Speech Reconstruction

Recent advancements in voice conversion systems have been largely driven by deep learning techniques, enabling the high-quality synthesis of human speech. However, existing models often fail to generate emotionally expressive speech, resulting in outputs that lack the richness of human communication. The main goal of this research is to fill this gap by incorporating a prosody-aware extension into embedding-guided neural voice conversion (EGNVC). This extension seeks to improve the emotional expressiveness and naturalness of the converted speech, while preserving both speaker identity and linguistic content. A new framework is introduced that integrates a content encoder, speaker embedder, and prosody extractor to address this challenge. The prosody extractor is designed to capture dynamic prosodic elements like pitch, energy, and timing from a reference audio signal. These features are then introduced into the decoding process through a prosody conditioning module, allowing for precise control over prosody during speech synthesis. FiLM (Feature-wise Linear Modulation) layers are employed to adjust intermediate features based on speaker and prosody maintained. The vocoder converts the generated mel-spectrograms into high-quality waveforms, ensuring a natural-sounding output. Experimental results show that the proposed model surpasses conventional voice conversion systems in maintaining linguistic content and prosodic features, with significant improvements in Mel-Cepstral Distortion (MCD), Pitch RMSE, and Energy RMSE. Subjective assessments, including Mean Opinion Score (MOS) and Emotion Classification Accuracy, highlight the model's enhanced emotional expressiveness and naturalness. These findings emphasize the critical role of prosody in high-quality voice conversion, offering a robust framework for creating more dynamic, expressive, and human-like speech synthesis systems.

This paper presents a zero-shot learning framework for speech-based emotion recognition across languages using contrastive learning. Traditional emotion AI models often depend on large, labeled corpora in a specific language, limiting global scalability. We introduce a contrastive embedding technique that aligns emotional cues across multilingual speech without explicit emotion labels in the target language. Trained on English and evaluated on German, Mandarin, and Arabic datasets, our approach achieves competitive accuracy compared to supervised baselines. The model's design enables practical deployment in customer service, digital therapy, and cross-cultural human-computer interaction.

The presented research focuses on the challenging task of creating lip-sync facial videos that align with a specified target speech segment. A novel deep-learning model has been developed to produce precise synthetic lip movements corresponding to the speech extracted from an audio source. Consequently, there are instances where portions of the visual data may fall out of sync with the updated audio and this challenge is handled through, a novel strategy, leveraging insights from a robust lip-sync discriminator. Additionally, this study introduces fresh criteria and evaluation benchmarks for assessing lip synchronization in unconstrained videos. LipChanger demonstrates improved PSNR values, indicative of enhanced image quality. Furthermore, it exhibits highly accurate lip synthesis, as evidenced by lower LMD values and higher SSIM values. These outcomes suggest that the LipChanger approach holds significant potential for enhancing lip synchronization in talking face videos, resulting in more realistic lip movements. The proposed LipChanger model and its associated evaluation benchmarks show promise and could potentially contribute to advancements in lip-sync technology for unconstrained talking face videos.

Lipreading involves using visual data to recognize spoken words by analyzing the movements of the lips and surrounding area. It is a hot research topic with many potential applications, such as human-machine interaction and enhancing audio speech recognition. Recent deep-learning based works aim to integrate visual features extracted from the mouth region with landmark points on the lip contours. However, employing a simple combination method such as concatenation may not be the most effective approach to get the optimal feature vector. To address this challenge, firstly, we propose a cross-attention fusion-based approach for large lexicon Arabic vocabulary to predict spoken words in videos. Our method leverages the power of cross-attention networks to efficiently integrate visual and geometric features computed on the mouth region. Secondly, we introduce the first large-scale Lipreading in the Wild for Arabic (LRW-AR) dataset containing 20,000 videos for 100-word classes, uttered by 36 speakers. The experimental results obtained on LRW-AR and ArabicVisual databases showed the effectiveness and robustness of the proposed approach in recognizing Arabic words. Our work provides insights into the feasibility and effectiveness of applying lipreading techniques to the Arabic language, opening doors for further research in this field. Link to the

Human speech can be characterized by different components, including semantic content, speaker identity and prosodic information. Significant progress has been made in disentangling representations for semantic content and speaker identity in Automatic Speech Recognition (ASR) and speaker verification tasks respectively. However, it is still an open challenging research question to extract prosodic information because of the intrinsic association of different attributes, such as timbre and rhythm, and because of the need for unsupervised training schemes to achieve robust large-scale and speakerindependent ASR. The aim of this paper is to address the disentanglement of emotional prosody from speech based on unsupervised reconstruction. Specifically, we identify, design, implement and integrate three crucial components in our proposed speech reconstruction model Prosody2Vec: (1) a unit encoder that transforms speech signals into discrete units for semantic content, (2) a pretrained speaker verification model to generate speaker identity embeddings, and (3) a trainable prosody encoder to learn prosody representations. We first pretrain the Prosody2Vec representations on unlabelled emotional speech corpora, then fine-tune the model on specific datasets to perform Speech Emotion Recognition (SER) and Emotional Voice Conversion (EVC) tasks. Both objective (weighted and unweighted accuracies) and subjective (mean opinion score) evaluations on the EVC task suggest that Prosody2Vec effectively captures general prosodic features that can be smoothly transferred to other emotional speech. In addition, our SER experiments on the IEMOCAP dataset reveal that the prosody features learned by Prosody2Vec are complementary and beneficial for the performance of widely used speech pretraining models and surpass the state-of-the-art methods when combining Prosody2Vec with HuBERT representations. Some audio samples can be found on our demo website 1 .

Computational speech reconstruction algorithms have the ultimate aim of returning natural sounding speech to aphonic and dysphonic individuals. These algorithms can also be used by unimpaired speakers for communicating sensitive or private information. When the glottis loses function due to disease or surgery, aphonic and dysphonic patients retain the power of vocal tract modulation to some degree but they are unable to speak anything more than hoarse whispers without prosthetic aid. While whispering can be seen as a natural and secondary aspect of speech communications for most people, it becomes the primary mechanism of communications for those who have impaired voice production mechanisms, such as laryngectomees. In this paper, by considering the current limitations of speech reconstruction methods, a novel algorithm for converting whispers to normal speech is proposed and the efficiency of the algorithm is discussed. The proposed algorithm relies upon twin mapping models and makes use of artificially generated whispers (called whisperised speech) to regenerate natural phonated speech from whispers. Through a training-based approach, the mapping models exploit whisperised speech to overcome frame to frame time alignment problem in the speech reconstruction process.

An articulatory speech sythesiser is presented that extends the Recurrent Gradient-based Motor Inference Model for Speech Resynthesis [1] with a classifier optimizing for semantic discrimination. The new design features show promising results in improving model performance.

A growing need for on-device machine learning has led to an increased interest in lightweight neural networks that lower model complexity while retaining performance. While a variety of general-purpose techniques exist in this context, very few approaches exploit domain-specific properties to further improve upon the capacity-performance trade-off. In this paper, extending our prior work [1], we train a network to emulate the behaviour of an audio codec and use this network to construct a loss. By approximating the psychoacoustic model underlying the codec, our approach enables lightweight neural networks to focus on perceptually relevant properties without wasting their limited capacity on imperceptible signal components. We adapt our method to two audio source separation tasks, demonstrate an improvement in performance for small-scale networks via listening tests, characterize the behaviour of the loss network in detail, and quantify the relationship between performance gain and model capacity. Our work illustrates the potential for incorporating perceptual principles into objective functions for neural networks.

Recent advances in brain decoding have made it possible to classify image categories based on neural activity. Increasing numbers of studies have further attempted to reconstruct the image itself. However, because images of objects and scenes inherently involve spatial layout information, the reconstruction usually requires retinotopically organized neural data with high spatial resolution, such as fMRI signals. In contrast, spatial layout does not matter in the perception of ‘texture’, which is known to be represented as spatially global image statistics in the visual cortex. This property of ‘texture’ enables us to reconstruct the perceived image from EEG signals, which have a low spatial resolution. Here, we propose an MVAE-based approach for reconstructing texture images from visual evoked potentials measured from observers viewing natural textures such as the textures of various surfaces and object ensembles. This approach allowed us to reconstruct images that perceptually rese...

Neural text-to-speech synthesis (NTTS) models have shown significant progress in generating high-quality speech, however they require a large quantity of training data. This makes creating models for multiple styles expensive and time-consuming. In this paper different styles of speech are analysed based on prosodic variations, from this a model is proposed to synthesise speech in the style of a newscaster, with just a few hours of supplementary data. We pose the problem of synthesising in a target style using limited data as that of creating a bi-style model that can synthesise both neutral-style and newscasterstyle speech via a one-hot vector which factorises the two styles. We also propose conditioning the model on contextual word embeddings, and extensively evaluate it against neutral NTTS, and neutral concatenativebased synthesis. This model closes the gap in perceived style-appropriateness between natural recordings for newscaster-style of speech, and neutral speech synthesis by approximately two-thirds.

Both acoustic and visual information influence human perception of speech. For this reason, the lack of audio in a video sequence determines an extremely low speech intelligibility for untrained lip readers. In this paper, we present a way to synthesise speech from the silent video of a talker using deep learning. The system learns a mapping function from raw video frames to acoustic features and reconstructs the speech with a vocoder synthesis algorithm. To improve speech reconstruction performance, our model is also trained to predict text information in a multi-task learning fashion and it is able to simultaneously reconstruct and recognise speech in real time. The results in terms of estimated speech quality and intelligibility show the effectiveness of our method, which exhibits an improvement over existing video-to-speech approaches.

Much of the recent literature on automatic speech recognition (ASR) is taking an end-to-end approach. Unlike English where the writing system is closely related to sound, Chinese characters (Hanzi) represent meaning, not sound. We propose factoring audio → Hanzi into two sub-tasks: (1) audio → Pinyin and (2) Pinyin → Hanzi, where Pinyin is a system of phonetic transcription of standard Chinese. Factoring the audio → Hanzi task in this way achieves 3.9% CER (character error rate) on the Aishell-1 corpus, the best result reported on this dataset so far.

We present SL-ReDu, a recently commenced innovative project that aims to exploit deep-learning progress to advance the state-of-theart in video-based automatic recognition of Greek Sign Language (GSL), while focusing on the use-case of GSL education as a second language. We first briefly overview the project goals, focal areas, and timeline. We then present our initial deep learning-based approach for GSL recognition that employs efficient visual tracking of the signer hands, convolutional neural networks for feature extraction, and attention-based encoder-decoder sequence modeling for sign prediction. Finally, we report experimental results for small-vocabulary, isolated GSL recognition on the single-signer "Polytropon" corpus. To our knowledge, this work constitutes the first application of deep-learning techniques to GSL. CCS CONCEPTS • Computing methodologies → Neural networks; • Humancentered computing → Accessibility technologies; • Applied computing → Computer-assisted instruction.

Dubbing contributes to a larger international distribution of multimedia documents. It aims to replace the original voice in a source language by a new one in a target language. For now, the target voice selection procedure, called voice casting, is manually performed by human experts. This selection is not exclusively based on acoustic similarity between the two voices. Actually, it is also supported by more subjective criteria such as the "color" of the voice, sociocultural choices.. . The objective of this work is to model a voice similarity metric able to embed all the concerned voice characteristics, including the observers' receptive interests. In this paper, we propose a Siamese Neural Networks-based approach, measuring proximity between the original and dubbed voices. We propose an adapted jackknifing cross-validation method to evaluate our similarity model on unseen voices. The results show that we successfully capture information allowing two voices to be associated, with respect to the character's or role's abstract dimension.

Speech-driven facial animation is useful for a variety of applications such as telepresence, chatbots, etc. The necessary attributes of having a realistic face animation are 1) audiovisual synchronization (2) identity preservation of the target individual (3) plausible mouth movements (4) presence of natural eye blinks. The existing methods mostly address the audiovisual lip synchronization, and few recent works have addressed synthesis of natural eye blinks for overall video realism. In this paper, we propose a method for identity-preserving realistic facial animation from speech. We first generate person-independent facial landmarks from audio using DeepSpeech features for invariance to different voices, accents, etc. To add realism, we impose eye blinks on facial landmarks using unsupervised learning and retarget the person-independent landmarks to person-specific landmarks to preserve the identity-related facial structure which helps in generation of plausible mouth shapes of the target identity. Finally, we use LSGAN to generate the facial texture from person-specific facial landmarks, using an attention mechanism that helps to preserve identity-related texture. An extensive comparison of our proposed method with the current state-of-the-art methods demonstrate a significant improvement in terms of lip synchronization accuracy, image reconstruction quality, sharpness, and identity-preservation. A user study also reveals improved realism of our animation results over the stateof-the-art methods. To the best of our knowledge, this is the first work in speech-driven 2D facial animation that simultaneously addresses all the above-mentioned attributes of a realistic speech driven face animation.

Previously proposed FullSubNet has achieved outstanding performance in Deep Noise Suppression (DNS) Challenge and attracted much attention. However, it still encounters issues such as inputoutput mismatch and coarse processing for frequency bands. In this paper, we propose an extended single-channel real-time speech enhancement framework called FullSubNet+ with following significant improvements. First, we design a lightweight multi-scale time sensitive channel attention (MulCA) module which adopts multi-scale convolution and channel attention mechanism to help the network focus on more discriminative frequency bands for noise reduction. Then, to make full use of the phase information in noisy speech, our model takes all the magnitude, real and imaginary spectrograms as inputs. Moreover, by replacing the long short-term memory (LSTM) layers in original full-band model with stacked temporal convolutional network (TCN) blocks, we design a more efficient full-band module called full-band extractor. The experimental results in DNS Challenge dataset show the superior performance of our FullSubNet+, which reaches the state-of-theart (SOTA) performance and outperforms other existing speech enhancement approaches.

Previously proposed FullSubNet has achieved outstanding performance in Deep Noise Suppression (DNS) Challenge and attracted much attention. However, it still encounters issues such as inputoutput mismatch and coarse processing for frequency bands. In this paper, we propose an extended single-channel real-time speech enhancement framework called FullSubNet+ with following significant improvements. First, we design a lightweight multi-scale time sensitive channel attention (MulCA) module which adopts multi-scale convolution and channel attention mechanism to help the network focus on more discriminative frequency bands for noise reduction. Then, to make full use of the phase information in noisy speech, our model takes all the magnitude, real and imaginary spectrograms as inputs. Moreover, by replacing the long short-term memory (LSTM) layers in original full-band model with stacked temporal convolutional network (TCN) blocks, we design a more efficient full-band module called full-band extractor. The experimental results in DNS Challenge dataset show the superior performance of our FullSubNet+, which reaches the state-of-theart (SOTA) performance and outperforms other existing speech enhancement approaches.

While Automatic Speech Recognition (ASR) models have shown significant advances with the introduction of unsupervised or selfsupervised training techniques, these improvements are still only limited to a subsection of languages and speakers. Transfer learning enables the adaptation of large-scale multilingual models to not only low-resource languages but also to more specific speaker groups. However, fine-tuning on data from new domains is usually accompanied by a decrease in performance on the original domain. Therefore, in our experiments, we examine how well the performance of large-scale ASR models can be approximated for smaller domains, with our own dataset of German Senior Voice Commands (SVC-de), and how much of the general speech recognition performance can be preserved by selectively freezing parts of the model during training. To further increase the robustness of the ASR model to vocabulary and speakers outside of the fine-tuned domain, we apply Experience Replay [20] for continual learning. By adding only a fraction of data from the original domain, we are able to reach Word-Error-Rates (WERs) below 5% on the new domain, while stabilizing performance for general speech recognition at acceptable WERs.

In the absence of large-scale in-domain supervised training data, ASR models can achieve reasonable performance through pre-training on additional data that is unlabeled, mismatched or both. Given such data constraints, we compare pre-training endto-end models on matched but unlabeled data (unsupervised) and on labeled but mismatched data (supervised), where the labeled data is mismatched in either domain or language. Across encoder architectures, pre-training methods and languages, our experiments indicate that both types of pre-training improve performance, with relative WER reductions of 15-30% in the domain mismatch case and up to 15% in the language mismatch condition. We further find that the advantage from unsupervised pre-training is most prominent when there is no matched and labeled fine-tuning data, provided that a sufficient amount of mismatched data is still available for supervised fine-tuning.

The aim of this work is to investigate the impact of 1 crossmodal self-supervised pre-training for speech reconstruc-2 tion (video-to-audio) by leveraging the natural co-occurrence 3 of audio and visual streams in videos. We propose LipSound2 4 that consists of an encoder-decoder architecture and location-5 aware attention mechanism to map face image sequences to 6 mel-scale spectrograms directly without requiring any human 7 annotations. The proposed LipSound2 model is first pre-trained 8 on ∼2400-h multilingual (e.g., English and German) audiovisual 9 data (VoxCeleb2). To verify the generalizability of the proposed 10 method, we then fine-tune the pre-trained model on domain-11 specific datasets (GRID and TCD-TIMIT) for English speech 12 reconstruction and achieve a significant improvement on speech 13 quality and intelligibility compared to previous approaches in 14 speaker-dependent and speaker-independent settings. In addition 15 to English, we conduct Chinese speech reconstruction on the 16 Chinese Mandarin Lip Reading (CMLR) dataset to verify the 17 impact on transferability. Finally, we train the cascaded lip 18 reading (video-to-text) system by fine-tuning the generated audios 19 on a pre-trained speech recognition system and achieve the state-20 of-the-art performance on both English and Chinese benchmark 21 datasets. 22 Index Terms-Lip reading, self-supervised pre-training, speech 23 recognition, speech reconstruction. 24 I. INTRODUCTION 25 I NSPIRED by human bimodal perception [1] in which both 26 sight and sound are used to improve the comprehension of 27 speech, a lot of effort has been spent on speech processing 28 tasks by leveraging visual information, for example, integrat-29 ing simultaneous lip movement sequences into speech recogni-30 tion [2], [3], guiding neural networks in isolating target speech 31 signals with a static face image for speech separation [4], [5], 32 and grounding speech recognition with visual objects and 33 scene information [6], [7]. Multimodal audiovisual methods 34 achieve significant improvement over single modality models 35 since the visual signals are invariant to acoustic noise and

We present SL-ReDu, a recently commenced innovative project that aims to exploit deep-learning progress to advance the state-of-theart in video-based automatic recognition of Greek Sign Language (GSL), while focusing on the use-case of GSL education as a second language. We first briefly overview the project goals, focal areas, and timeline. We then present our initial deep learning-based approach for GSL recognition that employs efficient visual tracking of the signer hands, convolutional neural networks for feature extraction, and attention-based encoder-decoder sequence modeling for sign prediction. Finally, we report experimental results for small-vocabulary, isolated GSL recognition on the single-signer "Polytropon" corpus. To our knowledge, this work constitutes the first application of deep-learning techniques to GSL. CCS CONCEPTS • Computing methodologies → Neural networks; • Humancentered computing → Accessibility technologies; • Applied computing → Computer-assisted instruction.

Computational speech reconstruction algorithms have the ultimate aim of returning natural sounding speech to aphonic and dysphonic individuals. These algorithms can also be used by unimpaired speakers for communicating sensitive or private information. When the glottis loses function due to disease or surgery, aphonic and dysphonic patients retain the power of vocal tract modulation to some degree but they are unable to speak anything more than hoarse whispers without prosthetic aid. While whispering can be seen as a natural and secondary aspect of speech communications for most people, it becomes the primary mechanism of communications for those who have impaired voice production mechanisms, such as laryngectomees. In this paper, by considering the current limitations of speech reconstruction methods, a novel algorithm for converting whispers to normal speech is proposed and the efficiency of the algorithm is discussed. The proposed algorithm relies upon twin mapping models and makes use of artificially generated whispers (called whisperised speech) to regenerate natural phonated speech from whispers. Through a training-based approach, the mapping models exploit whisperised speech to overcome frame to frame time alignment problem in the speech reconstruction process.

Personal assistants are becoming more pervasive in our envi-ronments but still do not provide natural interactions. Their lack of realism in term of expressiveness and their lack of visual feedback can create frustrating experiences and make users lose patience. In this sense, we propose an end-to-end trainable neural architecture for text-driven 3D mouth animations. Previous works showed such architectures provide better realism and could open the door for integrated affective Human Computer Interface (HCI). Our study shows that such visual feedback improves users’ comfort for 78%of the candidates significantly while slightly improving their time perception.

Our goal is to spot words in silent speech videos without explicitly recognizing the spoken words, where the lip motion of the speaker is clearly visible and audio is absent. Existing work in this domain has mainly focused on recognizing a fixed set of words in word-segmented lip videos, which limits the applicability of the learned model due to limited vocabulary and high dependency on the model's recognition performance. Our contribution is twofold: (1) we develop a pipeline for recognition-free retrieval and show its performance against recognition-based retrieval on a large-scale dataset and another set of out-of-vocabulary words. (2) We introduce a query expansion technique using pseudo-relevant feedback and propose a novel re-ranking method based on maximizing the correlation between spatiotemporal landmarks of the query and the top retrieval candidates. Our word spotting method achieves 35% higher mean average precision over recognition-based method on large-scale LRW dataset. We also demonstrate the application of the method by word spotting in a popular speech video ("The great dictator" by Charlie Chaplin) where we show that the word retrieval can be used to understand what was spoken perhaps in the silent movies. Finally, we compare our model against ASR in a noisy environment and analyze the effect of the performance of underlying lip-reader and input video quality on the proposed word spotting pipeline.

Lip-reading is a technique to understand speech by observing a speaker’s lips movement. It has numerous applications; for example, it is helpful for hearing impaired persons and understanding the speech in noisy environments. Most of the previous works of lips-reading focused on frontal and near frontal face lip-reading and some of them targeted multiple poses in high quality videos. However, their results are not satisfactory on low quality videos containing multiple poses. In this research work, a lip-reading framework is proposed for improving the recognition rate in low quality videos. In this work, a Multiple Pose (MP) dataset of low quality videos containing multiple extreme poses is built. The proposed framework decomposes the input video into frames and enhances them by applying the Contrast Limited Adaptive Histogram Equalization (CLAHE) method. Next, faces are detected from enhanced frames and frontalized the multiple poses using the face frontalization Generative Adversar...

This is the age of instant gratification. Browsing the entire Publication is dawdle , hence we have proposed an application that summarizes all your reading's in a snap of time using AI technologies. The system is composed of Optical Character Recognition(OCR) engine to convert the image to text and transformer to summarise the text, dispensing recurrent networks followed by feature prediction network that maps character embeddings to mel-spectrogram and a GAN based vocoder to convert spectrogram to time based waveforms. Through extensive experiments we demonstrate digest podcast ability to recognize, summarize, speech synthesis for summarized audio generation. Our code can be found at https://github.com/mitran27/Digest_Cassette.

Given an arbitrary face image and an arbitrary speech clip, the proposed work attempts to generate the talking face video with accurate lip synchronization. Existing works either do not consider temporal dependency across video frames thus yielding abrupt facial and lip movement or are limited to the generation of talking face video for a specific person thus lacking generalization capacity. We propose a novel conditional recurrent generation network that incorporates both image and audio features in the recurrent unit for  temporal dependency. To achieve both image- and video-realism, a pair of spatial-temporal discriminators are included in the network for better image/video quality. Since accurate lip synchronization is essential to the success of talking face video generation, we also construct a lip-reading discriminator to boost the accuracy of lip synchronization. We also extend the network to model the natural pose and expression of talking face on the Obama Dataset. Extensi...

Silent speech decoding (SSD), based on articulatory neuromuscular activities, has become a prevalent task of brain–computer interfaces (BCIs) in recent years. Many works have been devoted to decoding surface electromyography (sEMG) from articulatory neuromuscular activities. However, restoring silent speech in tonal languages such as Mandarin Chinese is still difficult. This paper proposes an optimized sequence-to-sequence (Seq2Seq) approach to synthesize voice from the sEMG-based silent speech. We extract duration information to regulate the sEMG-based silent speech using the audio length. Then, we provide a deep-learning model with an encoder–decoder structure and a state-of-the-art vocoder to generate the audio waveform. Experiments based on six Mandarin Chinese speakers demonstrate that the proposed model can successfully decode silent speech in Mandarin Chinese and achieve a character error rate (CER) of 6.41% on average with human evaluation.

In this work, a novel direct speech-to-speech methodology for translation is proposed; it is based on an LSTM neural network structure which has proven useful for translation in the classical way, i.e., the one consisting of three stages: speech-to-text conversion, text-to-text translation, and text-to-speech synthesis. In contrast with traditional approaches, the one in this work belongs to the recently appeared idea of direct translation without text representation, as this sort of training better corresponds to the way oral language learning takes place in humans. As a proof of concept digits are translated from an audio source in Spanish and pronounced as an audio signal in English. Advantages and disadvantages of the proposal when compared with traditional methodologies are discussed.

Talking head generation is to synthesize a lip-synchronized talking head video by inputting an arbitrary face image and corresponding audio clips. Existing methods ignore not only the interaction and relationship of cross-modal information, but also the local driving information of the mouth muscles. In this study, we propose a novel generative framework that contains a dilated non-causal temporal convolutional self-attention network as a multimodal fusion module to promote the relationship learning of cross-modal features. In addition, our proposed method uses both audio-and speech-related facial action units (AUs) as driving information. Speech-related AU information can guide mouth movements more accurately. Because speech is highly correlated with speech-related AUs, we propose an audio-to-AU module to predict speech-related AU information. We utilize pre-trained AU classifier to ensure that the generated images contain correct AU information. We verify the effectiveness of the proposed model on the GRID and TCD-TIMIT datasets. An ablation study is also conducted to verify the contribution of each component. The results of quantitative and qualitative experiments demonstrate that our method outperforms existing methods in terms of both image quality and lip-sync accuracy.

The task of talking head generation is to synthesize a lip synchronized talking head video by inputting an arbitrary face image and audio clips. Most existing methods ignore the local driving information of the mouth muscles. In this paper, we propose a novel recurrent generative network that uses both audio and speech-related facial action units (AUs) as the driving information. AU information related to the mouth can guide the movement of the mouth more accurately. Since speech is highly correlated with speech-related AUs, we propose an Audio-to-AU module in our system to predict the speech-related AU information from speech. In addition, we use AU classifier to ensure that the generated images contain correct AU information. Frame discriminator is also constructed for adversarial training to improve the realism of the generated face. We verify the effectiveness of our model on the GRID dataset and TCD-TIMIT dataset. We also conduct an ablation study to verify the contribution of each component in our model. Quantitative and qualitative experiments demonstrate that our method outperforms existing methods in both image quality and lip-sync accuracy. Recently, talking head generation has attracted more and more attention in the fields of academic and industry, which is essential in the applications of human-computer interaction, film making, virtual reality, computer games, etc. This research explores how to generate a talking head video by inputting anyone's image as an identity image and driving information related to mouth movement, e.g., speech audio, and text. Before deep learning became popular, many researchers in early work relied on Hidden Markov Models (HMM) to capture the dynamic relationship between audio and lip motion

Much of the recent literature on automatic speech recognition (ASR) is taking an end-to-end approach. Unlike English where the writing system is closely related to sound, Chinese characters (Hanzi) represent meaning, not sound. We propose factoring audio → Hanzi into two sub-tasks: (1) audio → Pinyin and (2) Pinyin → Hanzi, where Pinyin is a system of phonetic transcription of standard Chinese. Factoring the audio → Hanzi task in this way achieves 3.9% CER (character error rate) on the Aishell-1 corpus, the best result reported on this dataset so far.

Text-based voice editing (TBVE) uses synthetic output from textto-speech (TTS) systems to replace words in an original recording. Recent work has used neural models to produce edited speech that is similar to the original speech in terms of clarity, speaker identity, and prosody. However, one limitation of prior work is the usage of finetuning to optimise performance: this requires further model training on data from the target speaker, which is a costly process that may incorporate potentially sensitive data into server-side models. In contrast, this work focuses on the zero-shot approach which avoids finetuning altogether, and instead uses pretrained speaker verification embeddings together with a jointly trained reference encoder to encode utterance-level information that helps capture aspects such as speaker identity and prosody. Subjective listening tests find that both utterance embeddings and a reference encoder improve the continuity of speaker identity and prosody between the edited synthetic speech and unedited original recording in the zero-shot setting.

Supervised ASR models have reached unprecedented levels of accuracy, thanks in part to ever-increasing amounts of labelled training data. However, in many applications and locales, only moderate amounts of data are available, which has led to a surge in semi-and weakly-supervised learning research. In this paper, we conduct a large-scale study evaluating the effectiveness of weakly-supervised learning for speech recognition by using loosely related contextual information as a surrogate for ground-truth labels. For weakly supervised training, we use 50k hours of public English social media videos along with their respective titles and post text to train an encoder-decoder transformer model. Our best encoder-decoder models achieve an average of 20.8% WER reduction over a 1000 hours supervised baseline, and an average of 13.4% WER reduction when using only the weakly supervised encoder for CTC fine-tuning. Our results show that our setup for weak supervision improved both the encoder acoustic representations as well as the decoder language generation abilities.

Many semi-and weakly-supervised approaches have been investigated for overcoming the labeling cost of building highquality speech recognition systems. On the challenging task of transcribing social media videos in low-resource conditions, we conduct a large scale systematic comparison between two self-labeling methods on one hand, and weakly-supervised pretraining using contextual metadata on the other. We investigate distillation methods at the frame level and the sequence level for hybrid, encoder-only CTC-based, and encoder-decoder speech recognition systems on Dutch and Romanian languages using 27,000 and 58,000 hours of unlabeled audio respectively. Although all approaches improved upon their respective baseline WERs by more than 8%, sequence-level distillation for encoder-decoder models provided the largest relative WER reduction of 20% compared to the strongest data-augmented supervised baseline.

In this paper we present VDTTS, a Visually-Driven Textto-Speech model. Motivated by dubbing, VDTTS takes advantage of video frames as an additional input alongside text, and generates speech that matches the video signal. We demonstrate how this allows VDTTS to, unlike plain TTS models, generate speech that not only has prosodic variations like natural pauses and pitch, but is also synchronized to the input video. Experimentally, we show our model produces wellsynchronized outputs, approaching the video-speech synchronization quality of the ground-truth, on several challenging benchmarks including "in-the-wild" content from VoxCeleb2. We encourage the reader to view the demo videos demonstrating video-speech synchronization, robustness to speaker ID swapping, and prosody.

In this paper, we present an audio-visual model to perform speech super-resolution at large scale-factors (8× and 16×). Previous works attempted to solve this problem using only the audio modality as input and thus were limited to low scale-factors of 2× and 4×. In contrast, we propose to incorporate both visual and auditory signals to superresolve speech of sampling rates as low as 1kHz. In such challenging situations, the visual features assist in learning the content and improves the quality of the generated speech. Further, we demonstrate the applicability of our approach to arbitrary speech signals where the visual stream is not accessible. Our “pseudo-visual network” precisely synthesizes the visual stream solely from the low-resolution speech input. Extensive experiments and the demo video illustrate our method’s remarkable results and benefits over state-of-the-art audio-only speech super-resolution approaches. Figure 1: We present an audio-visual model for super-resolving v...

In this work, we rethink the task of speech enhancement in unconstrained real-world environments. Current state-of-the-art methods use only the audio stream and are limited in their performance in a wide range of real-world noises. Recent works using lip movements as additional cues improve the quality of generated speech over "audio-only" methods. But, these methods cannot be used for several applications where the visual stream is unreliable or completely absent. We propose a new paradigm for speech enhancement by exploiting recent breakthroughs in speech-driven lip synthesis. Using one such model as a teacher network, we train a robust student network to produce accurate lip movements that mask away the noise, thus acting as a "visual noise filter". The intelligibility of the speech enhanced by our pseudo-lip approach is comparable (< 3% difference) to the case of using real lips. This implies that we can exploit the advantages of using lip movements even in the absence of a real video stream. We rigorously evaluate our model using quantitative metrics as well as human evaluations. Additional ablation studies and a demo video on our website containing qualitative comparisons and results clearly illustrate the effectiveness of our approach. We provide a demo video which clearly illustrates the effectiveness of our proposed approach on our website: http://cvit.iiit.ac.in/research/projects/cvit-projects/ visual-speech-enhancement-without-a-real-visual-stream. The code and models are also released for future research: https://github.com/Sindhu-Hegde/ pseudo-visual-speech-denoising.

Figure 1: We address the problem of generating speech from silent lip videos for any speaker in the wild. Previous works train either on large amounts of data of isolated speakers or in laboratory settings with a limited vocabulary. Conversely, we can generate speech for the lip movements of arbitrary identities in any voice without additional speaker-specific fine-tuning. Our new VAE-GAN approach allows us to learn strong audiovisual associations despite the ambiguous nature of the task.

State-of-the-art text-to-speech (TTS) systems require several hours of recorded speech data to generate high-quality synthetic speech. When using reduced amounts of training data, standard TTS models suffer from speech quality and intelligibility degradations, making training low-resource TTS systems problematic. In this paper, we propose a novel extremely low-resource TTS method called Voice Filter that uses as little as one minute of speech from a target speaker. It uses voice conversion (VC) as a post-processing module appended to a pre-existing high-quality TTS system and marks a conceptual shift in the existing TTS paradigm, framing the few-shot TTS problem as a VC task. Furthermore, we propose to use a duration-controllable TTS system to create a parallel speech corpus to facilitate the VC task. Results show that the Voice Filter outperforms state-of-the-art few-shot speech synthesis techniques in terms of objective and subjective metrics on one minute of speech on a diverse set of voices, while being competitive against a TTS model built on 30 times more data. 1

This work introduces a predominantly parallel, end-to-end TTS model based on normalizing flows. It extends prior parallel approaches by additionally modeling speech rhythm as a separate generative distribution to facilitate variable token duration during inference. We further propose a robust framework for the on-line extraction of speech-text alignments – a critical yet highly unstable learning problem in end-to-end TTS frameworks. Our experiments demonstrate that our proposed techniques yield improved alignment quality, better output diversity compared to controlled baselines.

In this paper we propose WaveGlow: a flow-based network capable of generating high quality speech from melspectrograms. WaveGlow combines insights from Glow [1] and WaveNet [2] in order to provide fast, efficient and highquality audio synthesis, without the need for auto-regression. WaveGlow is implemented using only a single network, trained using only a single cost function: maximizing the likelihood of the training data, which makes the training procedure simple and stable. Our PyTorch implementation produces audio samples at a rate of more than 500 kHz on an NVIDIA V100 GPU. Mean Opinion Scores show that it delivers audio quality as good as the best publicly available WaveNet implementation. All code will be made publicly available online [3].

Audio Packet Loss Concealment (PLC) is the hiding of gaps in audio streams caused by data transmission failures in packet switched networks. This is a common problem, and of increasing importance as end-to-end VoIP telephony and teleconference systems become the default and ever more widely used form of communication in business as well as in personal usage. This paper presents the INTERSPEECH 2022 Audio Deep Packet Loss Concealment challenge. We first give an overview of the PLC problem, and introduce some classical approaches to PLC as well as recent work. We then present the open source dataset released as part of this challenge as well as the evaluation methods and metrics used to determine the winner. We also briefly introduce PLCMOS, a novel data-driven metric that can be used to quickly evaluate the performance PLC systems. Finally, we present the results of the INTERSPEECH 2022 Audio Deep PLC Challenge, and provide a summary of important takeaways.

One of the main challenges for end-to-end speech translation is data scarcity. We leverage pseudo-labels generated from unlabeled audio by a cascade and an end-to-end speech translation model. This provides 8.3 and 5.7 BLEU gains over a strong semi-supervised baseline on the MuST-C English-French and English-German datasets, reaching state-of-the art performance. The effect of the quality of the pseudo-labels is investigated. Our approach is shown to be more effective than simply pre-training the encoder on the speech recognition task. Finally, we demonstrate the effectiveness of self-training by directly generating pseudo-labels with an end-to-end model instead of a cascade model.

We present a voice conversion solution using recurrent sequence to sequence modeling for DNNs. Our solution takes advantage of recent advances in attention based modeling in the fields of Neural Machine Translation (NMT), Text-to-Speech (TTS) and Automatic Speech Recognition (ASR). The problem consists of converting between voices in a parallel setting when {\it $ $} audio pairs are available. Our seq2seq architecture makes use of a hierarchical encoder to summarize input audio frames. On the decoder side, we use an attention based architecture used in recent TTS works. Since there is a dearth of large multispeaker voice conversion databases needed for training DNNs, we resort to training the network with a large single speaker dataset as an autoencoder. This is then adapted for the smaller multispeaker voice conversion datasets available for voice conversion. In contrast with other voice conversion works that use $F_0$, duration and linguistic features, our system uses mel spectrog...

The task of talking head generation is to synthesize a lip synchronized talking head video by inputting an arbitrary face image and audio clips. Most existing methods ignore the local driving information of the mouth muscles. In this paper, we propose a novel recurrent generative network that uses both audio and speech-related facial action units (AUs) as the driving information. AU information related to the mouth can guide the movement of the mouth more accurately. Since speech is highly correlated with speech-related AUs, we propose an Audio-to-AU module in our system to predict the speech-related AU information from speech. In addition, we use AU classifier to ensure that the generated images contain correct AU information. Frame discriminator is also constructed for adversarial training to improve the realism of the generated face. We verify the effectiveness of our model on the GRID dataset and TCD-TIMIT dataset. We also conduct an ablation study to verify the contribution of ...

This paper introduces WaveGrad 2, a non-autoregressive generative model for text-to-speech synthesis. WaveGrad 2 is trained to estimate the gradient of the log conditional density of the waveform given a phoneme sequence. The model takes an input phoneme sequence, and through an iterative refinement process, generates an audio waveform. This contrasts to the original WaveGrad vocoder which conditions on mel-spectrogram features, generated by a separate model. The iterative refinement process starts from Gaussian noise, and through a series of refinement steps (e.g., 50 steps), progressively recovers the audio sequence. WaveGrad 2 offers a natural way to trade-off between inference speed and sample quality, through adjusting the number of refinement steps. Experiments show that the model can generate high fidelity audio, approaching the performance of a state-of-the-art neural TTS system. We also report various ablation studies over different model configurations. Audio samples are available at https://wavegrad.github.io/v2.

Audiovisual speech enhancement (AVSE) methods use both audio and visual features for the task of speech enhancement and the use of visual features has been shown to be particularly effective in multi-speaker scenarios. In the majority of deep neural network (DNN) based AVSE methods, the audio and visual data are first processed separately using different sub-networks, and then the learned features are fused to utilize the information from both modalities. There have been various studies on suitable audio input features and network architectures, however, to the best of our knowledge, there is no published study that has investigated which visual features are best suited for this specific task. In this work, we perform an empirical study of the most commonly used visual features for DNN based AVSE, the pre-processing requirements for each of these features, and investigate their influence on the performance. Our study shows that despite the overall better performance of embedding-based features, their computationally intensive pre-processing make their use difficult in low resource systems. For such systems, optical flow or raw pixels-based features might be better suited.

This paper presents an audio-visual approach for voice separation which outperforms state-of-theart methods at a low latency in two scenarios: speech and singing voice. The model is based on a two-stage network. Motion cues are obtained with a lightweight graph convolutional network that processes face landmarks. Then, both audio and motion features are fed to an audio-visual transformer which produces a fairly good estimation of the isolated target source. In a second stage, the predominant voice is enhanced with an audio-only network. We present different ablation studies and comparison to state-of-the-art methods. Finally, we explore the transferability of models trained for speech separation in the task of singing voice separation. The demos, code, and weights will be made publicly available at https://ipcv.github.io/VoViT/