Report on acoustic model building from 1M hours of unlabeled speech.
problem Building acoustic models from large amounts of unlabeled speech data.
method Student/teacher training, scheduled learning, BMUF, gradient threshold compression SGD.
result Extremely large amounts of unlabeled data lead to significant relative WER improvements.
AV-CPL uses continuous pseudo-labels for AVSR combining labeled and unlabeled data.
problem Improving AVSR performance with labeled and unlabeled data.
method Semi-supervised method using continuous pseudo-labels generated by the same AVSR model.
result Significant improvements in VSR performance on LRS3 dataset.
Improved SAR in asynchronous conversations using neural models and unlabeled data.
problem Lack of labeled data for SAR in asynchronous conversations.
method Hierarchical LSTM-CRF model, semi-supervised learning with word embeddings, adversarial training.
result Adversarial training improves SAR performance by leveraging labeled data from synchronous domains.
Bayesian SHMM discovers acoustic units from unlabeled speech.
problem Discovering language-specific acoustic units from unlabeled speech.
method Bayesian Subspace Hidden Markov Model (SHMM) trained on labeled data to find new acoustic units on target language.
result Significantly outperforms previous HMM-based systems and compares favorably with Variational Auto Encoder-HMM.
An ability to model a generative process and learn a latent representation for speech in an unsupervised fashion will be crucial to process vast quantities of unlabelled speech data. Recently, deep probabilistic generative models such as Variational Autoencoders (VAEs) have achieved tremendous success in modeling natur…
Automatically assessing emotional valence in human speech has historically been a difficult task for machine learning algorithms. The subtle changes in the voice of the speaker that are indicative of positive or negative emotional states are often "overshadowed" by voice characteristics relating to emotional intensity …
In conventional supervised pattern recognition tasks, model selection is typically accomplished by minimizing the classification error rate on a set of so-called development data, subject to ground-truth labeling by human experts or some other means. In the context of speech processing systems and other large-scale pra…
Study tackles hate speech against journalists on social media.
problem Hate speech against journalists on social media remains prevalent despite efforts.
method Defined journalist-specific hate speech, annotated tweets, trained deep learning models, and proposed an ensemble model.
result Proposed ensemble model outperforms individual models in detecting journalist-targeted hate speech.
The paper improves ASR accuracy using semi-supervised learning and dropout.
problem Improving ASR accuracy with limited labeled data.
method Training a seed model on limited labeled data, using dropout for uncertainty, and data selection for diversity.
result The approach significantly reduces ASR errors compared to baseline.
We propose the application of a semi-supervised learning method to improve the performance of acoustic modelling for automatic speech recognition based on deep neural net- works. As opposed to unsupervised initialisation followed by supervised fine tuning, our method takes advantage of both unlabelled and labelled data…
This paper introduces a method to select and weight pretext tasks for better self-supervised speech representation learning.
problem Combining pretext tasks for better performance in self-supervised speech representation learning.
method Estimating calibrated weights for partial losses corresponding to pretext tasks during self-supervised training.
result Groups of selected and weighted pretext tasks perform better than classic baselines in automatic speech recognition and speaker/emotion recognition.
Mic2Mic reduces microphone variability for speech systems.
problem Microphone variability impacts speech system accuracy.
method Cycle-consistent GANs for learning microphone variability reduction.
result Mic2Mic recovers 66% to 89% of lost accuracy due to microphone variability.
i-Mix improves contrastive learning across domains without domain-specific augmentations.
problem Improving contrastive representation learning for unlabeled data across diverse domains.
method i-Mix treats contrastive learning as a non-parametric classifier problem, mixing data in input and virtual label spaces.
result i-Mix consistently improves representation quality across image, speech, and tabular data domains.
End-to-end speech recognition using EEG without speech input.
problem Speech recognition without direct speech input.
method Implemented attention model and CTC-based ASR systems for EEG signals; fused EEG with noisy speech features.
result Demonstrated end-to-end speech recognition using EEG signals.
Paper shows continuous speech recognition with EEG features, no speech input.
problem Continuous speech recognition with limited vocabulary and noisy/no speech input.
method Connectionist temporal classification (CTC) model, EEG features, new deep learning architecture.
result Continuous speech recognition achieved on limited vocabulary with noisy/no speech input.
STRODE learns timings and dynamics from unlabeled time series data.
problem Learning dynamics of random event timings from unlabeled sensory inputs.
method Probabilistic Ordinary Differential Equation (STRODE) that samples from posterior point processes.
result Successfully infers event timings from synthetic and real-world datasets.
This paper presents sampling-based speech parameter generation using moment-matching networks for Deep Neural Network (DNN)-based speech synthesis. Although people never produce exactly the same speech even if we try to express the same linguistic and para-linguistic information, typical statistical speech synthesis pr…
Training a model to perform a task typically requires a large amount of data from the domains in which the task will be applied. However, it is often the case that data are abundant in some domains but scarce in others. Domain adaptation deals with the challenge of adapting a model trained from a data-rich source domai…
RUSLAN is a large Russian speech corpus for text-to-speech.
problem Lack of high-quality annotated Russian speech data for text-to-speech.
method Developed a large annotated Russian speech corpus and trained a neural network for text-to-speech synthesis.
result Synthesized speech quality evaluated with MOS scores: 4.05 for naturalness, 3.78 for intelligibility.
AeGAN improves speech clarity in noisy environments.
problem Improving speech recognition in crowded noisy environments.
method Generative adversarial networks (GAN) with a novel architecture.
result The proposed framework outperforms traditional and learning-based methods.
Detects AI-synthesized speech using cepstral and bispectral analysis.
problem Validating the authenticity of speech from AI-generated content.
method Integrates cepstral and bispectral analysis for distinguishing human from AI-synthesized speech.
result Higher-order statistics show less correlation for human speech compared to AI-synthesis, and cepstral analysis reveals unique power components.
The performance of automatic speech recognition systems(ASR) degrades in the presence of noisy speech. This paper demonstrates that using electroencephalography (EEG) can help automatic speech recognition systems overcome performance loss in the presence of noise. The paper also shows that distillation training of auto…
Improved visual speech synthesis using adapted ASR acoustic models.
problem Lack of synchronized audio, video, and depth data for speaker-independent speech-driven visual speech synthesis.
method Adapted an ASR acoustic model trained on audio-only data to the visual speech synthesis domain.
result Viewers significantly prefer animations generated from the adapted ASR acoustic model.
This paper presents a statistical method of single-channel speech enhancement that uses a variational autoencoder (VAE) as a prior distribution on clean speech. A standard approach to speech enhancement is to train a deep neural network (DNN) to take noisy speech as input and output clean speech. Although this supervis…
WaveCycleGAN2 improves speech synthesis quality by reducing aliasing.
problem Human ear can still distinguish synthesized speech from natural speech.
method WaveCycleGAN2 uses generators without down/up-sampling modules and combines discriminators from waveform and acoustic parameter domains.
result WaveCycleGAN2 achieves high-quality speech synthesis with comparable mean opinion scores to natural speech.
AudioPaLM combines text and speech models to improve speech processing and translation.
problem Improving speech understanding and generation models.
method Unified multimodal architecture combining text and speech models.
result AudioPaLM significantly outperforms existing speech translation systems.
SemiNAS reduces NAS cost by predicting accuracy of unlabeled architectures.
problem Costly evaluation of architectures limits NAS efficiency.
method SemiNAS uses unlabeled architectures to train an accuracy predictor.
result SemiNAS achieves comparable accuracy with less data.
Deep Neural Networks (DNNs) are universal function approximators providing state-of- the-art solutions on wide range of applications. Common perceptual tasks such as speech recognition, image classification, and object tracking are now commonly tackled via DNNs. Some fundamental problems remain: (1) the lack of a mathe…
Enhances spoken speech quality using EEG signals.
problem Improves speech clarity in noisy environments.
method Generative adversarial network (GAN), gated recurrent unit (GRU), temporal convolutional network (TCN) regression models.
result Significant improvement in speech enhancement quality compared to traditional methods.
Improved speech enhancement using diffusion models with MSE loss.
problem Efficient incorporation of noisy speech in generative speech enhancement.
method Augmented diffusion-based generative model with a MSE loss for enhanced speech.
result Proposed method improves speech enhancement performance compared to original diffusion model.
This work disentangles speech and non-speech components from found data.
problem Building robust acoustic models from found data with non-standard variations.
method Latent Stochastic Models and Multinode Latent Space Variational Autoencoders (VAE).
result Speech and music can be separated in the latent space of a VAE, independent of the language.
Continuous speech recognition from brain activity without vocalization.
problem Recognizing silent speech from EEG signals.
method Implemented a CTC ASR model using EEG signals.
result Demonstrated feasibility of EEG for continuous silent speech recognition.
Speech emotion recognition system using features and text.
problem Improving accuracy in emotion recognition from speech.
method Used speech features (Spectrogram, MFCC) and text, trained Deep Neural Networks.
result Combined MFCC-Text CNN model achieved highest accuracy.
Enhanced text-to-speech synthesizes expressive speech from a single example.
problem Creating a new expressive speech style from a single example of speech.
method Combines VAE and Normalizing Flows to improve disentanglement and naturalness.
result Reduces KL-divergence by 22% and improves perceptual metrics.
Diffusion models enhance speech without supervision.
problem Challenges in generalizing supervised speech enhancement methods to unseen conditions.
method Unsupervised speech enhancement using diffusion-based generative models.
result Demonstrates promising results compared to supervised and unsupervised baselines.
Improved speech recognition using EEG and video.
problem Enhancing continuous speech recognition systems.
method Implemented a CTC-based ASR model using EEG features.
result EEG features improve continuous visual speech recognition.
Speech synthesis from EEG features using RNN.
problem Speech synthesis from EEG data.
method Recurrent Neural Network (RNN) regression model to predict acoustic features from EEG features.
result Feasibility of synthesizing speech directly from EEG features demonstrated.
Survey on DNNs for speech processing, focusing on limited data challenges.
problem Challenges in training DNNs for speech tasks with limited data.
method Overview of techniques for few-shot speech processing.
result Promising few-shot techniques for speech processing.
Speech enhancement improved by adapting to unknown speakers without auxiliary signals.
problem Improving speech enhancement accuracy for unknown speakers.
method Adopting multi-task learning for speech enhancement and speaker identification, using multi-head self-attention.
result Achieved state-of-the-art performance and improved subjective quality.
Synthetic speech data improves keyword spotting models with fewer real examples.
problem Training models for recognizing spoken keywords with limited real data.
method Used a pre-trained speech embedding model to extract features for training a small keyword spotting model.
result A model trained on synthetic speech data can detect 10 keywords with the same accuracy as a model trained on over 500 real examples.
Paper explores EEG-based speech recognition using transformers, showing faster training and better performance for smaller vocabularies.
problem Continuous speech recognition using EEG features.
method Transformer-based ASR model compared to RNN-based models.
result Transformer models perform better for smaller vocabularies but RNN models outperform them for larger vocabularies.
Paper proposes a CNN for speech emotion recognition using center loss and reconstruction.
problem Speech emotion recognition (SER) in audio signals.
method Convolutional Neural Network (CNN) with center loss and reconstruction as regularizers.
result Proposed method achieves highly discriminative features for SER.
VoiceFilter-Lite separates speech from background in real-time for on-device speech recognition.
problem Separate speech from background in real-time for on-device speech recognition.
method Asymmetric loss, adaptive runtime suppression, quantization to 8-bit.
result VoiceFilter-Lite achieves real-time speech separation and maintains speech recognition performance.
Novel BCI system classifies imagined speech with high accuracy.
problem Classifying imagined speech from brain signals.
method Hierarchical deep learning with CNN and autoencoder.
result Achieved 83.42% average accuracy across six phonological tasks.
We propose a learning-based filter that allows us to directly modify a synthetic speech waveform into a natural speech waveform. Speech-processing systems using a vocoder framework such as statistical parametric speech synthesis and voice conversion are convenient especially for a limited number of data because it is p…
This research improves neural synthesizers for music sounds from speech data.
problem Applying speech synthesis techniques to musical instrument sounds.
method Comparison of three neural synthesizers in three scenarios: training, zero-shot learning, and fine-tuning.
result Neural synthesizers trained on speech data and fine-tuned on music data perform better.
Enhanced transformer converts whispered speech to natural speech.
problem Machine recognition of whispered speech is challenging.
method Proposes an enhanced transformer architecture trained end-to-end using supervised learning.
result Similar formant distributions of converted speech to groundtruth.
This paper converts speech to match a face image and vice versa.
problem Matching speech to a face image and vice versa.
method Proposes a model with speech converter, face encoder/decoder, and voice encoder.
result Trained model converts speech to match a face image and generates a face image that matches the voice of input speech.