09 seconds

These findings confirm that LATTE achieves real-time inference with minimal computational overhead, substantiating its suitability for deployment in resource-constrained and latency-sensitive speech recognition environments. The architecture’s streamlined acoustic-to-textual processing pipeline ensures superior speed, scalability, and precision, distinguishing LATTE as a robust framework for real-time dysarthric speech communication.

Evaluation and Results

To evaluate the performance of our proposed dysarthric speech recognition model, we employ several widely used metrics: accuracy, loss, character error rate (CER), WER, phoneme error rate (PER), and F1-score. These metrics together provide a comprehensive view of transcription performance, spanning overall prediction correctness, character- and phoneme-level accuracy, and segment-level balance between precision and recall. LATTE is evaluated on a corpus of UA-Speech English dysarthric speech, which contains diverse speakers and exhibits irregular articulation patterns, making it an ideal benchmark for assessing transcription robustness.

Accuracy measures the proportion of correctly predicted samples among all predictions, while loss quantifies the difference between predicted and true labels. Accuracy and categorical cross-entropy loss are defined as:

Accuracy = \frac{Number o f Correct Predictions}{Total Number o f Predictions} \times 100 %, Loss = {\frac{1}{N}}_{i = 1} \sum_{N} y_{i} \log (p_{i})

(18)

where N is the total number of samples, y_i is the binary indicator for the correct class of sample, and p_i is the predicted probability of sample i belonging to the positive class.

WER, CER, and PER evaluate transcription errors at the word, character, and phoneme levels, respectively. These metrics are especially important in dysarthric speech due to irregular phoneme production and atypical articulation. They are calculated as:

WER = \frac{S + D + I}{N}, CER = \frac{S + D + I}{N}, PER = \frac{S + D + I}{N}

(19)

where S, D, and I denote substitutions, deletions, and insertions, and N is the total number of words (for WER), characters (for CER), or phonemes (for PER) in the reference transcription. These metrics allow evaluation from coarse word-level errors down to fine-grained phoneme-level transcription mistakes, highlighting LATTE’s ability to capture subtle speech patterns.

F1-score provides a harmonic mean of precision and recall, balancing over- and under-prediction for segment-level recognition:

F 1 = 2 \cdot \frac{\Pr ecision \cdot Re call}{\Pr ecision + Re call}

(20)

where

\Pr ecision = \frac{True Positives}{True Positives + False Positives}

and

Re call = \frac{True Positives}{True Positives + False Positives}

. High F1-score indicates both accurate and reliable transcription at the segment level, which is critical in real-time speech-to-text applications.

Training and validation performance

LATTE was trained for 10 epochs. Table 7 summarizes epoch-wise training and validation performance.

Table 7.

Epoch-wise training and validation performance of low-latency acoustic transcription and textual encoding

Epoch	Training accuracy	Training loss	Validation accuracy	Validation loss
1	0.9563	0.3580	0.9604	0.2240
2	0.9604	0.2171	0.9608	0.1990
3	0.9612	0.1919	0.9625	0.1710
4	0.9626	0.1683	0.9654	0.1444
5	0.9655	0.1419	0.9699	0.1141
6	0.9686	0.1203	0.9732	0.0967
7	0.9716	0.1028	0.9776	0.0786
8	0.9741	0.0908	0.9790	0.0728
9	0.9762	0.0821	0.9813	0.0632
10	0.9780	0.0747	0.9827	0.0572

As observed in Figure 3, training accuracy improves consistently across epochs, reflecting progressive learning of dysarthric speech patterns. Validation accuracy rises in parallel, indicating strong generalization. Concurrently, training and validation losses decrease steadily, demonstrating effective optimization and feature extraction at character, phoneme, and word levels.

The accuracy curves as indicated in Fig. 3 show that LATTE is progressively refining its ability to map complex dysarthric speech to correct textual output. Similarly, the loss curves confirm effective convergence and reduction of transcription errors. Table 8 presents a detailed evaluation across all metrics.

Table 8.

Error metrics of the proposed low-latency acoustic transcription and textual encoding model

Metric	Value (%)
Word error rate	12.5
Character error rate	1.0
Phoneme error rate	8.3
F1-score	96.1

The low CER and PER as illlustrated in Fig. 4 highlight LATTE’s ability to accurately capture fine-grained phonetic and character-level patterns. The WER demonstrates effective word-level transcription, even with irregular articulation, while the high F1-score confirms a balance between precision and recall at the segment-level transcription.

Ablation study analysis

In this study, we conducted a series of experiments to evaluate the individual contributions of CNNs and BiLSTM networks in dysarthric speech recognition. These experiments were performed on the UA-Speech dataset. We first employed a CNN-based model for dysarthric speech recognition. The model achieved a WER of 93.87%, indicating that while CNNs are effective in capturing local features in speech signals, they may not fully capture the temporal dependencies inherent in dysarthric speech. Next, we applied a BiLSTM-based model. The BiLSTM model achieved a WER of 77.5%, demonstrating its capability to model temporal dependencies in speech. However, the performance was still suboptimal, suggesting that solely relying on temporal modeling may not be sufficient for accurate dysarthric speech recognition. Subsequently, we integrated both CNN and BiLSTM components into our proposed LATTE framework. This hybrid approach achieved a WER of 12.5%, significantly outperforming both the CNN-only and BiLSTM-only models. The integration of CNN allowed for the extraction of local features, while the BiLSTM component effectively modeled the temporal dependencies, leading to improved performance. These results underscore the complementary strengths of CNNs and BiLSTMs in dysarthric speech recognition. While CNNs excel at capturing local features, BiLSTMs are adept at modeling temporal dependencies. The combination of both in the LATTE framework leverages these strengths, leading to superior performance in recognizing dysarthric speech patterns. An overview of ablation study is presented in Fig. 5.

Comparison with baseline architectures

To validate LATTE, we compared it with two recent strong baselines: Hu et al. 2024 (Multilingual Wav2Vec2.0) and Mahum et al. 2025 (Swin Transformer). Table 9 summarizes the comparison.

Table 9.

Comparison of proposed low-latency acoustic transcription and textual encoding with recent baselines

Model	WER (%)	CER (%)	Inference (ms/sample)
Hu et al. (2024)³²	17.6	2.3	35
Mahum et al. (2025)²⁷	16.4	1.8	40
Proposed LATTE	12.5	1.0	18

CER, character error rate.

Bold value highlights the result of the proposed framework.

LATTE achieves lower WER and CER than baselines and maintains nearly half the inference time, demonstrating superior transcription performance and low-latency applicability. Figure 6 illustrates this comparison visually. Overall, LATTE shows strong generalization, consistent low error rates across characters, words, and phonemes, and efficient computation, confirming its suitability for real-time dysarthric speech-to-text transcription.

FIG. 3.

Training and validation performance of LATTE. Left: accuracy curves; right: loss curves across 10 epochs. The minimal gap indicates strong generalization. LATTE, low-latency acoustic transcription and textual encoding.

FIG. 4.

Box plot of WER for LATTE on test samples, illustrating low error variance and robust transcription. WER, word error rate.

FIG. 5.

Component-wise ablation study analysis of LAT architecture on dysarthric speech. CER, character error rate; PER, phoneme error rate.

FIG. 6.

Bar chart comparison of WER and CER for LATTE and recent baselines.

Ethical Considerations and Societal Impact

This research adheres to the highest ethical standards, particularly given the sensitivity of working with dysarthric speech data. All data from the UA-Speech corpus were handled with strict privacy protocols, ensuring that participant identities were anonymized and protected. Proper informed consent was ensured wherever applicable, especially for those with speech impairments, allowing participants to fully understand the purpose and use of their data in developing assistive technologies. Bias in model development was another key consideration. Dysarthric speech varies significantly across individuals, and we aimed to develop a model that could recognize a wide range of speech patterns inclusively. This helps ensure that no specific group of individuals with speech disorders is disproportionately disadvantaged by the system’s performance. Additionally, ethical safeguards are in place to prevent misuse of the technology, particularly in contexts where privacy concerns or harm could arise. The societal impact of this research is significant, particularly in supporting individuals with speech impairments in clinical and social settings. Improved speech-to-text transcription can assist clinicians in diagnosing and monitoring speech disorders more effectively, while also enhancing social integration for individuals with dysarthria by enabling clearer and more accurate communication. This work can substantially reduce the communication barriers faced by dysarthric patients, contributing positively to their quality of life and their broader societal inclusion.

Error Analysis

The recognition of dysarthric speech inherently involves significant acoustic and articulatory variability, especially across different severity levels of impairment. In the LATTE framework, three primary sources of transcriptional errors were observed: (i) variability in articulatory precision, (ii) spectral–temporal distortion due to inconsistent phoneme production, and (iii) dataset imbalance across severity categories. To provide a deeper insight, we further quantified performance variation across dysarthria severity levels, categorized as mild, moderate, and severe based on the intelligibility annotations of the UA-Speech corpus. Table 10 presents the WER, PER, and CER across these levels.

Table 10.

Severity-wise error analysis of dysarthric speech recognition using low-latency acoustic transcription and textual encoding

Severity level	WER (%)	PER (%)	CER (%)
Mild dysarthria	7.4	5.6	0.6
Moderate dysarthria	11.8	8.9	0.9
Severe dysarthria	18.9	11.3	1.5
Overall (average)	12.5	8.3	1.0

PER, phoneme error rate.

As shown, the error rate progressively increases with dysarthria severity. LATTE maintains strong performance for mild and moderately impaired speakers due to its CNN–BiLSTM hybrid design, where convolutional layers capture short-term acoustic cues, and the BiLSTM units effectively model temporal coarticulation dynamics. However, in severe dysarthria, significant phoneme reduction, elongated articulatory transitions, and reduced formant stability cause higher phoneme substitution and deletion errors, resulting in a noticeable rise in WER and PER. The hybrid network exhibits limitations in capturing extremely distorted articulatory trajectories where intra-speaker variability dominates.

Additionally, the scarcity of training samples for severe cases within the UA-Speech dataset introduces class imbalance, affecting model generalization. Background noise and microphone inconsistencies further contribute to temporal misalignment between predicted and reference phoneme boundaries. These observations underline the importance of enhanced preprocessing, intelligibility-based data augmentation, and adaptive acoustic modeling to improve robustness in low-intelligibility scenarios. Future extensions of LATTE will explore intelligibility-aware loss functions, attention-based refinement, and transformer-assisted spectral encoding to minimize such degradation across severity levels.

Conclusion and Future Work

Dysarthric speech, characterized by slow, slurred, and often indiscernible patterns, presents a significant challenge for ASR systems. The intricacies of dysarthric speech patterns, combined with its sensitive nature, make it difficult for conventional models to accurately capture and transcribe. Existing research in the field has often reported higher WERs, underscoring the limitations of previous approaches in recognizing such speech accurately. This limitation poses a significant barrier for dysarthric patients, whose ability to communicate through automated systems is hindered by these recognition challenges. This study has proposed an advanced LATTE architecture aimed at addressing these difficulties in dysarthric speech recognition. Our approach has demonstrated a marked improvement in accuracy and a substantial reduction in the WER compared with existing models. By focusing on the specific challenges of dysarthric speech, such as its unpredictable variations and slurred articulation, our model was able to produce more reliable speech-to-text transcriptions.

While the results are promising, certain limitations remain. The experiments were conducted on the UA-Speech dataset, which primarily contains English-language recordings; therefore, the generalizability of the proposed model to other languages or datasets with differing phonetic and acoustic structures may require further validation. Moreover, as the model leverages deep architectures combining convolutional and recurrent components, the computational complexity can be significant, particularly for real-time or resource-constrained applications. Future research will explore model compression, cross-linguistic adaptation, and dataset diversification to enhance the scalability and applicability of the LATTE framework across diverse linguistic and computational settings.

The implications of this work extend beyond technical accuracy. Accurate speech transcription for dysarthric patients can significantly enhance their ability to communicate, offering vital support in both personal and clinical settings. By reducing the gap between spoken words and their textual counterparts, this technology can alleviate some of the social and psychological impacts experienced by individuals with dysarthria. Ultimately, our proposed framework contributes to an improved quality of life for dysarthric individuals, fostering greater independence and social inclusion, while also providing clinicians with a more robust tool for supporting their patients. This research underscores the importance of continued innovation in the field of dysarthric speech recognition and highlights the potential for technological advances to have a profound positive impact on society.

Footnotes

Author Disclosure Statement

No competing financial interests exist.

Funding Information

No funding was received for this article.

Authors’sContributions

Qurat Ul Ain led the overall research and made the primary intellectual contribution to this study. She conceptualized the research problem, designed the methodological framework, carried out data preprocessing and feature extraction, developed and implemented the proposed models, conducted all experimental evaluations, and performed an in-depth analysis of the results. She also drafted the original manuscript and finalized it through multiple revisions in accordance with journal standards. Hammad Afzal supervised the research and provided substantial academic and technical guidance throughout the study. He contributed to refining the research objectives, validating the methodological design, and critically reviewing the experimental setup and performance evaluation strategy. Fazli Subhan supported the preliminary analysis and assisted in reviewing the manuscript and provided suggestions that helped improve clarity and presentation. Mazliham Mohd Suud contributed to the literature review process and assisted in manuscript proofreading and formatting. Youghyun Jung provided high-level academic insight and reviewed the manuscript to ensure technical accuracy, coherence, and alignment with the scope of the journal.

Abbreviations

References

Freed

. Motor speech disorders: diagnosis and treatment. Plural Publishing, 2023.

Enderby

. Disorders of communication: Dysarthria. Handb Clin Neurol, 2013; 110:273–281.

Castillo Guerra

, Lovey

. A modern approach to dysarthria classification. Proceedings of the 25th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (IEEE Cat. No. 03CH37439). Vol. 3. IEEE, 2003.

Albaqshi

, Sagheer

. Dysarthric speech recognition using convolutional recurrent neural networks. International Journal of Intelligent Engineering Systems, 2020; 13(6):384–392.

Barrett

, Hu

, Howell

. Systematic review of machine learning approaches for detecting developmental stuttering. IEEE/ACM Trans Audio Speech Lang Process, 2022; 30:1160–1172.

Joshy

, Rajan

. Automated dysarthria severity classification: A study on acoustic features and deep learning techniques. IEEE Trans Neural Syst Rehabil Eng, 2022; 30:1147–1157.

Qian

, Xiao

. A Survey of automatic speech recognition for dysarthric speech. Electronics (Basel), 2023; 12(20):4278.

Saksamudre

, Shrishrimal

, Deshmukh

. A review on different approaches for speech recognition system. Int J Comput Appl, 2015; 115(22):23–28.

Okolo

, et al. Data-driven approaches to bridging the gap in health communication disparities: A systematic review. World Journal of Advanced Research and Reviews, 2024; 21(2):1435–1445.

10.

Singhania

, Reddy

. Improving preventative care and health outcomes for patients with chronic diseases using big data-driven insights and predictive modeling. International Journal of Applied Health Care Analytics, 2024; 9(2):1–14.

11.

Veetil

, V

, Orozco-Arroyave

, et al. Robust language independent voice data driven Parkinson’s disease detection. Eng Appl Artif Intell, 2024; 129:107494.

12.

Salameh

, Surakhi

OLAM

, Khanafseh

. A comprehensive survey on the data-driven approaches used for tackling the covid-19 pandemic. WSEAS Transactions on Biology and Biomedicine, 2024; 21:200–217.

13.

Zhang

, Wu

, Qiu

, et al. Intelligent speech technologies for transcription, disease diagnosis, and medical equipment interactive control in smart hospitals: A review. Comput Biol Med, 2023; 153:106517.

14.

Yang

, Li

, Ding

, et al. Deep learning-based speech analysis for Alzheimer’s disease detection: A literature review. Alzheimers Res Ther, 2022; 14(1):186.

15.

Lee

, Seong

, Ozlu

, et al. Biosignal sensors and deep learning-based speech recognition: A review. Sensors (Basel), 2021; 21(4):1399.

16.

Khan

, Kader

, Islam

SMR

, et al. Machine learning and deep learning approaches for brain disease diagnosis: Principles and recent advances. Ieee Access, 2021; 9:37622–37655.

17.

Fernández-Díaz

, Gallardo-Antolín

. An attention Long Short-Term Memory based system for automatic classification of speech intelligibility. Eng Appl Artif Intell, 2020; 96:103976.

18.

Sidi Yakoub

, Selouani

S-A

, Zaidi

B-F

, et al. Improving dysarthric speech recognition using empirical mode decomposition and convolutional neural network. J Audio Speech Music Proc, 2020; 2020(1):1–7.

19.

Sheng

, Kuang

, Bai

, et al. Deep learning for visual speech analysis: A survey. IEEE Trans Pattern Anal Mach Intell, 2024; 46(9):6001–6022.

20.

Chen

, Zhang

, Yang

, et al. Improving cross-lingual low-resource speech recognition by Task-based Meta PolyLoss. Computer Speech & Language, 2024; 87 Article:101648.

21.

Liu

, Yang

, Qu

. Exploration of Whisper fine-tuning strategies for low-resource ASR. J Audio Speech Music PROC, 2024 Article number;2024(1); doi: 10.1186/s13636-024-00349-3

22.

Geng

, et al. Investigation of data augmentation techniques for disordered speech recognition. ArXiv, 2022 preprint arXiv:2201.05562.

23.

Celin

TAM

, Nagarajan

, Vijayalakshmi

. Data augmentation using virtual microphone array synthesis and multi-resolution feature extraction for isolated word dysarthric speech recognition. in IEEE J Sel Top Signal Process, 2020; 14(2):346–354; doi: 10.1109/JSTSP.2020.2972161

24.

Chandrakala

, Malini

, Veni

. Histogram of states based assistive system for speech impairment due to neurological disorders. IEEE Trans Neural Syst Rehabil Eng, 2021; 29:2425–2434; doi: 10.1109/TNSRE.2021.3125314

25.

, Su

, Qian

. Multi-stage audio-visual fusion for dysarthric speech recognition with pre-trained models. in IEEE Trans Neural Syst Rehabil Eng, 2023; 31:1912–1921; doi: 10.1109/TNSRE.2023.3262001

26.

Xiong

, Barker

, Christensen

. Deep learning of articulatory-based representations and applications for improving dysarthric speech recognition, Speech Communication; 13th ITG-Symposium, Oldenburg, Germany, 2018, pp. 1–5.

27.

Mahum

, Ganiyu

, Hidri

, et al. A novel Swin transformer based framework for speech recognition for dysarthria. Sci Rep, 2025; 15(1):20070.

28.

, Phooi Seng

, Ang

. Collaborative AI dysarthric speech recognition system with data augmentation using generative adversarial neural network. IEEE Trans Neural Syst Rehabil Eng, 2025; 33:2097–2111.

29.

Yue

, et al. Multi-modal dysarthric speech recognition based on audio and lip features. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2025; 33:1123–1132.

30.

Peng

, et al. Decoding dysarthric speech using improved conformer with phoneme representation. IEEE/ACM Trans Audio Speech Lang Process, 2025; 33:1784–1796.

31.

Shahamiri

, Mandal

, Sarkar

. Dysarthric speech recognition: An investigation on using depthwise separable convolutions and residual connections. Neural Comput & Applic, 2025; 37(12):7991–8005.

32.

, Xie

, Geng

, et al. Self-supervised ASR models and features for dysarthric and elderly speech recognition. IEEE/ACM Trans Audio Speech Lang Process, 2024; 32:3561–3575.

33.

Mehmood

, et al. An automatic multilingual dysarthric speech detection system based on prosodic and acoustic cues. IEEE Access, 2025; 13:55622–55633.

34.

, et al. Exploring self-supervised pre-trained ASR models for dysarthric and elderly speech recognition. ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023.

35.

Menéndez-Pidal

, Polikoff

, Peters

, et al. The Nemours database of dysarthric speech, in Proc. Intl. Conf. Spoken Language Processing (ICSLP), Philadelphia, PA, USA, 1996, pp. 1962–1965; doi: 10.1109/ICSLP.1996.608020

36.

Rudzicz

, Namasivayam

, Wolff

. The TORGO database of acoustic and articulatory speech from speakers with dysarthria. Lang Resources & Evaluation, 2012; 46(4):523–541.

37.

Kim

, et al. Dysarthric speech database for universal access research, in Proc. 9th Annu. Conf. Int. Speech Commun. Assoc. (INTERSPEECH), Brisbane, QLD, Australia, 2008, pp. 1741–1744.

38.

MacDonald

, et al. Disordered speech data collection: Lessons learned at 1 million utterances from project Euphonia, in Proc. Annu. Conf. Int. Speech Commun. Assoc. (INTERSPEECH), 2021, pp. 4833–4837, doi: 10.21437/Interspeech.2021-697

39.

Liu

, Geng

, Hu

, et al. Recent progress in the CUHK dysarthric speech recognition system. IEEE/ACM Trans Audio Speech Lang Process, 2021; 29:2267–2281; doi: 10.1109/TASLP.2021.3091805