CHI ’25, Yokohama, Japan

We propose a unified framework for Singing Voice Synthesis (SVS) and Conversion (SVC), addressing the limitations of existing approaches in cross-domain SVS/SVC, poor output musicality, and scarcity of singing data. Our framework enables control over multiple aspects, including language content based on lyrics, performance attributes based on a musical score, singing style and vocal techniques based on a selector, and voice identity based on a speech sample. The proposed zero-shot learning paradigm consists of one SVS model and two SVC models, utilizing pre-trained content embeddings and a diffusion-based generator. The proposed framework is also trained on mixed datasets comprising both singing and speech audio, allowing singing voice cloning based on speech reference. Experiments show substantial improvements in timbre similarity and musicality over state-of-the-art baselines, providing insights into other low-data music tasks such as instrumental style transfer. Examples can be found at: everyone-can-sing.github.io.

Neural codecs have demonstrated strong performance in high-fidelity compression of audio signals at low bitrates. The token-based representations produced by these codecs have proven particularly useful for generative modeling. While much research has focused on improvements in compression ratio and perceptual transparency, recent works have largely overlooked another desirable codec property – idempotence, the stability of compressed outputs under multiple rounds of encoding. We find that state-of-the-art neural codecs exhibit varied degrees of idempotence, with some degrading audio outputs significantly after as few as three encodings. We investigate possible causes of low idempotence and devise a method for improving idempotence through fine-tuning a codec model. We then examine the effect of idempotence on a simple conditional generative modeling task, and find that increased idempotence can be achieved without negatively impacting downstream modeling performance – potentially extending the usefulness of neural codecs for practical file compression and iterative generative modeling workflows.

Large Vision-Language Models (LVLMs) often produce responses that misalign with factual information, a phenomenon known as hallucinations. While hallucinations are well-studied, the exact causes behind them remain underexplored. In this paper, we first investigate the root causes of hallucinations in LVLMs. Our findings reveal that existing mitigation techniques primarily reduce hallucinations for visual recognition prompts—those that require simple descriptions of visual elements—but fail for cognitive prompts that demand deliberate reasoning. We identify the core issue as a lack of true visual perception in LVLMs: although they can accurately recognize visual elements, they struggle to fully interpret these elements in the context of the input prompt and effectively link this recognition to their internal knowledge, which is critical for reasoning. To address this gap, we introduce Visual Description Grounded Decoding (VDGD), a simple, robust, and training-free method designed to enhance visual perception and improve reasoning capabilities in LVLMs. VDGD works by first generating a detailed description of the image and appending it as a prefix to the instruction. During response generation, tokens are sampled based on their KL divergence to the description, favoring candidates with lower divergence. Experimental results on multiple visual reasoning benchmarks and LVLMs demonstrate that VDGD consistently outperforms existing baselines 2% - 33%. Finally, we introduce VaLLu, a benchmark designed for comprehensive evaluation of the cognitive capabilities of LVLMs.

Diffusion models have demonstrated significant potential in speech synthesis tasks, including text-to-speech (TTS) and voice cloning. However, their iterative denoising processes are computationally intensive, and previous distillation attempts have shown consistent quality degradation. Moreover, existing TTS approaches are limited by non-differentiable components or iterative sampling that prevent true end-to-end optimization with perceptual metrics. We introduce DMOSpeech, a distilled diffusion-based TTS model that uniquely achieves both faster inference and superior performance compared to its teacher model. By enabling direct gradient pathways to all model components, we demonstrate the first successful end-to-end optimization of differentiable metrics in TTS, incorporating Connectionist Temporal Classification (CTC) loss and Speaker Verification (SV) loss. Our comprehensive experiments, validated through extensive human evaluation, show significant improvements in naturalness, intelligibility, and speaker similarity while reducing inference time by orders of magnitude. This work establishes a new framework for aligning speech synthesis with human auditory preferences through direct metric optimization. The audio samples are available at https://dmospeech.github.io/demo

Automatic video commentary systems are widely used on multimedia social media platforms to extract factual information about video content. However, current systems may overlook essential paralinguistic cues, including emotion and attitude, which are critical for fully conveying the meaning of visual content. The absence of these cues can limit user understanding or, in some cases, distort the video’s original intent. Expressive speech effectively conveys these cues and enhances the user’s comprehension of videos. Building on these insights, this paper explores the usage of vision-context-aware expressive speech in enhancing users’ understanding of videos in video commentary systems1. Firstly, our formatting study indicates that semantic-only speech can lead to ambiguity, and misaligned emotions between speech and visuals may distort content interpretation. To address this, we propose a method called vision-context-aware speech synthesis (V-CASS). It analyzes para-linguistic cues from visuals using a vision-language model and leverages a knowledge-infused language model to guide the expressive speech model in generating context-aligned speech. User studies show that V-CASS enhances emotional and attitudinal resonance, as well as user audio-visual understanding and engagement, with 74.68% of participants preferring the system. Finally, we explore the potential of our method in helping blind and low-vision users navigate web videos, improving universal accessibility.

Imagine James Bond speaking like Mr. Bean---such a mismatch would create a jarring dissonance and break the viewer's immersion. Current research on virtual avatar animation has focused on modeling 3D geometry, appearance, motion generation, however, neglecting the harmony between speech prosody and the avatar's visual presentation and contextual environment. In this paper, we seek to bridge this gap by firstly identifying and defining the key elements necessary for achieving audiovisual harmony, such as appearance, expression, body posture, backgrounds and colors. Subsequently, we propose a method that jointly models semantic consistency in avatar animation, named HarmoniVox, specifically on crafting prosodic speech consistent with the avatar's essence from given visual image. To achieve this, we implement a technical framework with a mutual modal contrastive learning strategy, enhancing multimodal alignment in a coarse-to-fine fashion. To support this method, we establish a experimental dataset HarAvaSpeech comprising 28,929 image-audio pairs, designed to encompass expressive speech prosody and rich avatar visual presentations across a wide range of contexts. Leveraging this dataset, our experiments demonstrate that the proposed method outperforms the baselines in manipulating the nuanced tone and harmonious rhythm of speech with the avatar visual presentations, and reveal generalizability on out-of-domain cases. Demo would be provided in https://harmonivox.github.io/harmonivox/.

Audio source separation aims to extract individual sound sources from an audio mixture. Recent studies on source separation focus primarily on minimizing signal-level distance, typically measured by source-to-distortion ratio (SDR). However, scant attention has been given to the perceptual quality of the separated tracks. In this paper, we propose MDX-GAN, an efficient and high-fidelity audio source separator based on MDX-Net for multiple sound classes. We leverage different training objectives to enhance the perceptual quality of audio source separation. Specifically, we adopt perceptually-motivated loss functions on top of the waveform loss, including multi-resolution STFT and Mel-spectrogram losses, and employ the adversarial training paradigm with multi-domain and multi-scale discriminators to refine the perceptual quality of separation. Additionally, we extend the model to support multiple sound classes within a single network via feature-wise linear modulation (FiLM). We conduct both objective and subjective experiments to evaluate MDX-GAN on real-world settings, and assess the impacts of design components on the perceptual quality and SDR scores. Results demonstrate that MDX-GAN accurately separates the sound source and achieves superior perceptual quality.

High-quality speech synthesis models may be used to spread misinformation or impersonate voices. Audio watermarking can combat misuse by embedding a traceable signature in generated audio. However, existing audio watermarks typically demonstrate robustness to only a small set of transformations of the watermarked audio. To address this, we propose MaskMark, a neural network-based digital audio watermarking technique optimized for speech. MaskMark embeds a secret key vector in audio via a multiplicative spectrogram mask, allowing the detection of watermarked speech segments even under substantial signal-processing or neural network-based transformations. Comparisons to a state-of-the-art baseline on natural and synthetic speech corpora and a human subjects evaluation demonstrate MaskMark’s superior robustness in detecting watermarked speech while maintaining high perceptual transparency.

Research in generative self-supervised learning (SSL) has largely focused on local embeddings for tokenized sequences. We introduce a generative SSL framework that learns a global representation that is disentangled from local embeddings. We apply this technique to jointly learn a global speaker embedding and a zero-shot voice converter. The converter modifies recorded speech to sound as if it were spoken by a different person while preserving the content, using only a short reference clip unavailable to the model during training. Listening experiments conducted on an unseen dataset show that our models significantly outperform SOTA baselines in both quality and speaker similarity for various datasets and unseen languages.

Instruction Tuning (IT), the process of training large language models (LLMs) using instruction-response pairs, has emerged as the predominant method for transforming base pre-trained LLMs into open-domain conversational agents. While IT has achieved notable success and widespread adoption, its limitations and shortcomings remain underexplored. In this paper, through rigorous experiments and an in-depth analysis of the changes LLMs undergo through IT, we reveal various limitations of IT. In particular, we show that (1) IT fails to enhance knowledge or skills in LLMs. LoRA fine-tuning is limited to learning response initiation and style tokens, and full-parameter fine-tuning leads to knowledge degradation. (2) Copying response patterns from IT datasets derived from knowledgeable sources leads to a decline in response quality. (3) Full-parameter fine-tuning increases hallucination by inaccurately borrowing tokens from conceptually similar instances in the IT dataset for generating responses. (4) Popular methods to improve IT do not lead to performance improvements over a simple LoRA fine-tuned model. Our findings reveal that responses generated solely from pre-trained knowledge consistently outperform responses by models that learn any form of new knowledge from IT on open-source datasets. We hope the insights and challenges revealed in this paper inspire future work in related directions.

Recent AIGC systems possess the capability to generate digital multimedia content based on human language instructions, such as text, image and video. However, when it comes to speech, existing methods related to human instruction-to-speech generation exhibit two limitations. Firstly, they require the division of inputs into content prompt (transcript) and description prompt (style and speaker), instead of directly supporting human instruction. This division is less natural in form and does not align with other AIGC models. Secondly, the practice of utilizing an independent description prompt to model speech style, without considering the transcript content, restricts the ability to control speech at a fine-grained level. To address these limitations, we propose VoxInstruct, a novel unified multilingual codec language modeling framework that extends traditional text-to-speech tasks into a general human instruction-to-speech task. Our approach enhances the expressiveness of human instruction-guided speech generation and aligns the speech generation paradigm with other modalities. To enable the model to automatically extract the content of synthesized speech from raw text instructions, we introduce speech semantic tokens as an intermediate representation for instruction-to-content guidance. We also incorporate multiple Classifier-Free Guidance (CFG) strategies into our codec language model, which strengthens the generated speech following human instructions. Furthermore, our model architecture and training strategies allow for the simultaneous support of combining speech prompt and descriptive human instruction for expressive speech synthesis, which is a first-of-its-kind attempt. Codes, models and demos are at: https://github.com/thuhcsi/VoxInstruct.

Speech-language multi-modal learning presents a significant challenge due to the fine nuanced information inherent in speech styles. Therefore, a large-scale dataset providing elaborate comprehension of speech style is urgently needed to facilitate insightful interplay between speech audio and natural language. However, constructing such datasets presents a major trade-off between large-scale data collection and high-quality annotation. To tackle this challenge, we propose an automatic speech annotation system for expressiveness interpretation that annotates in-the-wild speech clips with expressive and vivid human language descriptions. Initially, speech audios are processed by a series of expert classifiers and captioning models to capture diverse speech characteristics, followed by a fine-tuned LLaMA for customized annotation generation. Unlike previous tag/templet-based annotation frameworks with limited information and diversity, our system provides in-depth understandings of speech style through tailored natural language descriptions, thereby enabling accurate and voluminous data generation for large model training. With this system, we create SpeechCraft, a fine-grained bilingual expressive speech dataset. It is distinguished by highly descriptive natural language style prompts, containing approximately 2,000 hours of audio data and encompassing over two million speech clips. Extensive experiments demonstrate that the proposed dataset significantly boosts speech-language task performance in stylist speech synthesis and speech style understanding.

Lyrics and music are both significant for a singer to perform a song. Therefore, it is important in singer's motion generation to model both semantic and acoustic correlation with motions at the same time. In this paper, we propose HoloSinger, a novel comprehensive system that synthesizes singing motions according to the given song. Additionally, we present singing avatar with octahedral holographic projection. For singing motion generation, we introduce a Transformer-VAE generative model to decompose lyrics and music, then fuse their impacts to synthesize singer's motions. Extensive experiments and user studies show that our method automatically generates realistic motions that adhere to musical choreography and reflect the lyric semantics appropriately. Furthermore, we design a desktop-level holographic projection device with an octahedral structure. It achieves high-definition holographic projection effects with smaller volume, larger imaging area ratio, and the ability of real-time AI interaction.

Consumer-grade music recordings such as those captured by mobile devices typically contain distortions in the form of background noise, reverb, and microphone-induced EQ. This paper presents a deep learning approach to enhance low-quality music recordings by combining (i) an image-to-image translation model for manipulating audio in its mel-spectrogram representation and (ii) a music vocoding model for mapping synthetically generated mel-spectrograms to perceptually realistic waveforms. We find that this approach to music enhancement outperforms baselines which use classical methods for mel-spectrogram inversion and an end-to-end approach directly mapping noisy waveforms to clean waveforms. Additionally, in evaluating the proposed method with a listening test, we analyze the reliability of common audio enhancement evaluation metrics when used in the music domain.

Automatic speech quality assessment remains challenging, as we lack complete models of human auditory perception. Many existing full-reference models correlate well with human perception, but cannot be used in real-world scenarios where ground truth clean reference recordings are not available. On the other hand no-reference metrics typically suffer from several shortcomings, such as lack of robustness to unseen perturbations and reliance on (limited) labeled data for training. Moreover, noise or large variance among the labels makes it difficult to learn generalizable representations, especially for recordings with subtle differences. This paper proposes a learning framework for estimating the quality of a recording without any reference, and without any human judgments. The main component of this framework is a pairwise quality-preference strategy that reduces label noise, thereby making learning more robust. From pairwise preferences, we first learn a content invariant quality ordering; and then we re-target the model to predict quality on an absolute scale. We show that the resulting learned metric is well-calibrated with human judgments. Since it is a deep network, the metric is differentiable, making it suitable as a loss function for downstream tasks. For example, we show that adding this metric to an existing speech enhancement method yields significant improvement.

Speech representation learning transforms speech into features that are suitable for downstream tasks, e.g. speech recognition, phoneme classification, or speaker identification. For such recognition tasks, a representation can be lossy (non-invertible), which is typical of BERT-like self-supervised models. However, when used for synthesis tasks, we find these lossy representations prove to be insufficient to plausibly reconstruct the input signal. This paper introduces a method for invertible and controllable speech representation learning based on disentanglement. The representation can be decoded into a signal perceptually identical to the original. Moreover, its disentangled components (content, pitch, speaker identity, and energy) can be controlled independently to alter the synthesis result. Our model builds upon a zero-shot voice conversion model AutoVC-F0, in which we introduce alteration invariant content loss (AIC loss) and adversarial training (GAN). Through objective measures and subjective tests, we show that our formulation offers significant improvement in voice conversion sound quality as well as more precise control over the disentangled features.

Many audio processing tasks require perceptual assessment.However, the time and expense of obtaining "gold standard" human judgments limit the availability of such data.Most applications incorporate full reference or other similarity-based metrics (e.g.PESQ) that depend on a clean reference.Researchers have relied on such metrics to evaluate and compare various proposed methods, often concluding that small, measured differences imply one is more effective than another.This paper demonstrates several practical scenarios where similarity metrics fail to agree with human perception, because they: (1) vary with clean references; (2) rely on attributes that humans factor out when considering quality, and (3) are sensitive to imperceptible signal level differences.In those scenarios, we show that no-reference metrics do not suffer from such shortcomings and correlate better with human perception.We conclude therefore that similarity serves as an unreliable proxy for audio quality.

Stylized pattern generation is challenging and has received increasing attention in recent studies. However, it requires further exploration in pattern generation that matches the given scenes. This paper proposes a demonstration that automatically generates carpet patterns with input home scenes and other user preferences. Besides meeting the individual needs of users and providing highly editable output, the critical challenge of the system is to make the output pattern coordinate with the input home scene, which distinguishes our approach from others. The carpets generated by the system also permit easy modification or extension to various interior design styles.

Following the prevalence of short-form video, short-form voice content has emerged on social media platforms like Twitter and Facebook. A challenge that creators face is hard constraints on the content length. If the initial recording is not short enough, they need to re-record or edit their content. Both are time-consuming, and the latter, if supported, can have a learning curve. Moreover, creators need to manually create multiple versions to publish content on platforms with different length constraints. To simplify this process, we present ROPE1 (Record Once, Post Everywhere). Creators can record voice content once, and our system will automatically shorten it to all length limits by removing parts of the recording for each target. We formulate this as a combinatorial optimization problem and propose a novel algorithm that automatically selects optimal sentence combinations from the original content to comply with each length constraint. Creators can customize the algorithmically shortened content by specifying sentences to include or exclude. Our system can also use the user-specified constraints to recompute and provides a new version. We conducted a user study comparing ROPE with a sentence-based manual editing baseline. The results show that ROPE can generate high-quality edits, alleviating the cognitive loads of creators for shortening content. While our system and user study address short-form voice content specifically, we believe that the same concept can also be applied to other media such as video with narration and dialog.

In this paper, we consider the large-scale precoding matrix indicator (PMI) selection problem at the receiver in wireless communications. The selection is based on the channel capacity of the PMI matrix in a pre-designed codebook. The quality of the PMI matrix is essential in achieving higher spectral efficiency. We first derive two novel formulations including a partial permutation-matrix model and an indicator-vector model for the original problem. The discrete constraints in the formulations make the problem NP-hard. Then we propose a stochastic projected gradient method augmented by block coordinate descent under various strategies. We show that the algorithms terminate in finite steps and produce sufficient descent at each iteration when the step size is chosen properly. Extensive experiments demonstrate that our proposed algorithms are able to find better PMI matrices more efficiently compared to the existing methods.

Many speech processing methods based on deep learning require an automatic and differentiable audio metric for the loss function. The DPAM approach of Manocha et al. [1] learns a full-reference metric trained directly on human judgments, and thus correlates well with human perception. However, it requires a large number of human annotations and does not generalize well outside the range of perturbations on which it was trained. This paper introduces CDPAM –a metric that builds on and advances DPAM. The primary improvement is to combine contrastive learning and multi-dimensional representations to build robust models from limited data. In addition, we collect human judgments on triplet comparisons to improve generalization to a broader range of audio perturbations. CDPAM correlates well with human responses across nine varied datasets. We also show that adding this metric to existing speech synthesis and enhancement methods yields significant improvement, as measured by objective and subjective tests.

Text-based speech editors expedite the process of editing speech recordings by permitting editing via intuitive cut, copy, and paste operations on a speech transcript. A major drawback of current systems, however, is that edited recordings often sound unnatural because of prosody mismatches around edited regions. In our work, we propose a new context-aware method for more natural sounding text-based editing of speech. To do so, we 1) use a series of neural networks to generate salient prosody features that are dependent on the prosody of speech surrounding the edit and amenable to fine-grained user control 2) use the generated features to control a standard pitch-shift and time-stretch method and 3) apply a denoising neural network to remove artifacts induced by the signal manipulation to yield a high-fidelity result. We evaluate our approach using a subjective listening test, provide a detailed comparative analysis, and conclude several interesting insights.

Speech generation and enhancement have seen recent breakthroughs in quality thanks to deep learning. These methods typically operate at a limited sampling rate of 16-22kHz due to computational complexity and available datasets. This limitation imposes a gap between the output of such methods and that of high-fidelity (≥44kHz) real-world audio applications. This paper proposes a new bandwidth extension (BWE) method that expands 8-16kHz speech signals to 48kHz. The method is based on a feed-forward WaveNet architecture trained with a GAN-based deep feature loss. A mean-opinion-score (MOS) experiment shows significant improvement in quality over state-of-the-art BWE methods. An AB test reveals that our 16-to-48kHz BWE is able to achieve fidelity that is typically indistinguishable from real high-fidelity recordings. We use our method to enhance the output of recent speech generation and denoising methods, and experiments demonstrate significant improvement in sound quality over these baselines. We propose this as a general approach to narrow the gap between generated speech and recorded speech, without the need to adapt such methods to higher sampling rates.

Short online videos have become the dominating media on social platforms. However, finding suitable music to accompany videos can be a challenging task to some video creators, due to copyright constraints, limitations in search engines, and required audio-editing expertise. One possible solution to these problems is to use AI music generation. In this paper we present a user interface (UI) paradigm that allows users to input a song to an AI engine and then interactively regenerate and mix AI-generated music. To arrive at this design, we conducted user studies with a total of 104 video creators at several stages of our design and development process. User studies supported the effectiveness of our approach and provided valuable insights about human-AI interaction as well as the design and evaluation of mixed-initiative interfaces in creative practice.

Music similarity search is useful for a variety of creative tasks such as replacing one music recording with another recording with a similar "feel", a common task in video editing. For this task, it is typically necessary to define a similarity metric to compare one recording to another. Music similarity, however, is hard to define and depends on multiple simultaneous notions of similarity (i.e. genre, mood, instrument, tempo). While prior work ignore this issue, we embrace this idea and introduce the concept of multidimensional similarity and unify both global and specialized similarity metrics into a single, semantically disentangled multidimensional similarity metric. To do so, we adapt a variant of deep metric learning called conditional similarity networks to the audio domain and extend it using track-based information to control the specificity of our model. We evaluate our method and show that our single, multidimensional model outperforms both specialized similarity spaces and alternative baselines. We also run a user-study and show that our approach is favored by human annotators as well.

Non-parallel many-to-many voice conversion remains an interesting but challenging speech processing task. Many style-transfer-inspired methods such as generative adversarial networks (GANs) and variational autoencoders (VAEs) have been proposed. Recently, AutoVC, a conditional autoencoders (CAEs) based method achieved state-of-the-art results by disentangling the speaker identity and speech content using information-constraining bottlenecks, and it achieves zero-shot conversion by swapping in a different speaker's identity embedding to synthesize a new voice. However, we found that while speaker identity is disentangled from speech content, a significant amount of prosodic information, such as source F0, leaks through the bottleneck, causing target F0 to fluctuate unnaturally. Furthermore, AutoVC has no control of the converted F0 and thus unsuitable for many applications. In the paper, we modified and improved autoencoder-based voice conversion to disentangle content, F0, and speaker identity at the same time. Therefore, we can control the F0 contour, generate speech with F0 consistent with the target speaker, and significantly improve quality and similarity. We support our improvement through quantitative and qualitative analysis.

The goal of acoustic matching is to transform an audio recording made in one acoustic environment to sound as if it had been recorded in a different environment, based on reference audio from the target environment. This paper introduces a deep learning solution for two parts of the acoustic matching problem. First, we characterize acoustic environments by mapping audio into a low-dimensional embedding invariant to speech content and speaker identity. Next, a waveform-to-waveform neural network conditioned on this embedding learns to transform an input waveform to match the acoustic qualities encoded in the target embedding. Listening tests on both simulated and real environments show that the proposed approach improves on state-of-the-art baseline methods.

Many audio processing tasks require perceptual assessment.The "gold standard" of obtaining human judgments is timeconsuming, expensive, and cannot be used as an optimization criterion.On the other hand, automated metrics are efficient to compute but often correlate poorly with human judgment, particularly for audio differences at the threshold of human detection.In this work, we construct a metric by fitting a deep neural network to a new large dataset of crowdsourced human judgments.Subjects are prompted to answer a straightforward, objective question: are two recordings identical or not?These pairs are algorithmically generated under a variety of perturbations, including noise, reverb, and compression artifacts; the perturbation space is probed with the goal of efficiently identifying the just-noticeable difference (JND) level of the subject.We show that the resulting learned metric is well-calibrated with human judgments, outperforming baseline methods.Since it is a deep network, the metric is differentiable, making it suitable as a loss function for other tasks.Thus, simply replacing an existing loss (e.g., deep feature loss) with our metric yields significant improvement in a denoising network, as measured by subjective pairwise comparison.

Speech synthesis has recently seen significant improvements in fidelity, driven by the advent of neural vocoders and neural prosody generators. However, these systems lack intuitive user controls over prosody, making them unable to rectify prosody errors (e.g., misplaced emphases and contextually inappropriate emotions) or generate prosodies with diverse speaker excitement levels and emotions. We address these limitations with a user-controllable, context-aware neural prosody generator. Given a real or synthesized speech recording, our model allows a user to input prosody constraints for certain time frames and generates the remaining time frames from input text and contextual prosody. We also propose a pitch-shifting neural vocoder to modify input speech to match the synthesized prosody. Through objective and subjective evaluations we show that we can successfully incorporate user control into our prosody generation model without sacrificing the overall naturalness of the synthesized speech.

Real-world audio recordings are often degraded by factors such as noise, reverberation, and equalization distortion.This paper introduces HiFi-GAN, a deep learning method to transform recorded speech to sound as though it had been recorded in a studio.We use an end-to-end feed-forward WaveNet architecture, trained with multi-scale adversarial discriminators in both the time domain and the time-frequency domain.It relies on the deep feature matching losses of the discriminators to improve the perceptual quality of enhanced speech.The proposed model generalizes well to new speakers, new speech content, and new environments.It significantly outperforms state-of-the-art baseline methods in both objective and subjective experiments.

An artist faces two challenges when creating a 2D animated character to mimic a specific human performance. First, the artist must design and draw a collection of artwork depicting portions of the character in a suitable set of poses, for example arm and hand poses that can be selected and combined to express the range of gestures typical for that person. Next, to depict a specific performance, the artist must select and position the appropriate set of artwork at each moment of the animation. This paper presents a system that addresses these challenges by leveraging video of the target human performer. Our system tracks arm and hand poses in an example video of the target. The UI displays clusters of these poses to help artists select representative poses that capture the actor's style and personality. From this mapping of pose data to character artwork, our system can generate an animation from a new performance video. It relies on a dynamic programming algorithm to optimize for smooth animations that match the poses found in the video. Artists used our system to create four 2D characters and were pleased with the final automatically animated results. We also describe additional applications addressing audio-driven or text-based animations.

We introduce a perceptually motivated approach to bandwidth expansion for speech. Our method pairs a new 3-way split variant of the FFTNet neural vocoder structure with a perceptual loss function, combining objectives from both the time and frequency domains. Mean opinion score tests show that it outperforms baseline methods from both domains, even for extreme bandwidth expansion.

This paper introduces a deep learning approach to enhance speech recordings made in a specific environment. A single neural network learns to ameliorate several types of recording artifacts, including noise, reverberation, and non-linear equalization. The method relies on a new perceptual loss function that combines adversarial loss with spectrogram features. Both subjective and objective evaluations show that the proposed approach improves on state-of-the-art baseline methods.

Editing talking-head video to change the speech content or to remove filler words is challenging. We propose a novel method to edit talking-head video based on its transcript to produce a realistic output video in which the dialogue of the speaker has been modified, while maintaining a seamless audio-visual flow (i.e. no jump cuts). Our method automatically annotates an input talking-head video with phonemes, visemes, 3D face pose and geometry, reflectance, expression and scene illumination per frame. To edit a video, the user has to only edit the transcript, and an optimization strategy then chooses segments of the input corpus as base material. The annotated parameters corresponding to the selected segments are seamlessly stitched together and used to produce an intermediate video representation in which the lower half of the face is rendered with a parametric face model. Finally, a recurrent video generation network transforms this representation to a photorealistic video that matches the edited transcript. We demonstrate a large variety of edits, such as the addition, removal, and alteration of words, as well as convincing language translation and full sentence synthesis.

We introduce FFTNet, a deep learning approach synthesizing audio waveforms. Our approach builds on the recent WaveNet project, which showed that it was possible to synthesize a natural sounding audio waveform directly from a deep convolutional neural network. FFTNet offers two improvements over WaveNet. First it is substantially faster, allowing for real-time synthesis of audio waveforms. Second, when used as a vocoder, the resulting speech sounds more natural, as measured via a “mean opinion score” test.

Editing audio narration using conventional software typically involves many painstaking low-level manipulations. Some state of the art systems allow the editor to work in a text transcript of the narration, and perform select, cut, copy and paste operations directly in the transcript; these operations are then automatically applied to the waveform in a straightforward manner. However, an obvious gap in the text-based interface is the ability to type new words not appearing in the transcript, for example inserting a new word for emphasis or replacing a misspoken word. While high-quality voice synthesizers exist today, the challenge is to synthesize the new word in a voice that matches the rest of the narration. This paper presents a system that can synthesize a new word or short phrase such that it blends seamlessly in the context of the existing narration. Our approach is to use a text to speech synthesizer to say the word in a generic voice, and then use voice conversion to convert it into a voice that matches the narration. Offering a range of degrees of control to the editor, our interface supports fully automatic synthesis, selection among a candidate set of alternative pronunciations, fine control over edit placements and pitch profiles, and even guidance by the editors own voice. The paper presents studies showing that the output of our method is preferred over baseline methods and often indistinguishable from the original voice.

State-of-the art voice conversion methods re-synthesize voice from spectral representations such as MFCCs and STRAIGHT, thereby introducing muffled artifacts. We propose a method that circumvents this concern using concatenative synthesis coupled with exemplar-based unit selection. Given parallel speech from source and target speakers as well as a new query from the source, our method stitches together pieces of the target voice. It optimizes for three goals: matching the query, using long consecutive segments, and smooth transitions between the segments. To achieve these goals, we perform unit selection at the frame level and introduce triphone-based preselection that greatly reduces computation and enforces selection of long, contiguous pieces. Our experiments show that the proposed method has better quality than baseline methods, while preserving high individuality.