Phoneme Error Rate Research Articles

A Deep Learning Automatic Speech Recognition Model for Shona Language

This study presented the development of a deep learning-based Automatic Speech Recognition (ASR) system for Shona, a low-resource language characterized by unique tonal and grammatical complexities. The research aimed to address the challenges posed by limited training data, a lack of labelled data, and the intricate tonal nuances present in Shona speech, with the objective of achieving significant improvements in recognition accuracy compared to traditional statistical models. Motivated by the limitations of existing approaches, the research addressed three key questions. Firstly, it explored the feasibility of using deep learning to develop an accurate ASR system for Shona. Secondly, it investigated the specific challenges involved in designing and implementing deep learning architectures for Shona speech recognition and proposed strategies to mitigate these challenges. Lastly, it compared the performance of the deep learning-based model with existing statistical models in terms of accuracy. The developed ASR system utilized a hybrid architecture consisting of a Convolutional Neural Network (CNN) for acoustic modelling and a Long Short-Term Memory (LSTM) network for language modelling. To overcome the scarcity of data, the research employed data augmentation techniques and transfer learning. Attention mechanisms were also incorporated to accommodate the tonal nature of Shona speech. The resulting ASR system achieved impressive results, with a Word Error Rate (WER) of 29%, Phoneme Error Rate (PER) of 12%, and an overall accuracy of 74%. These metrics indicated a significant improvement over existing statistical models, highlighting the potential of deep learning to enhance ASR accuracy for under-resourced languages like Shona. This research contributed to the advancement of ASR technology for under-resourced languages like Shona, ultimately fostering improved accessibility and communication for Shona speakers worldwide.

Automatic speech recognition (ASR) for the diagnosis of pronunciation of speech sound disorders in Korean children.

This study presents a model of automatic speech recognition (ASR) that is designed to diagnose pronunciation issues in children with speech sound disorders (SSDs) to replace manual transcriptions in clinical procedures. Because ASR models trained for general purposes mainly predict input speech into standard spelling words, well-known high-performance ASR models are not suitable for evaluating pronunciation in children with SSDs. We fine-tuned the wav2vec2.0 XLS-R model to recognise words as they are pronounced by children, rather than converting the speech into their standard spelling words. The model was fine-tuned with a speech dataset of 137 children with SSDs pronouncing 73 Korean words that are selected for actual clinical diagnosis. The model's Phoneme Error Rate (PER) was only 10% when its predictions of children's pronunciations were compared to human annotations of pronunciations as heard. In contrast, despite its robust performance on general tasks, the state-of-the-art ASR model Whisper showed limitations in recognising the speech of children with SSDs, with a PER of approximately 50%. While the model still requires improvement in terms of the recognition of unclear pronunciation, this study demonstrates that ASR models can streamline complex pronunciation error diagnostic procedures in clinical fields.

Pronunciation error detection model based on feature fusion

Mispronunciation detection and diagnosis (MDD) is a specific speech recognition task that aims to recognize the phoneme sequence produced by a user, compare it with the standard phoneme sequence, and identify the type and location of any mispronunciations. However, the lack of large amounts of phoneme-level annotated data limits the performance improvement of the model. In this paper, we propose a joint training approach, Acoustic Error_Type Linguistic (AEL) that utilizes the error type information, acoustic information, and linguistic information from the annotated data, and achieves feature fusion through multiple attention mechanisms. To address the issue of uneven distribution of phonemes in the MDD data, which can cause the model to make overconfident predictions when using the CTC loss, we propose a new loss function, Focal Attention Loss, to improve the performance of the model, such as F1 score accuracy and other metrics. The proposed method in this paper was evaluated on the TIMIT and L2-Arctic public corpora. In ideal conditions, it was compared with the baseline model CNN-RNN-CTC. The F1 score, diagnostic accuracy, and precision were improved by 31.24%, 16.6%, and 17.35% respectively. Compared to the baseline model, our model reduced the phoneme error rate from 29.55% to 8.49% and showed significant improvements in other metrics. Furthermore, experimental results demonstrated that when we have a model capable of accurately obtaining pronunciation error types, our model can achieve results close to the ideal conditions.

Conditional Random Fields Applied to Arabic Orthographic-Phonetic Transcription

Orthographic-To-Phonetic (O2P) Transcription is the process of learning the relationship between the written word and its phonetic transcription. It is a necessary part of Text-To-Speech (TTS) systems and it plays an important role in handling Out-Of-Vocabulary (OOV) words in Automatic Speech Recognition systems. The O2P is a complex task, because for many languages, the correspondence between the orthography and its phonetic transcription is not completely consistent. Over time, the techniques used to tackle this problem have evolved, from earlier rules based systems to the current more sophisticated machine learning approaches. In this paper, we propose an approach for Arabic O2P Conversion based on a probabilistic method: Conditional Random Fields (CRF). We discuss the results and experiments of this method apply on a pronunciation dictionary of the Most Commonly used Arabic Words, a database that we called (MCAW-Dic). MCAW-Dic contains over 35 000 words in Modern Standard Arabic (MSA) and their pronunciation, a database that we have developed by ourselves assisted by phoneticians and linguists from the University of Tlemcen. The results achieved are very satisfactory and point the way towards future innovations. Indeed, in all our tests, the score was between 11 and 15% error rate on the transcription of phonemes (Phoneme Error Rate). We could improve this result by including a large context, but in this case, we encountered memory limitations and calculation difficulties.

Disordered speech recognition considering low resources and abnormal articulation

The success of automatic speech recognition (ASR) benefits a great number of healthy people, but not people with disorders. The speech disordered may truly need support from technology, while they actually gain little. The difficulties of disordered ASR arise from the limited availability of data and the abnormal nature of speech, e.g, unclear, unstable, and incorrect pronunciations. To realize the ASR of disordered speech, this study addresses the problems of disordered speech in two respects, low resources, and articulatory abnormality. In order to solve the problem of low resources, this study proposes staged knowledge distillation (KD), which provides different references to the student models according to their mastery of knowledge, so as to avoid feature overfitting. To tackle the articulatory abnormalities in dysarthria, we propose an intended phonological perception method (IPPM) by applying the motor theory of speech perception to ASR, in which pieces of intended phonological features are estimated and provided to ASR. And further, we solve the challenges of disordered ASR by combining the staged KD and the IPPM. TORGO database and UASEECH corpus are two commonly used datasets of dysarthria which is the main cause of speech disorders. Experiments on the two datasets validated the effectiveness of the proposed methods. Compared with the baseline, the proposed method achieves 35.14%∼38.12% relative phoneme error rate reductions (PERRs) for speakers with varying degrees of dysarthria on the TORGO database and relative 8.17%∼13.00% PERRs on the UASPEECH corpus. The experiments demonstrated that addressing disordered speech from both low resources and speech abnormality is an effective way to solve the problems, and the proposed methods significantly improved the performance of ASR for disordered speech.

Silent speech recognition using data augmentation based on a three-dimensional lip model

In this study, we proposed a data augmentation method using a three-dimensional model of a speaker's face for machine lip reading, which estimates the content of speech without using speech data. The proposed method converted a speaker's face into a three-dimensional model using DECA (Detailed Expression Capture and Animation), and generated a large amount of learning data by rotating the three-dimensional model in different directions. Then we obtained the facial features around the lips as time series data using dlib, which is an OpenCV module. The time series data is fed into a recognition model. We introduced end-to-end recognition model which is used in continuous speech recognition. The recognition model is based on DeepSpeech2, and we modified the filter size of CNN (convolutional neural network) and the number of layers of BiGRU (bidirectional gated recurrent unit) to adapt to our goal. We evaluated the proposed method on ten ordinary Japanese words. We used the phoneme error rate as the evaluation index. As a result of the evaluation, the error rate of about 0.32 was achieved even for data where the speaker of the evaluation data was not included in the speaker of the training data.

Complex Dynamic Neurons Improved Spiking Transformer Network for Efficient Automatic Speech Recognition

The spiking neural network (SNN) using leaky-integrated-and-fire (LIF) neurons has been commonly used in automatic speech recognition (ASR) tasks. However, the LIF neuron is still relatively simple compared to that in the biological brain. Further research on more types of neurons with different scales of neuronal dynamics is necessary. Here we introduce four types of neuronal dynamics to post-process the sequential patterns generated from the spiking transformer to get the complex dynamic neuron improved spiking transformer neural network (DyTr-SNN). We found that the DyTr-SNN could handle the non-toy automatic speech recognition task well, representing a lower phoneme error rate, lower computational cost, and higher robustness. These results indicate that the further cooperation of SNNs and neural dynamics at the neuron and network scales might have much in store for the future, especially on the ASR tasks.

MagTrack: A Wearable Tongue Motion Tracking System for Silent Speech Interfaces.

Current electromagnetic tongue tracking devices are not amenable for daily use and thus not suitable for silent speech interface and other applications. We have recently developed MagTrack, a novel wearable electromagnetic articulograph tongue tracking device. This study aimed to validate MagTrack for potential silent speech interface applications. We conducted two experiments: (a) classification of eight isolated vowels in consonant-vowel-consonant form and (b) continuous silent speech recognition. In these experiments, we used data from healthy adult speakers collected with MagTrack. The performance of vowel classification was measured by accuracies. The continuous silent speech recognition was measured by phoneme error rates. The performance was then compared with results using data collected with commercial electromagnetic articulograph in a prior study. The isolated vowel classification using MagTrack achieved an average accuracy of 89.74% when leveraging all MagTrack signals (x, y, z coordinates; orientation; and magnetic signals), which outperformed the accuracy using commercial electromagnetic articulograph data (only y, z coordinates) in our previous study. The continuous speech recognition from two subjects using MagTrack achieved phoneme error rates of 73.92% and 66.73%, respectively. The commercial electromagnetic articulograph achieved 64.53% from the same subject (66.73% using MagTrack data). MagTrack showed comparable results with the commercial electromagnetic articulograph when using the same localized information. Adding raw magnetic signals would improve the performance of MagTrack. Our preliminary testing demonstrated the potential for silent speech interface as a lightweight wearable device. This work also lays the foundation to support MagTrack's potential for other applications including visual feedback-based speech therapy and second language learning.

Machine learning for Arabic phonemes recognition using electrolarynx speech

<p><span lang="EN-US">Automatic speech recognition system is one of the essential ways of interaction with machines. Interests in speech based intelligent systems have grown in the past few decades. Therefore, there is a need to develop more efficient methods for human speech recognition to ensure the reliability of communication between individuals and machines. This paper is concerned with Arabic phoneme recognition of electrolarynx device. Electrolarynx is a device used by cancer patients having vocal laryngeal cords removed. Speech recognition here is considered to find the preferred machine learning model that can classify phonemes produced by electrolarynx device. The phonemes recognition employs different machine learning schemes, including convolutional neural network, recurrent neural network, artificial neural network (ANN), random forest, extreme gradient boosting (XGBoost), and long short-term memory. Modern standard Arabic is utilized for testing and training phases of the recognition system. The dataset covers both an ordinary speech and electrolarynx device speech recorded by the same person. Mel frequency cepstral coefficients are considered as speech features. The results show that the ANN machine learning method outperformed other methods with an accuracy rate of 75%, a precision value of 77%, and a phoneme error rate (PER) of 21.85%.</span></p>

Mispronunciation Detection and Diagnosis with Articulatory-Level Feedback Generation for Non-Native Arabic Speech

A high-performance versatile computer-assisted pronunciation training (CAPT) system that provides the learner immediate feedback as to whether their pronunciation is correct is very helpful in learning correct pronunciation and allows learners to practice this at any time and with unlimited repetitions, without the presence of an instructor. In this paper, we propose deep learning-based techniques to build a high-performance versatile CAPT system for mispronunciation detection and diagnosis (MDD) and articulatory feedback generation for non-native Arabic learners. The proposed system can locate the error in pronunciation, recognize the mispronounced phonemes, and detect the corresponding articulatory features (AFs), not only in words but even in sentences. We formulate the recognition of phonemes and corresponding AFs as a multi-label object recognition problem, where the objects are the phonemes and their AFs in a spectral image. Moreover, we investigate the use of cutting-edge neural text-to-speech (TTS) technology to generate a new corpus of high-quality speech from predefined text that has the most common substitution errors among Arabic learners. The proposed model and its various enhanced versions achieved excellent results. We compared the performance of the different proposed models with the state-of-the-art end-to-end technique of MDD, and our system had a better performance. In addition, we proposed using fusion between the proposed model and the end-to-end model and obtained a better performance. Our best model achieved a 3.83% phoneme error rate (PER) in the phoneme recognition task, a 70.53% F1-score in the MDD task, and a detection error rate (DER) of 2.6% for the AF detection task.

Parallel Gated Recurrent Unit Networks as an Encoder for Speech Recognition

Listen, Attend and Spell (LAS) network is one of the end-to-end approaches for speech recognition, which does not require an explicit language model. It consists of two parts; the encoder part which receives acoustic features as inputs, and the decoder network which produces one character at a time step, based on the encoder output and an attention mechanism. Multi-layer recurrent neural networks (RNN) are used in both decoder and encoder parts. Hence, the LAS architecture can be simplified as one RNN for the decoder, and another RNN for the encoder. Their shapes and layer sizes can be different. In this work, we examined the performance of using multi RNNs for the encoder part. Our baseline LAS network uses an RNN with a hidden size of 256. We used 2 and 4 RNNs with hidden sizes of 128 and 64 for each case. The main idea behind the proposed approach is to focus the RNNs to different patterns (phonemes in this case) in the data. At the output of the encoder, their outputs are concatenated and fed to the decoder. TIMIT database is used to compare the performance of the mentioned networks, using phoneme error rate as the performance metric. The experimental results showed that proposed approach can achieve a better performance than the baseline network. However, increasing the number of RNNs does not guarantee further improvements.

Identifying the influence of transfer learning method in developing an end-to-end automatic speech recognition system with a low data level

Ensuring the best quality and performance of modern speech technologies, today, is possible based on the widespread use of machine learning methods. The idea of this project is to study and implement an end-to-end system of automatic speech recognition using machine learning methods, as well as to develop new mathematical models and algorithms for solving the problem of automatic speech recognition for agglutinative (Turkic) languages. Many research papers have shown that deep learning methods make it easier to train automatic speech recognition systems that use an end-to-end approach. This method can also train an automatic speech recognition system directly, that is, without manual work with raw signals. Despite the good recognition quality, this model has some drawbacks. These disadvantages are based on the need for a large amount of data for training. This is a serious problem for low-data languages, especially Turkic languages such as Kazakh and Azerbaijani. To solve this problem, various methods are needed to apply. Some methods are used for end-to-end speech recognition of languages belonging to the group of languages of the same family (agglutinative languages). Method for low-resource languages is transfer learning, and for large resources – multi-task learning. To increase efficiency and quickly solve the problem associated with a limited resource, transfer learning was used for the end-to-end model. The transfer learning method helped to fit a model trained on the Kazakh dataset to the Azerbaijani dataset. Thereby, two language corpora were trained simultaneously. Conducted experiments with two corpora show that transfer learning can reduce the symbol error rate, phoneme error rate (PER), by 14.23 % compared to baseline models (DNN+HMM, WaveNet, and CNC+LM). Therefore, the realized model with the transfer method can be used to recognize other low-resource languages.

Stemming for Better Indonesian Text-to-Phoneme

Stemming, the process of getting a root from a derivative word by removing the affixes (which can be prefixes, infixes, and suffixes), plays an important role in computational linguistics. This paper investigates the use of a stemmer to boost a nearest neighbor-based text-to-phoneme (TTP) or phonemicization for Indonesian language, which has some problems regarding derivatives with prefixes ⟨ber⟩, ⟨me⟩, ⟨per⟩, and ⟨ter⟩ since the graphene ⟨e⟩ is sometimes incorrectly phonemicized. The confix-stripping-based stemming (CS) is added as the preprocessing procedure for a fuzzy nearest neighbor-based phonemicization. It is then investigated to see its impact in solving such errors. The 5-fold cross-validation on 50k words indicates that it is capable of eliminating most errors produced by the Indonesian phonemicization caused by the four prefixes. The averaged phoneme error rate (PER) is decreased from 1.34% to 1.04%. A deep investigation informs that CS can solve some errors from roots and derivatives consisting of graphene sequences ⟨ber⟩, ⟨me⟩, and ⟨per⟩. But, it still frequently fails to solve the errors caused by some words containing a graphene sequence ⟨ter⟩. Compared to a state-of-the-art deep learning-based phonemicization system that exploits a Transformer network, the proposed model gives lower PER and complexity.

Vau Da Muntanialas: Energy-Efficient Multi-Die Scalable Acceleration of RNN Inference

Recurrent neural networks such as Long Short-Term Memories (LSTMs) learn temporal dependencies by keeping an internal state, making them ideal for time-series problems such as speech recognition. However, the output-to-input feedback creates distinctive memory bandwidth and scalability challenges in designing accelerators for RNNs. We present Muntaniala, an RNN accelerator architecture for LSTM inference with a silicon-measured energy-efficiency of 3.25$TOP/s/W$ and performance of 30.53$GOP/s$ in UMC 65 $nm$ technology. The scalable design of Muntaniala allows running large RNN models by combining multiple tiles in a systolic array. We keep all parameters stationary on every die in the array, drastically reducing the I/O communication to only loading new features and sharing partial results with other dies. For quantifying the overall system power, including I/O power, we built Vau da Muntanialas, to the best of our knowledge, the first demonstration of a systolic multi-chip-on-PCB array of RNN accelerator. Our multi-die prototype performs LSTM inference with 192 hidden states in 330$\mu s$ with a total system power of 9.0$mW$ at 10$MHz$ consuming 2.95$\mu J$. Targeting the 8/16-bit quantization implemented in Muntaniala, we show a phoneme error rate (PER) drop of approximately 3% with respect to floating-point (FP) on a 3L-384NH-123NI LSTM network on the TIMIT dataset.

Mlphon: A Multifunctional Grapheme-Phoneme Conversion Tool Using Finite State Transducers

In this article we present the design and the development of a knowledge based computational linguistic tool, Mlphon [em.el.foːɳ] for Malayalam language. Mlphon computationally models linguistic rules using finite state transducers and performs multiple functions including grapheme to phoneme (g2p) and phoneme to grapheme (p2g) conversions, syllabification, phonetic feature analysis and script grammar check. This open source software tool, released under MIT license, is developed as a one-stop solution to handle different speech related text processing tasks for automatic speech recognition, text to speech synthesis and non-speech natural language processing tasks including syllable subword based language modeling, phoneme diversity analysis and text sanity check. The tool is evaluated on a manually crafted gold standard lexicon. Mlphon performs orthographic syllabification with 99% accuracy with a syllable error rate of 0.62% on the gold standard lexicon. For grapheme to phoneme conversion task, overall phoneme recognition accuracy of 99% with a phoneme error rate of 0.55% is obtained on gold standard lexicon. Additionally an extrinsic evaluation of Mlphon is performed by employing the pronunciation lexicon created using Mlphon, in Malayalam automatic speech recognition (ASR) task. Performance analysis in terms of the computation time of lexicon creation process and the word error rate (WER) on ASR task are presented along with a comparison over other automated tools for lexicon creation. Pronunciation lexicons with more than 100k commonly used Malayalam words in phonemised and syllabified forms is created and they are published as open language resources along with this work. We also demonstrate the usage of Mlphon on different natural language processing applications - syllable subword ASR, assisted pronunciation learning, phoneme diversity analysis and text sanity check. Being a knowledge based solution with open source code, Mlphon can be adapted to other languages of similar script nature.

The use of speech recognition technology by people living with amyotrophic lateral sclerosis: a scoping review

Background More than 80% of people living with Amyotrophic Lateral Sclerosis (plwALS) develop difficulties with their speech, affecting communication, self-identity and quality of life. Automatic speech recognition technology (ASR) is becoming a common way to interact with a broad range of devices, to find information and control the environment. ASR can be problematic for people with acquired neurogenic motor speech difficulties (dysarthria). Given that the field is rapidly developing, a scoping review is warranted. Aims This study undertakes a scoping review on the use of ASR technology by plwALS and identifies research gaps in the existing literature. Materials and Methods Electronic databases and relevant grey literature were searched from 1990 to 2020. Eleven research papers and articles were identified that included participants living with ALS using ASR technology. Relevant data were extracted from the included sources, and a narrative summary of the findings presented. Outcomes and Results: Eleven publications used recordings of plwALS to assess word recognition rate (WRR) word error rate (WER) or phoneme error rate (PER) and appropriacy of responses by ASR devices. All were found to be linked to severity of dysarthria and the ASR technology used. One article examined how speech modification may improve ASR accuracy. The final article completed thematic analysis of Amazon.com reviews for the Amazon Echo and plwALS were reported to use ASR devices to control the environment and summon assistance. Conclusions There are gaps in the evidence base: understanding expectations of plwALS and how they use ASR technology; how WER/PER/WRR relates to usability; how ASR use changes as ALS progresses. Implications for rehabilitation Devices that people can interact with using speech are becoming ubiquitous. As movement and mobility are likely to be affected by ALS and progress over time, speech interaction could be very helpful for accessing information and environmental control. However, many people living with ALS (plwALS) also have impaired speech (dysarthria) and experience trouble using voice interaction technology because it may not understand them. Although advances in automated speech recognition (ASR) technology promise better understanding of dysarthric speech, future research needs to investigate how plwALS use ASR, how accurate it needs to be to be functionally useful, and how useful it may be over time as the disease progresses.

Presentation Attack Detection on Limited-Resource Devices Using Deep Neural Classifiers Trained on Consistent Spectrogram Fragments.

The presented paper is concerned with detection of presentation attacks against unsupervised remote biometric speaker verification, using a well-known challenge–response scheme. We propose a novel approach to convolutional phoneme classifier training, which ensures high phoneme recognition accuracy even for significantly simplified network architectures, thus enabling efficient utterance verification on resource-limited hardware, such as mobile phones or embedded devices. We consider Deep Convolutional Neural Networks operating on windows of speech Mel-Spectrograms as a means for phoneme recognition, and we show that one can boost the performance of highly simplified neural architectures by modifying the principle underlying training set construction. Instead of generating training examples by slicing spectrograms using a sliding window, as it is commonly done, we propose to maximize the consistency of phoneme-related spectrogram structures that are to be learned, by choosing only spectrogram chunks from the central regions of phoneme articulation intervals. This approach enables better utilization of the limited capacity of the considered simplified networks, as it significantly reduces a within-class data scatter. We show that neural architectures comprising as few as dozens of thousands parameters can successfully—with accuracy of up to 76%, solve the 39-phoneme recognition task (we use the English language TIMIT database for experimental verification of the method). We also show that ensembling of simple classifiers, using a basic bagging method, boosts the recognition accuracy by another 2–3%, offering Phoneme Error Rates at the level of 23%, which approaches the accuracy of the state-of-the-art deep neural architectures that are one to two orders of magnitude more complex than the proposed solution. This, in turn, enables executing reliable presentation attack detection, based on just few-syllable long challenges on highly resource-limited computing hardware.

Automated grapheme-to-phoneme conversion for Central Kurdish based on optimality theory

The writing system of Central Kurdish features three cases in which there is no one-to-one mapping between the orthographical letters and the phonemes of the language. Consequently, the written words including these cases may be pronounced in multiple ways. The process of finding the correct pronunciation of written words is called Grapheme-to-Phoneme (G2P) conversion and is a key step in natural language processing tasks such as speech synthesis. As Central Kurdish is a low-resourced language, we present a G2P conversion method based on the phonological rules of the language, rather than pronunciation dictionaries and data-driven learning methods. After reviewing the phonology and alphabet of the language through the framework of Optimality Theory, we generate all possible pronunciations. Then, by specifying and applying ranked constraints, we eliminate undesirable candidates so as to keep only one well-formed pronunciation per word. The evaluation of our proposed method on two datasets resulted in 0.75% of overall Phoneme Error Rate (PER) and achieved 94.71% precision in the detection of the short vowel /i/ and 100% of accuracy in the conversion of the letters “ی” and “و”. Analyzing these results suggests that there is no need for additional new letters in the current orthographic system of Central Kurdish. This approach also enables us to have a ranked suggestion list for the manual checking of the few unresolved ambiguous situations.

Generalizing neural signal-to-text brain-computer interfaces

Objective: Brain-Computer Interfaces (BCI) may help patients with faltering communication abilities due to neurodegenerative diseases produce text or speech by direct neural processing. However, their practical realization has proven difficult due to limitations in speed, accuracy, and generalizability of existing interfaces. The goal of this study is to evaluate the BCI performance of a robust speech decoding system that translates neural signals evoked by speech to a textual output. While previous studies have approached this problem by using neural signals to choose from a limited set of possible words, we employ a more general model that can type any word from a large corpus of English text. Approach: In this study, we create an end-to-end BCI that translates neural signals associated with overt speech into text output. Our decoding system first isolates frequency bands in the input depth-electrode signal encapsulating differential information regarding production of various phonemic classes. These bands form a feature set that then feeds into a Long Short-Term Memory (LSTM) model which discerns at each time point probability distributions across all phonemes uttered by a subject. Finally, a particle filtering algorithm temporally smooths these probabilities by incorporating prior knowledge of the English language to output text corresponding to the decoded word. The generalizability of our decoder is driven by the lack of a vocabulary constraint on this output word. Main result: This method was evaluated using a dataset of 6 neurosurgical patients implanted with intra-cranial depth electrodes to identify seizure foci for potential surgical treatment of epilepsy. We averaged 32% word accuracy and on the phoneme-level obtained 46% precision, 51% recall and 73.32% average phoneme error rate while also achieving significant increases in speed when compared to several other BCI approaches. Significance: Our study employs a more general neural signal-to-text model which could facilitate communication by patients in everyday environments.

Improving phoneme recognition of throat microphone speech recordings using transfer learning

Throat microphones (TM) are a type of skin-attached non-acoustic sensors, which are robust to environmental noise but carry a lower signal bandwidth characterization than the traditional close-talk microphones (CM). Attaining high-performance phoneme recognition is a challenging task when the training data from a degrading channel, such as TM, is limited. In this paper, we address this challenge for the TM speech recordings using a transfer learning approach based on the stacked denoising auto-encoders (SDA). The proposed transfer learning approach defines an SDA-based domain adaptation framework to map the source domain CM representations and the target domain TM representations into a common latent space, where the mismatch across TM and CM is eliminated to better train an acoustic model and to improve the TM phoneme recognition. For the phoneme recognition task, we use the convolutional neural network (CNN) and the hidden Markov model (HMM) based CNN/HMM hybrid system, which delivers better acoustic modeling performance compared to the conventional Gaussian mixture model (GMM) based models. In the experimental evaluations, we observed more than 12% relative phoneme error rate (PER) improvement for the TM recordings with the proposed transfer learning approach compared to baseline performances.