Speech-centric machine learning systems have revolutionized many leading domains ranging from transportation and healthcare to education and defense, profoundly changing how people live, work, and interact with each other. However, recent studies have demonstrated that many speech-centric ML systems may need to be considered more trustworthy for broader deployment. Specifically, concerns over privacy breaches, discriminating performance, and vulnerability to adversarial attacks have all been discovered in ML research fields. In order to address the above challenges and risks, a significant number of efforts have been made to ensure these ML systems are trustworthy, especially private, safe, and fair. In this paper, we conduct the first comprehensive survey on speech-centric trustworthy ML topics related to privacy, safety, and fairness. In addition to serving as a summary report for the research community, we point out several promising future research directions to inspire the researchers who wish to explore further in this area.
We present a method to remove unknown convolutive noise introduced to speech by reverberations of recording environments, utilizing some amount of training speech data from the reverberant environment, and any available non-reverberant speech data. Using Fourier transform computed over long temporal windows, which ideally cover the entire room impulse response, we convert room induced convolution to additions in the log spectral domain. Next, we compute a spectral normalization vector from statistics gathered over reverberated as well as over clean speech in the log spectral domain. During operation, this normalization vectors are used to alleviate reverberations from complex speech spectra recorded under the same reverberant conditions . Such dereverberated complex speech spectra are used to compute complex FDLP-spectrograms for use in automatic speech recognition.
Transformers are among the state of the art for many tasks in speech, vision, and natural language processing, among others. Self-attentions, which are crucial contributors to this performance have quadratic computational complexity, which makes training on longer input sequences challenging. Prior work has produced state-of-the-art transformer variants with linear attention, however, current models sacrifice performance to achieve efficient implementations. In this work, we develop a novel linear transformer by examining the properties of the key-query product within self-attentions. Our model outperforms state of the art approaches on speech recognition and speech summarization, resulting in 1 % absolute WER improvement on the Librispeech-100 speech recognition benchmark and a new INTERVIEW speech recognition benchmark, and 5 points on ROUGE for summarization with How2.
We present MooseNet, a trainable speech metric that predicts listeners' Mean Opinion Score (MOS). We report improvements to the challenge baselines using easy-to-use modeling techniques, which also scales for larger self-supervised learning (SSL) model. We present two models. The first model is a Neural Network (NN). As a second model, we propose a PLDA generative model on the top layers of the first NN model, which improves the pure NN model. Ensembles from our two models achieve the top 3 or 4 VoiceMOS leaderboard places on all system and utterance level metrics.
With the rapid advancement in synthetic speech generation technologies, great interest in differentiating spoof speech from the natural speech is emerging in the research community. The identification of these synthetic signals is a difficult task not only for the cutting-edge classification models but also for humans themselves. To prevent potential adverse effects, it becomes crucial to detect spoof signals. From a forensics perspective, it is also important to predict the algorithm which generated them to identify the forger. This needs an understanding of the underlying attributes of spoof signals which serve as a signature for the synthesizer. This study emphasizes the segments of speech signals critical in identifying their authenticity by utilizing the Vocal Tract System(\textit{VTS}) and Voice Source(\textit{VS}) features. In this paper, we propose a system that detects spoof signals as well as identifies the corresponding speech-generating algorithm. We achieve 99.58\% in algorithm classification accuracy. From experiments, we found that a VS feature-based system gives more attention to the transition of phonemes, while, a VTS feature-based system gives more attention to stationary segments of speech signals. We perform model fusion techniques on the VS-based and VTS-based systems to exploit the complementary information to develop a robust classifier. Upon analyzing the confusion plots we found that WaveRNN is poorly classified depicting more naturalness. On the other hand, we identified that synthesizer like Waveform Concatenation, and Neural Source Filter is classified with the highest accuracy. Practical implications of this work can aid researchers from both forensics (leverage artifacts) and the speech communities (mitigate artifacts).
This paper proposes an approach to build a high-quality text-to-speech (TTS) system for technical domains using data augmentation. An end-to-end (E2E) system is trained on hidden Markov model (HMM) based synthesized speech and further fine-tuned with studio-recorded TTS data to improve the timbre of the synthesized voice. The motivation behind the work is that issues of word skips and repetitions are usually absent in HMM systems due to their ability to model the duration distribution of phonemes accurately. Context-dependent pentaphone modeling, along with tree-based clustering and state-tying, takes care of unseen context and out-of-vocabulary words. A language model is also employed to reduce synthesis errors further. Subjective evaluations indicate that speech produced using the proposed system is superior to the baseline E2E synthesis approach in terms of intelligibility when combining complementing attributes from HMM and E2E frameworks. The further analysis highlights the proposed approach's efficacy in low-resource scenarios.
Several trade-offs need to be balanced when employing monaural speech separation (SS) models in conversational automatic speech recognition (ASR) systems. A larger SS model generally achieves better output quality at an expense of higher computation, meanwhile, a better SS model for overlapping speech often produces distorted output for non-overlapping speech. This paper addresses these trade-offs with a sparsely-gated mixture-of-experts (MoE). The sparsely-gated MoE architecture allows the separation models to be enlarged without compromising the run-time efficiency, which also helps achieve a better separation-distortion trade-off. To further reduce the speech distortion without compromising the SS capability, a multi-gate MoE framework is also explored, where different gates handle non-overlapping and overlapping frames differently. ASR experiments are conducted by using a simulated dataset for measuring both the speech separation accuracy and the speech distortion. Two advanced SS models, Conformer and WavLM-based models, are used as baselines. The sparsely-gated MoE models show a superior SS capability with less speech distortion, meanwhile marginally increasing the run-time computational cost. Experimental results using real conversation recordings are also presented, showing MoE's effectiveness in an end-to-end evaluation setting.
In this paper, we describe the systems developed by the SJTU X-LANCE team for LIMMITS 2023 Challenge, and we mainly focus on the winning system on naturalness for track 1. The aim of this challenge is to build a multi-speaker multi-lingual text-to-speech (TTS) system for Marathi, Hindi and Telugu. Each of the languages has a male and a female speaker in the given dataset. In track 1, only 5 hours data from each speaker can be selected to train the TTS model. Our system is based on the recently proposed VQTTS that utilizes VQ acoustic feature rather than mel-spectrogram. We introduce additional speaker embeddings and language embeddings to VQTTS for controlling the speaker and language information. In the cross-lingual evaluations where we need to synthesize speech in a cross-lingual speaker's voice, we provide a native speaker's embedding to the acoustic model and the target speaker's embedding to the vocoder. In the subjective MOS listening test on naturalness, our system achieves 4.77 which ranks first.
This paper presents a source free domain adaptation method for steady-state visually evoked potential (SSVEP) based brain-computer interface (BCI) spellers. SSVEP-based BCI spellers help individuals experiencing speech difficulties, enabling them to communicate at a fast rate. However, achieving a high information transfer rate (ITR) in the current methods requires an extensive calibration period before using the system, leading to discomfort for new users. We address this issue by proposing a method that adapts the deep neural network (DNN) pre-trained on data from source domains (participants of previous experiments conducted for labeled data collection), using only the unlabeled data of the new user (target domain). This adaptation is achieved by minimizing our proposed custom loss function composed of self-adaptation and local-regularity loss terms. The self-adaptation term uses the pseudo-label strategy, while the novel local-regularity term exploits the data structure and forces the DNN to assign the same labels to adjacent instances. Our method achieves striking 201.15 bits/min and 145.02 bits/min ITRs on the benchmark and BETA datasets, respectively, and outperforms the state-of-the-art alternative techniques. Our approach alleviates user discomfort and shows excellent identification performance, so it would potentially contribute to the broader application of SSVEP-based BCI systems in everyday life.
Automatic speech recognition research focuses on training and evaluating on static datasets. Yet, as speech models are increasingly deployed on personal devices, such models encounter user-specific distributional shifts. To simulate this real-world scenario, we introduce LibriContinual, a continual learning benchmark for speaker-specific domain adaptation derived from LibriVox audiobooks, with data corresponding to 118 individual speakers and 6 train splits per speaker of different sizes. Additionally, current speech recognition models and continual learning algorithms are not optimized to be compute-efficient. We adapt a general-purpose training algorithm NetAug for ASR and create a novel Conformer variant called the DisConformer (Disentangled Conformer). This algorithm produces ASR models consisting of a frozen 'core' network for general-purpose use and several tunable 'augment' networks for speaker-specific tuning. Using such models, we propose a novel compute-efficient continual learning algorithm called DisentangledCL. Our experiments show that the DisConformer models significantly outperform baselines on general ASR i.e. LibriSpeech (15.58% rel. WER on test-other). On speaker-specific LibriContinual they significantly outperform trainable-parameter-matched baselines (by 20.65% rel. WER on test) and even match fully finetuned baselines in some settings.