Automatic Speech Recognition Models Research Articles

Adversarial Attack and Defense for Commercial Black-box Chinese-English Speech Recognition Systems

The attacker can generate adversarial examples (AEs) to stealthily mislead automatic speech recognition (ASR) models, raising significant concerns about the security of intelligent voice control (IVC) devices. Existing adversarial attacks mainly generate AEs to mislead ASR models to output specific target English commands (e.g., open the door). However, it remains unknown whether AEs can be used to issue commands in other languages to attack commercial black-box ASR models. In this paper, taking Chinese phrases (e.g., 支付宝付款) and “Chinese-English code-switching” phrases (e.g., 关闭GPS) as the target commands, we propose adversarial attacks for commercial multilingual ASR models. In particular, if a multilingual speech recognition model can recognize Chinese and English, we call it a Chinese-English speech recognition model. In English, the meaning of “支付宝付款” and “关闭GPS” are “Alipay payment” and “turn off GPS”, respectively. In detail, we generate transferable AEs based on the open-sourced conventional DataTang Mandarin ASR model. Given 55 target commands, the success rate for generating AEs of them is up to 96% and 80% for Aliyun ASR API and Tencentyun ASR API, respectively. Our AEs can trigger actual attack actions on voice assistants (e.g., Apple Siri, Xiaomi Xiaoaitongxue) or spread malicious messages through ASR API services, while the target commands in the AEs are inaudible to human beings. Finally, by analyzing the spectrum differences between benign audio clips and AEs, we propose a general defense against adversarial audio attacks.

Sub-layer feature fusion applied to transformer model for automatic speech recognition

Sub-layer feature fusion applied to transformer model for automatic speech recognition

Heterogeneous Hierarchical Fusion Network for Multimodal Sentiment Analysis in Real-World Environments

Multimodal sentiment analysis models can determine users’ sentiments by utilizing rich information from various sources (e.g., textual, visual, and audio). However, there are two key challenges when deploying the model in real-world environments: (1) the limitations of relying on the performance of automatic speech recognition (ASR) models can lead to errors in recognizing sentiment words, which may mislead the sentiment analysis of the textual modality, and (2) variations in information density across modalities complicate the development of a high-quality fusion framework. To address these challenges, this paper proposes a novel Multimodal Sentiment Word Optimization Module and a heterogeneous hierarchical fusion (MSWOHHF) framework. Specifically, the proposed Multimodal Sentiment Word Optimization Module optimizes the sentiment words extracted from the textual modality by the ASR model, thereby reducing sentiment word recognition errors. In the multimodal fusion phase, a heterogeneous hierarchical fusion network architecture is introduced, which first utilizes a Transformer Aggregation Module to fuse the visual and audio modalities, enhancing the high-level semantic features of each modality. A Cross-Attention Fusion Module then integrates the textual modality with the audiovisual fusion. Next, a Feature-Based Attention Fusion Module is proposed that enables fusion by dynamically tuning the weights of both the combined and unimodal representations. It then predicts sentiment polarity using a nonlinear neural network. Finally, the experimental results on the MOSI-SpeechBrain, MOSI-IBM, and MOSI-iFlytek datasets show that the MSWOHHF outperforms several baselines, demonstrating better performance.

Speaker-Attributed Training for Multi-Speaker Speech Recognition Using Multi-Stage Encoders and Attention-Weighted Speaker Embedding

Automatic speech recognition (ASR) aims at understanding naturally spoken human speech to be used as text inputs to machines. In multi-speaker environments, where multiple speakers are talking simultaneously with a large amount of overlap, a significant performance degradation may occur with conventional ASR systems if they are trained by recordings of single talkers. This paper proposes a multi-speaker ASR method that incorporates speaker embedding information as an additional input. The embedding information for each of the speakers in the training set was extracted as numeric vectors, and all of the embedding vectors were stacked to construct a total speaker profile matrix. The speaker profile matrix from the training dataset enables finding embedding vectors that are close to the speakers of the input recordings in the test conditions, and it helps to recognize the individual speakers’ voices mixed in the input. Furthermore, the proposed method efficiently reuses the decoder from the existing speaker-independent ASR model, eliminating the need for retraining the entire system. Various speaker embedding methods such as i-vector, d-vector, and x-vector were adopted, and the experimental results show 0.33% and 0.95% absolute (3.9% and 11.5% relative) improvements without and with the speaker profile in the word error rate (WER).

Effect of simulated hearing loss on automatic speech recognition for an android robot-patient.

The importance of simulating patient behavior for medical assessment training has grown in recent decades due to the increasing variety of simulation tools, including standardized/simulated patients, humanoid and android robot-patients. Yet, there is still a need for improvement of current android robot-patients to accurately simulate patient behavior, among which taking into account their hearing loss is of particular importance. This paper is the first to consider hearing loss simulation in an android robot-patient and its results provide valuable insights for future developments. For this purpose, an open-source dataset of audio data and audiograms from human listeners was used to simulate the effect of hearing loss on an automatic speech recognition (ASR) system. The performance of the system was evaluated in terms of both word error rate (WER) and word information preserved (WIP). Comparing different ASR models commonly used in robotics, it appears that the model size alone is insufficient to predict ASR performance in presence of simulated hearing loss. However, though absolute values of WER and WIP do not predict the intelligibility for human listeners, they do highly correlate with it and thus could be used, for example, to compare the performance of hearing aid algorithms.

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.

Refining maritime Automatic Speech Recognition by leveraging synthetic speech

Maritime transport serves as a critical component of global trade and logistics, enabling the movement of goods and resources across oceans and waterways. Especially in busy waterways and ports, effective and accurate communication is essential, as it ensures the seamless exchange of information and the coordinated execution of port activities. However, comprehensibility is often hindered by factors such as poor audio quality, background noise, and diverse languages and accents. Automatic Speech Recognition (ASR) systems can mitigate these issues by providing real-time transcription and enabling the implementation of automated, value-adding services to enhance situational awareness. While pre-trained ASR models excel on general speech, maritime ASR faces unique challenges due to a lack of annotated data, diverse accents, and specialized terminology.To this end, we focus on improving the transcription quality of pre-trained ASR models for maritime communication with a particular focus on accurately recognizing maritime-specific terminology such as vessel and location names. Due to the scarcity of transcribed maritime communication, we create a synthetic training dataset tailored to regional maritime terminology. The synthetic audio is augmented with general human speech and used to fine-tune an end-to-end ASR model under various settings. The evaluation of the models employs a proprietary dataset of regional maritime radio communication from the port of Hamburg.The experimental results demonstrate a notable enhancement in ASR performance. Specifically, our approach yields an absolute improvement over the pre-trained baseline of 13.46% Word-Error-Rate and an increase of 41.57% recall for vessel names and 38.65% recall for locations. Our findings underscore the efficacy of integrating synthetic training data to address the challenges encountered in maritime ASR, paving the way for more robust and accurate speech recognition systems tailored to maritime applications.

Cascaded cross-modal transformer for audio–textual classification

Speech classification tasks often require powerful language understanding models to grasp useful features, which becomes problematic when limited training data is available. To attain superior classification performance, we propose to harness the inherent value of multimodal representations by transcribing speech using automatic speech recognition models and translating the transcripts into different languages via pretrained translation models. We thus obtain an audio–textual (multimodal) representation for each data sample. Subsequently, we combine language-specific Bidirectional Encoder Representations from Transformers with Wav2Vec2.0 audio features via a novel cascaded cross-modal transformer (CCMT). Our model is based on two cascaded transformer blocks. The first one combines text-specific features from distinct languages, while the second one combines acoustic features with multilingual features previously learned by the first transformer block. We employed our system in the Requests Sub-Challenge of the ACM Multimedia 2023 Computational Paralinguistics Challenge. CCMT was declared the winning solution, obtaining an unweighted average recall of 65.41% and 85.87% for complaint and request detection, respectively. Moreover, we applied our framework on the Speech Commands v2 and HVB dialog data sets, surpassing previous studies reporting results on these benchmarks. Our code is freely available for download at: https://github.com/ristea/ccmt.

Decoupled structure for improved adaptability of end-to-end models

Although end-to-end (E2E) trainable automatic speech recognition (ASR) has shown great success by jointly learning acoustic and linguistic information, it still suffers from the effect of domain shifts, thus limiting potential applications. The E2E ASR model implicitly learns an internal language model (LM) which characterises the training distribution of the source domain, and the E2E trainable nature makes the internal LM difficult to adapt to the target domain with text-only data. To solve this problem, this paper proposes decoupled structures for attention-based encoder–decoder (Decoupled-AED) and neural transducer (Decoupled-Transducer) models, which can achieve flexible domain adaptation in both offline and online scenarios while maintaining robust intra-domain performance. To this end, the acoustic and linguistic parts of the E2E model decoder (or prediction network) are decoupled, making the linguistic component (i.e. internal LM) replaceable. When encountering a domain shift, the internal LM can be directly replaced during inference by a target-domain LM, without re-training or using domain-specific paired speech-text data. Experiments for E2E ASR models trained on the LibriSpeech-100h corpus showed that the proposed decoupled structure gave 15.1% and 17.2% relative word error rate reductions on the TED-LIUM 2 and AESRC2020 corpora while still maintaining performance on intra-domain data. It is also shown that the decoupled structure can be used to boost cross-domain speech translation quality while retaining the intra-domain performance.

Perception of Phonological Assimilation by Neural Speech Recognition Models

Abstract Human listeners effortlessly compensate for phonological changes during speech perception, often unconsciously inferring the intended sounds. For example, listeners infer the underlying /n/ when hearing an utterance such as “clea[m] pan”, where [m] arises from place assimilation to the following labial [p]. This article explores how the neural speech recognition model Wav2Vec2 perceives assimilated sounds, and identifies the linguistic knowledge that is implemented by the model to compensate for assimilation during Automatic Speech Recognition (ASR). Using psycholinguistic stimuli, we systematically analyze how various linguistic context cues influence compensation patterns in the model’s output. Complementing these behavioral experiments, our probing experiments indicate that the model shifts its interpretation of assimilated sounds from their acoustic form to their underlying form in its final layers. Finally, our causal intervention experiments suggest that the model relies on minimal phonological context cues to accomplish this shift. These findings represent a step towards better understanding the similarities and differences in phonological processing between neural ASR models and humans.

Recognition of target domain Japanese speech using language model replacement

End-to-end (E2E) automatic speech recognition (ASR) models, which consist of deep learning models, are able to perform ASR tasks using a single neural network. These models should be trained using a large amount of data; however, collecting speech data which matches the targeted speech domain can be difficult, so speech data is often used that is not an exact match to the target domain, resulting in lower performance. In comparison to speech data, in-domain text data is much easier to obtain. Thus, traditional ASR systems use separately trained language models and HMM-based acoustic models. However, it is difficult to separate language information from an E2E ASR model because the model learns both acoustic and language information in an integrated manner, making it very difficult to create E2E ASR models for specialized target domain which are able to achieve sufficient recognition performance at a reasonable cost. In this paper, we propose a method of replacing the language information within pre-trained E2E ASR models in order to achieve adaptation to a target domain. This is achieved by deleting the “implicit” language information contained within the ASR model by subtracting the source-domain language model trained with a transcription of the ASR’s training data in a logarithmic domain. We then integrate a target domain language model through addition in the logarithmic domain. This subtraction and addition to replace of the language model is based on Bayes’ theorem. In our experiment, we first used two datasets of the Corpus of Spontaneous Japanese (CSJ) to evaluate the effectiveness of our method. We then we evaluated our method using the Japanese Newspaper Article Speech (JNAS) and CSJ corpora, which contain audio data from the read speech and spontaneous speech domain, respectively, to test the effectiveness of our proposed method at bridging the gap between these two language domains. Our results show that our proposed language model replacement method achieved better ASR performance than both non-adapted (baseline) ASR models and ASR models adapted using the conventional Shallow Fusion method.

Enhancing Air Traffic Control Communication Systems with Integrated Automatic Speech Recognition: Models, Applications and Performance Evaluation.

In air traffic control (ATC), speech communication with radio transmission is the primary way to exchange information between the controller and the pilot. As a result, the integration of automatic speech recognition (ASR) systems holds immense potential for reducing controllers' workload and plays a crucial role in various ATC scenarios, which is particularly significant for ATC research. This article provides a comprehensive review of ASR technology's applications in the ATC communication system. Firstly, it offers a comprehensive overview of current research, including ATC corpora, ASR models, evaluation measures and application scenarios. A more comprehensive and accurate evaluation methodology tailored for ATC is proposed, considering advancements in communication sensing systems and deep learning techniques. This methodology helps researchers in enhancing ASR systems and improving the overall performance of ATC systems. Finally, future research recommendations are identified based on the primary challenges and issues. The authors sincerely hope this work will serve as a clear technical roadmap for ASR endeavors within the ATC domain and make a valuable contribution to the research community.

Automatic transcription system for parliamentary debates in the context of assembly of the republic of Portugal

The transcription of parliamentary proceedings is essential for democratic governance. Traditional methods are manual and time-consuming. This work introduces an Automatic Transcription System for the Assembly of the Republic of Portugal (STAAR) that uses an automatic speech recognition model and speaker diarization technologies. STAAR was developed after analyzing existing technologies and the Assembly’s specific needs, leading to an effective solution that integrates with current processes. STAAR stands out for its efficiency in transcribing debates and adapting to parliamentary language nuances. It significantly exceeded expectations by presenting a low transcription error rate, ranging from 1.7 to 11.3%, depending on the context and speech style, reducing the time required to produce the official parliamentary debates journal, and improving overall transcription efficiency. Additionally, STAAR enabled the transcription of previously undocumented parliamentary committee meetings, enhancing the documentation of parliamentary activities. This achievement marks a significant step in modernizing parliamentary processes, increasing transparency and accessibility of political information, and positions the Portuguese Parliament at the forefront of technological innovation in parliamentary debates transcription.

Smart reception: An artificial intelligence driven bangla language based receptionist system employing speech, speaker, and face recognition for automating reception services

In recent times, service robots (SR) have become widely accepted in a variety of fields as an alternative to traditional reception methods. Artificial Intelligence (AI) driven systems are seen as efficient labor alternatives, resulting in several SR models already being developed, primarily designed for English-speaking environments. Despite a large Bangla-speaking population, the exploration and implementation of Bangla language support within AI reception systems remain unexplored due to limited resources, posing significant challenges for deployment. This study presents a novel AI-enabled receptionist framework tailored to Bangla-speaking environments, addressing the limitation of Bangla language resources for developing automated reception systems. Leveraging advanced AI technologies including Face Recognition, Speaker Recognition, Automatic Speech Recognition (ASR), Text-to-Speech Synthesis, and Question Answering System, our integrated system demonstrates promise to automate Bangla reception systems. Our suite of models yields positive performance, with a face recognition accuracy of 99.38%, a speaker recognition system with 5.83 Equal Error Rate (ERR), ASR model achieving a Word Error Rate (WER) of 9.005%, TTS model scoring a Mean Opinion Score (MOS) of 4.10, and a question-answering system with a validation loss of 0.03%. Real-world evaluation among 1664 users in a university setting achieved over 75% user satisfaction, with 88% expressing interest in real-life implementation, showcasing usability across different domains. Limitations such as limited training data, scalability, and environmental sensitivity persist, underscoring the need for further development. Nevertheless, our framework demonstrates potential for real-life implementation, fostering human-robot interaction in Bangla-speaking contexts and paving the way for future innovations in AI-driven Bangla receptionist systems.

Automatic Speech Recognition of Conversational Speech in Individuals With Disordered Speech.

This study examines the effectiveness of automatic speech recognition (ASR) for individuals with speech disorders, addressing the gap in performance between read and conversational ASR. We analyze the factors influencing this disparity and the effect of speech mode-specific training on ASR accuracy. Recordings of read and conversational speech from 27 individuals with various speech disorders were analyzed using both (a) one speaker-independent ASR system trained and optimized for typical speech and (b) multiple ASR models that were personalized to the speech of the participants with disordered speech. Word error rates were calculated for each speech model, read versus conversational, and subject. Linear mixed-effects models were used to assess the impact of speech mode and disorder severity on ASR accuracy. We investigated nine variables, classified as technical, linguistic, or speech impairment factors, for their potential influence on the performance gap. We found a significant performance gap between read and conversational speech in both personalized and unadapted ASR models. Speech impairment severity notably impacted recognition accuracy in unadapted models for both speech modes and in personalized models for read speech. Linguistic attributes of utterances were the most influential on accuracy, though atypical speech characteristics also played a role. Including conversational speech samples in model training notably improved recognition accuracy. We observed a significant performance gap in ASR accuracy between read and conversational speech for individuals with speech disorders. This gap was largely due to the linguistic complexity and unique characteristics of speech disorders in conversational speech. Training personalized ASR models using conversational speech significantly improved recognition accuracy, demonstrating the importance of domain-specific training and highlighting the need for further research into ASR systems capable of handling disordered conversational speech effectively.

End-To-End deep neural models for Automatic Speech Recognition for Polish Language

This article concerns research on deep learning models (DNN) used for automatic speech recognition (ASR). In such systems, recognition is based on Mel Frequency Cepstral Coefficients (MFCC) acoustic features and spectrograms. The latest ASR technologies are based on convolutional neural networks (CNNs), recurrent neural networks (RNNs) and Transformers. The article presents an analysis of modern artificial intelligence algorithms adapted for automatic recognition of the Polish language. The differences between conventional architectures and ASR DNN End-To-End (E2E) models are discussed. Preliminary tests of five selected models (QuartzNet, FastConformer, Wav2Vec 2.0 XLSR, Whisper and ESPnet Model Zoo) on Mozilla Common Voice, Multilingual LibriSpeech and VoxPopuli databases are demonstrated. Tests were conducted for clean audio signal, signal with bandwidth limitation and degraded. The tested models were evaluated on the basis of Word Error Rate (WER).

Advanced accent/dialect identification and accentedness assessment with multi-embedding models and automatic speech recognition.

The ability to accurately classify accents and assess accentedness in non-native speakers are challenging tasks due primarily to the complexity and diversity of accent and dialect variations. In this study, embeddings from advanced pretrained language identification (LID) and speaker identification (SID) models are leveraged to improve the accuracy of accent classification and non-native accentedness assessment. Findings demonstrate that employing pretrained LID and SID models effectively encodes accent/dialect information in speech. Furthermore, the LID and SID encoded accent information complement an end-to-end (E2E) accent identification (AID) model trained from scratch. By incorporating all three embeddings, the proposed multi-embedding AID system achieves superior accuracy in AID. Next, leveraging automatic speech recognition (ASR) and AID models is investigated to explore accentedness estimation. The ASR model is an E2E connectionist temporal classification model trained exclusively with American English (en-US) utterances. The ASR error rate and en-US output of the AID model are leveraged as objective accentedness scores. Evaluation results demonstrate a strong correlation between scores estimated by the two models. Additionally, a robust correlation between objective accentedness scores and subjective scores based on human perception is demonstrated, providing evidence for the reliability and validity of using AID-based and ASR-based systems for accentedness assessment in non-native speech. Such advanced systems would benefit accent assessment in language learning as well as speech and speaker assessment for intelligibility, quality, and speaker diarization and speech recognition advancements.

Automatic speech recognition for Indonesian medical dictation in cloud environment

<span lang="EN-US">This paper introduces SPWPM, an automatic speech recognition (ASR) system designed specifically for Indonesian medical dictation. The main objective of SPWPM is to assist medical professionals in producing medical reports and diagnosing patients. Deployed within a cloud computing service architecture, SPWPM strives to achieve a minimum speech recognition accuracy of 95%. The ASR model of SPWPM is developed using Kaldi and PyChain technologies—creating a comprehensive training dataset involving collaboration with PT Dua-Empat-Tujuh and Harapan Kita Hospital. Several optimization techniques were applied, including language modeling with smoothing, lexicon generation using the Grapheme-to-Phoneme Converter, and data augmentation. The readiness of this technology to assist hospital users was assessed through two evaluations: the SPWPM architecture test and the SPWPM speech recognition test. The results demonstrate the system's preparedness in accurately transcribing medical dictation, showcasing its potential to enhance medical reporting for healthcare professionals in hospital environments.</span>

System dedicated to Polish Automatic Speech Recognition - overview of solutions

The paper presents the analysis of modern Artificial Intelligence algorithms for the automated system supporting human beings during their conversation in Polish language. Their task is to perform Automatic Speech Recognition (ASR) and process it further, for instance fill the computer-based form or perform the Natural Language Processing (NLP) to assign the conversation to one of predefined categories. The State-of-the-Art review is required to select the optimal set of tools to process speech in the difficult conditions, which degrade accuracy of ASR. The paper presents the top-level architecture of the system applicable for the task. Characteristics of Polish language are discussed. Next, existing ASR solutions and architectures with the End-To-End (E2E) deep neural network (DNN) based ASR models are presented in detail. Differences between Recurrent Neural Networks (RNN), Convolutional Neural Networks (CNN) and Transformers in the context of ASR technology are also discussed.

Cluster-Based Pairwise Contrastive Loss for Noise-Robust Speech Recognition.

This paper addresses a joint training approach applied to a pipeline comprising speech enhancement (SE) and automatic speech recognition (ASR) models, where an acoustic tokenizer is included in the pipeline to leverage the linguistic information from the ASR model to the SE model. The acoustic tokenizer takes the outputs of the ASR encoder and provides a pseudo-label through K-means clustering. To transfer the linguistic information, represented by pseudo-labels, from the acoustic tokenizer to the SE model, a cluster-based pairwise contrastive (CBPC) loss function is proposed, which is a self-supervised contrastive loss function, and combined with an information noise contrastive estimation (infoNCE) loss function. This combined loss function prevents the SE model from overfitting to outlier samples and represents the pronunciation variability in samples with the same pseudo-label. The effectiveness of the proposed CBPC loss function is evaluated on a noisy LibriSpeech dataset by measuring both the speech quality scores and the word error rate (WER). The experimental results reveal that the proposed joint training approach using the described CBPC loss function achieves a lower WER than the conventional joint training approaches. In addition, it is demonstrated that the speech quality scores of the SE model trained using the proposed training approach are higher than those of the standalone-SE model and SE models trained using conventional joint training approaches. An ablation study is also conducted to investigate the effects of different combinations of loss functions on the speech quality scores and WER. Here, it is revealed that the proposed CBPC loss function combined with infoNCE contributes to a reduced WER and an increase in most of the speech quality scores.