Accurate three-dimensional (3D) human motion tracking is essential for immersive augmented reality$(A R)$and virtual reality (VR) applications, allowing users to engage with virtual environments through realistic full-body avatars. Achieving this level of detail, however, is challenging when the driving signals are sparse, typically coming only from upper-body sensors, such as head-mounted devices and hand controllers. To address this challenge, we propose EgoMDM (Egocentric Motion Diffusion Model), an end-to-end diffusion-based framework designed to reconstruct full-body motion from sparse tracking signals. EgoMDM models human motion in a conditional autoregressive manner using a unidirectional recurrent neural network, making it well-suited for real-time applications. By embedding local-to-global translation, forward and inverse kinematics, and foot-contact detection within the diffusion framework, EgoMDM achieves seamless, end-to-end motion synthesis, effectively reducing artifacts like foot sliding and ground penetration. Additionally, EgoMDM is conditioned on the user's body scale, allowing it to generalize across a diverse population and produce consistent avatar shapes over time. In our extensive experiments on the AMASS motion capture dataset, EgoMDM achieves state-of-the-art performance in both motion tracking accuracy and synthesis quality, demonstrating its robustness and adaptability across various human motion scenarios. Furthermore, EgoMDM significantly outperforms the existing models when tested on real signal inputs, highlighting its robustness and applicability to the real-world data. See the project page at: https://yohanshin.github.io/egomdm.github.io/

We introduce SoundVista, a method to generate the ambient sound of an arbitrary scene at novel viewpoints. Given a pre-acquired recording of the scene from sparsely distributed microphones, SoundVista can synthesize the sound of that scene from an unseen target viewpoint. The method learns the underlying acoustic transfer function that relates the signals acquired at the distributed microphones to the signal at the target viewpoint, using a limited number of known recordings. Unlike existing works, our method does not require constraints or prior knowledge of sound source details. Moreover, our method efficiently adapts to diverse room layouts, reference microphone configurations and unseen environments. To enable this, we introduce a visual-acoustic binding module that learns visual embeddings linked with local acoustic properties from panoramic RGB and depth data. We first leverage these embeddings to optimize the placement of reference microphones in any given scene. During synthesis, we leverage multiple embeddings extracted from reference locations to get adaptive weights for their contribution, conditioned on target viewpoint. We benchmark the task on both publicly available data and real-world settings. We demonstrate significant improvements over existing methods.

We present REWIND (Real-Time Egocentric Whole-Body Motion Diffusion), a one-step diffusion model for real-time, high-fidelity human motion estimation from egocentric image inputs. While an existing method for egocentric whole-body (i.e., body and hands) motion estimation is non-real-time and acausal due to diffusion-based iterative motion refinement to capture correlations between body and hand poses, REWIND operates in a fully causal and real-time manner. To enable real-time inference, we introduce (1) cascaded body-hand denoising diffusion, which effectively models the correlation between egocentric body and hand motions in a fast, feed-forward manner, and (2) diffusion distillation, which enables high-quality motion estimation with a single denoising step. Our denoising diffusion model is based on a modified Transformer architecture, designed to causally model output motions while enhancing generalizability to unseen motion lengths. Additionally, REWIND optionally supports identity-conditioned motion estimation when identity prior is available. To this end, we propose a novel identity conditioning method based on a small set of pose exemplars of the target identity, which further enhances motion estimation quality. Through extensive experiments, we demonstrate that REWIND significantly outperforms the existing baselines both with and without exemplar-based identity conditioning.

This paper introduces a novel neural network model for rendering binaural audio directly from ambisonic recordings. We optimized the model end-to-end to learn a direct mapping between ambisonic and binaural signals. Our approach eliminates traditional processing steps that were required to mitigate artifacts due to spherical harmonic order truncation and spatial aliasing, as well as other complex filtering needed to compensate for near-field sound sources. To showcase the advantage of neural network-based rendering over traditional signal processing approaches, we introduce a new dataset that includes challenging near-field sound sources, including speech and background noises. We demonstrate that our model can produce binaural audio results that closely match the fidelity of ground truth binaural recordings. Our comprehensive validation shows that the proposed method outperforms existing methods on several error metrics as well as in subjective evaluations. Model code, demos and datasets are available on our project webpage.

Neural audio codecs have been widely adopted in audio-generative tasks because their compact and discrete representations are suitable for both large-language-model-style and regression-based generative models. However, most neural codecs struggle to model out-of-domain audio, resulting in error propagations to downstream generative tasks. In this paper, we first argue that information loss from codec compression degrades out-of-domain robustness. Then, we propose full-band 48 kHz ComplexDec with complex spectral input and output to ease the information loss while adopting the same 24 kbps bitrate as the baseline AuidoDec and ScoreDec. Objective and subjective evaluations demonstrate the out-of-domain robustness of ComplexDec trained using only the 30-hour VCTK corpus.

We introduce AV-Flow, an audio-visual generative model that animates photo-realistic 4D talking avatars given only text input. In contrast to prior work that assumes an existing speech signal, we synthesize speech and vision jointly. We demonstrate human-like speech synthesis, synchronized lip motion, lively facial expressions and head pose; all generated from just text characters. The core premise of our approach lies in the architecture of our two parallel diffusion transformers. Intermediate highway connections ensure communication between the audio and visual modalities, and thus, synchronized speech intonation and facial dynamics (e.g., eyebrow motion). Our model is trained with flow matching, leading to expressive results and fast inference. In case of dyadic conversations, AV-Flow produces an always-on avatar, that actively listens and reacts to the audio-visual input of a user. Through extensive experiments, we show that our method outperforms prior work, synthesizing natural-looking 4D talking avatars. Project page: https://aggelinacha.github.io/AV-Flow/

We propose FlowDec, a neural full-band audio codec for general audio sampled at 48 kHz that combines non-adversarial codec training with a stochastic postfilter based on a novel conditional flow matching method. Compared to the prior work ScoreDec which is based on score matching, we generalize from speech to general audio and move from 24 kbit/s to as low as 4 kbit/s, while improving output quality and reducing the required postfilter DNN evaluations from 60 to 6 without any fine-tuning or distillation techniques. We provide theoretical insights and geometric intuitions for our approach in comparison to ScoreDec as well as another recent work that uses flow matching, and conduct ablation studies on our proposed components. We show that FlowDec is a competitive alternative to the recent GAN-dominated stream of neural codecs, achieving FAD scores better than those of the established GAN-based codec DAC and listening test scores that are on par, and producing qualitatively more natural reconstructions for speech and harmonic structures in music.

Binaural rendering aims to synthesize binaural audio that mimics natural hearing based on a mono audio and the locations of the speaker and listener. Although many methods have been proposed to solve this problem, they struggle with rendering quality and streamable inference. Synthesizing high-quality binaural audio that is indistinguishable from real-world recordings requires precise modeling of binaural cues, room reverb, and ambient sounds. Additionally, real-world applications demand streaming inference. To address these challenges, we propose a flow matching based streaming binaural speech synthesis framework called BinauralFlow. We consider binaural rendering to be a generation problem rather than a regression problem and design a conditional flow matching model to render high-quality audio. Moreover, we design a causal U-Net architecture that estimates the current audio frame solely based on past information to tailor generative models for streaming inference. Finally, we introduce a continuous inference pipeline incorporating streaming STFT/ISTFT operations, a buffer bank, a midpoint solver, and an early skip schedule to improve rendering continuity and speed. Quantitative and qualitative evaluations demonstrate the superiority of our method over SOTA approaches. A perceptual study further reveals that our model is nearly indistinguishable from real-world recordings, with a 42% confusion rate.

We present an audio-driven real-time system for animating photorealistic 3D facial avatars with minimal latency, designed for social interactions in virtual reality for anyone. Central to our approach is an encoder model that transforms audio signals into latent facial expression sequences in real time, which are then decoded as photorealistic 3D facial avatars. Leveraging the generative capabilities of diffusion models, we capture the rich spectrum of facial expressions necessary for natural communication while achieving real-time performance (<15ms GPU time). Our novel architecture minimizes latency through two key innovations: an online transformer that eliminates dependency on future inputs and a distillation pipeline that accelerates iterative denoising into a single step. We further address critical design challenges in live scenarios for processing continuous audio signals frame-by-frame while maintaining consistent animation quality. The versatility of our framework extends to multimodal applications, including semantic modalities such as emotion conditions and multimodal sensors with head-mounted eye cameras on VR headsets. Experimental results demonstrate significant improvements in facial animation accuracy over existing offline state-of-the-art baselines, achieving 100 to 1000 × faster inference speed. We validate our approach through live VR demonstrations and across various scenarios such as multilingual speeches.

We present a new dataset called Real Acoustic Fields (RAF) that captures real acoustic room data from multiple modali-ties. The dataset includes high-quality and densely captured room impulse response data paired with multi-view images, and precise 6DoF pose tracking data for sound emitters and listeners in the rooms. We used this dataset to evaluate existing methods for novel-view acoustic synthesis and impulse re-sponse generation which previously relied on synthetic data. In our evaluation, we thoroughly assessed existing audio and audio- visual models against multiple criteria and proposed settings to enhance their performance on real-world data. We also conducted experiments to investigate the impact of incorporating visual data (i.e., images and depth) into neu-ral acoustic field models. Additionally, we demonstrated the effectiveness of a simple sim2real approach, where a model is pre-trained with simulated data and fine-tuned with sparse real-world data, resulting in significant improvements in the few-shot learning approach. RAF is the first dataset to provide densely captured room acoustic data, making it an ideal resource for researchers working on audio and audio-visual neural acoustic field modeling techniques. Demos and datasets are available on our project page.

We present a framework for generating full-bodied pho-torealistic avatars that gesture according to the conversational dynamics of a dyadic interaction. Given speech au-dio, we output multiple possibilities of gestural motion for an individual, including face, body, and hands. The key be-hind our method is in combining the benefits of sample di-versity from vector quantization with the high-frequency de-tails obtained through diffusion to generate more dynamic, expressive motion. We visualize the generated motion using highly photorealistic avatars that can express crucial nu-ances in gestures (e.g. sneers and smirks). To facilitate this line of research, we introduce a first-of-its-kind multi-view conversational dataset that allows for photorealistic reconstruction. Experiments show our model generates appropri-ate and diverse gestures, outperforming both diffusion- and VQ-only methods. Furthermore, our perceptual evaluation highlights the importance of photorealism (vs. meshes) in accurately assessing subtle motion details in conversational gestures. Code and dataset available on project page.

Although recent mainstream waveform-domain end-to-end (E2E) neural audio codecs achieve impressive coded audio quality with a very low bitrate, the quality gap between the coded and natural audio is still significant. A generative adversarial network (GAN) training is usually required for these E2E neural codecs because of the difficulty of direct phase modeling. However, such adversarial learning hinders these codecs from preserving the original phase information. To achieve human-level naturalness with a reasonable bitrate, preserve the original phase, and get rid of the tricky and opaque GAN training, we develop a score-based diffusion post-filter (SPF) in the complex spectral domain and combine our previous AudioDec with the SPF to propose ScoreDec, which can be trained using only spectral and score-matching losses. Both the objective and subjective experimental results show that ScoreDec with a 24 kbps bitrate encodes and decodes full-band 48 kHz speech with human-level naturalness and well-preserved phase information.

We introduce the novel-view acoustic synthesis (NVAS) task: given the sight and sound observed at a source view-point, can we synthesize the sound of that scene from an unseen target viewpoint? We propose a neural rendering approach: Visually-Guided Acoustic Synthesis (ViGAS) network that learns to synthesize the sound of an arbitrary point in space by analyzing the input audio-visual cues. To benchmark this task, we collect two first-of-their-kind large-scale multi-view audio-visual datasets, one synthetic and one real. We show that our model successfully reasons about the spatial cues and synthesizes faithful audio on both datasets. To our knowledge, this work represents the very first formulation, dataset, and approach to solve the novel-view acoustic synthesis task, which has exciting potential applications ranging from AR/VR to art and design. Unlocked by this work, we believe that the future of novel-view synthesis is in multi-modal learning from videos.

We propose NORD: a novel framework for estimating the relative depth between two binaural speech recordings. In contrast to existing depth estimation techniques, ours only requires audio signals as input. We trained the framework to solve depth preference (i.e. which input perceptually sounds closer to the listener’s head), and quantification tasks (i.e. quantifying the depth difference between the inputs). In addition, training leverages recent advances in metric and multi-task learning, which allows the framework to be invariant to both signal content (i.e. non-matched reference) and directional cues (i.e. azimuth and elevation). Our framework has additional useful qualities that make it suitable for use as an objective metric to benchmark binaural audio systems, particularly depth perception and sound externalization, which we demonstrate through experiments. We also show that NORD generalizes well under different reverberation and environments. The results from preference and quantification tasks correlate well with measured results.

A good audio codec for live applications such as telecommunication is characterized by three key properties: (1) compression, i.e. the bitrate that is required to transmit the signal should be as low as possible; (2) latency, i.e. encoding and decoding the signal needs to be fast enough to enable communication without or with only minimal noticeable delay; and (3) reconstruction quality of the signal. In this work, we propose an open-source, streamable, and real-time neural audio codec that achieves strong performance along all three axes: it can reconstruct highly natural sounding 48 kHz speech signals while operating at only 12 kbps and running with less than 6 ms (GPU)/10 ms (CPU) latency. An efficient training paradigm is also demonstrated for developing such neural audio codecs for real-world scenarios. Both objective and subjective evaluations using the VCTK corpus are provided. To sum up, AudioDec is a well-developed plug-and-play benchmark for audio codec applications.

While 3D human body modeling has received much attention in computer vision, modeling the acoustic equivalent, i.e. modeling 3D spatial audio produced by body motion and speech, has fallen short in the community. To close this gap, we present a model that can generate accurate 3D spatial audio for full human bodies. The system consumes, as input, audio signals from headset microphones and body pose, and produces, as output, a 3D sound field surrounding the transmitter's body, from which spatial audio can be rendered at any arbitrary position in the 3D space. We collect a first-of-its-kind multimodal dataset of human bodies, recorded with multiple cameras and a spherical array of 345 microphones. In an empirical evaluation, we demonstrate that our model can produce accurate body-induced sound fields when trained with a suitable loss. Dataset and code are available online.

Since facial actions such as lip movements contain significant information about speech content, it is not surprising that audio-visual speech enhancement methods are more accurate than their audio-only counterparts. Yet, state-of-the-art approaches still struggle to generate clean, realistic speech without noise artifacts and unnatural distortions in challenging acoustic environments. In this paper, we propose a novel audio-visual speech enhancement framework for high-fidelity telecommunications in AR/VR. Our approach leverages audio-visual speech cues to generate the codes of a neural speech codec, enabling efficient synthesis of clean, realistic speech from noisy signals. Given the importance of speaker-specific cues in speech, we focus on developing personalized models that work well for individual speakers. We demonstrate the efficacy of our approach on a new audio-visual speech dataset collected in an unconstrained, large vocabulary setting, as well as existing audio-visual datasets, outperforming speech enhancement baselines on both quantitative metrics and human evaluation studies. Please see the supplemental video for qualitative results11https://github.com/facebookresearch/facestar/releases/download/paper_materials/video.mp4.

Speech enhancement is a critical component of many user-oriented audio applications, yet current systems still suffer from distorted and unnatural outputs. While generative models have shown strong potential in speech synthesis, they are still lagging behind in speech enhancement. This work leverages recent advances in diffusion probabilistic models, and proposes a novel speech enhancement algorithm that incorporates characteristics of the observed noisy speech signal into the diffusion and reverse processes. More specifically, we propose a generalized formulation of the diffusion probabilistic model named conditional diffusion probabilistic model that, in its reverse process, can adapt to non-Gaussian real noises in the estimated speech signal. In our experiments, we demonstrate strong performance of the proposed approach compared to representative generative models, and investigate the generalization capability of our models to other datasets with noise characteristics unseen during training.

Impulse response estimation in high noise and in-the-wild settings, with minimal control of the underlying data distributions, is a challenging problem. We propose a novel framework for parameterizing and estimating impulse responses based on recent advances in neural representation learning. Our framework is driven by a carefully designed neural network that jointly estimates the impulse response and the (apriori unknown) spectral noise characteristics of an observed signal given the source signal. We demonstrate robustness in estimation, even under low signal-to-noise ratios, and show strong results when learning from spatio-temporal real-world speech data. Our framework provides a natural way to interpolate impulse responses on a spatial grid, while also allowing for efficiently com-pressing and storing them for real-time rendering applications in augmented and virtual reality.

In this work, we present an end-to-end binaural speech synthesis system that combines a low-bitrate audio codec with a powerful binaural decoder that is capable of accurate speech binauralization while faithfully reconstructing environmental factors like ambient noise or reverb.The network is a modified vectorquantized variational autoencoder, trained with several carefully designed objectives, including an adversarial loss.We evaluate the proposed system on an internal binaural dataset with objective metrics and a perceptual study.Results show that the proposed approach matches the ground truth data more closely than previous methods.In particular, we demonstrate the capability of the adversarial loss in capturing environment effects needed to create an authentic auditory scene.

An introductory course on distributed systems typically exposes students to some basic synchronization distributed algorithms. This is often the first exposure for these students to the topic of distributed algorithms. In addition, in a systems course, these algorithms are typically covered in an informal way, avoiding proofs of correctness or complexity analysis. Hence, this first exposure can be challenging to students. Visualization of these algorithms can help alleviate some of these challenges. We present a suite to visualize six basic algorithms on total ordering, critical sections, and leader election.

Estimation of accurate head-related transfer functions (HRTFs) is crucial to achieve realistic binaural acoustic experiences. HRTFs depend on source/listener locations and are therefore expensive and cumbersome to measure; traditional approaches require listener-dependent measurements of HRTFs at thousands of distinct spatial directions in an anechoic chamber. In this work, we present a data-driven approach to learn HRTFs implicitly with a neural network that achieves state of the art results compared to traditional approaches but relies on a much simpler data capture that can be performed in arbitrary, non-anechoic rooms. Despite that simpler and less acoustically ideal data capture, our deep learning based approach learns HRTF of high quality. We show in a perceptual study that the produced binaural audio is ranked on par with traditional DSP approaches by humans and illustrate that interaural time differences (ITDs), interaural level differences (ILDs) and spectral clues are accurately estimated.

This paper presents a generic method for generating full facial 3D animation from speech. Existing approaches to audio-driven facial animation exhibit uncanny or static upper face animation, fail to produce accurate and plausible co-articulation or rely on person-specific models that limit their scalability. To improve upon existing models, we propose a generic audio-driven facial animation approach that achieves highly realistic motion synthesis results for the entire face. At the core of our approach is a categorical latent space for facial animation that disentangles audio-correlated and audio-uncorrelated information based on a novel cross-modality loss. Our approach ensures highly accurate lip motion, while also synthesizing plausible animation of the parts of the face that are uncorrelated to the audio signal, such as eye blinks and eye brow motion. We demonstrate that our approach outperforms several baselines and obtains state-of-the-art quality both qualitatively and quantitatively. A perceptual user study demonstrates that our approach is deemed more realistic than the current state-of-the-art in over 75% of cases. We recommend watching the supplemental video before reading the paper: https://github.com/facebookresearch/meshtalk

Codec Avatars are a recent class of learned, photorealistic face models that accurately represent the geometry and texture of a person in 3D (i.e., for virtual reality), and are almost indistinguishable from video [28]. In this paper we describe the first approach to animate these parametric models in real-time which could be deployed on commodity virtual reality hardware using audio and/or eye tracking. Our goal is to display expressive conversations between individuals that exhibit important social signals such as laughter and excitement solely from la-tent cues in our lossy input signals. To this end we collected over 5 hours of high frame rate 3D face scans across three participants including traditional neutral speech as well as expressive and conversational speech. We investigate a multimodal fusion approach that dynamically identifies which sensor encoding should animate which parts of the face at any time. See the supplemental video which demonstrates our ability to generate full face motion far beyond the typically neutral lip articulations seen in competing work: https://research.fb.com/videos/audio-and-gaze-driven-facial-animation-of-codec-avatars/.

Action recognition has become a rapidly developing research field within the last decade. But with the increasing demand for large scale data, the need of hand annotated data for the training becomes more and more impractical. One way to avoid frame-based human annotation is the use of action order information to learn the respective action classes. In this context, we propose a hierarchical approach to address the problem of weakly supervised learning of human actions from ordered action labels by structuring recognition in a coarse-to-fine manner. Given a set of videos and an ordered list of the occurring actions, the task is to infer start and end frames of the related action classes within the video and to train the respective action classifiers without any need for hand labeled frame boundaries. We address this problem by combining a framewise RNN model with a coarse probabilistic inference. This combination allows for the temporal alignment of long sequences and thus, for an iterative training of both elements. While this system alone already generates good results, we show that the performance can be further improved by approximating the number of subactions to the characteristics of the different action classes as well as by the introduction of a regularizing length prior. The proposed system is evaluated on two benchmark datasets, the Breakfast and the Hollywood extended dataset, showing a competitive performance on various weak learning tasks such as temporal action segmentation and action alignment.

Analyzing human actions in videos has gained increased attention recently. While most works focus on classifying and labeling observed video frames or anticipating the very recent future, making long-term predictions over more than just a few seconds is a task with many practical applications that has not yet been addressed. In this paper, we propose two methods to predict a considerably large amount of future actions and their durations. Both, a CNN and an RNN are trained to learn future video labels based on previously seen content. We show that our methods generate accurate predictions of the future even for long videos with a huge amount of different actions and can even deal with noisy or erroneous input information.

Action detection and temporal segmentation of actions in videos are topics of increasing interest. While fully supervised systems have gained much attention lately, full annotation of each action within the video is costly and impractical for large amounts of video data. Thus, weakly supervised action detection and temporal segmentation methods are of great importance. While most works in this area assume an ordered sequence of occurring actions to be given, our approach only uses a set of actions. Such action sets provide much less supervision since neither action ordering nor the number of action occurrences are known. In exchange, they can be easily obtained, for instance, from meta-tags, while ordered sequences still require human annotation. We introduce a system that automatically learns to temporally segment and label actions in a video, where the only supervision that is used are action sets. An evaluation on three datasets shows that our method still achieves good results although the amount of supervision is significantly smaller than for other related methods.

Video learning is an important task in computer vision and has experienced increasing interest over the recent years. Since even a small amount of videos easily comprises several million frames, methods that do not rely on a frame-level annotation are of special importance. In this work, we propose a novel learning algorithm with a Viterbi-based loss that allows for online and incremental learning of weakly annotated video data. We moreover show that explicit context and length modeling leads to huge improvements in video segmentation and labeling tasks and include these models into our framework. On several action segmentation benchmarks, we obtain an improvement of up to 10% compared to current state-of-the-art methods.

We present an approach for weakly supervised learning of human actions. Given a set of videos and an ordered list of the occurring actions, the goal is to infer start and end frames of the related action classes within the video and to train the respective action classifiers without any need for hand labeled frame boundaries. To address this task, we propose a combination of a discriminative representation of subactions, modeled by a recurrent neural network, and a coarse probabilistic model to allow for a temporal alignment and inference over long sequences. While this system alone already generates good results, we show that the performance can be further improved by approximating the number of subactions to the characteristics of the different action classes. To this end, we adapt the number of subaction classes by iterating realignment and reestimation during training. The proposed system is evaluated on two benchmark datasets, the Breakfast and the Hollywood extended dataset, showing a competitive performance on various weak learning tasks such as temporal action segmentation and action alignment.

While current approaches to action recognition on presegmented video clips already achieve high accuracies, temporal action detection is still far from comparably good results. Automatically locating and classifying the relevant action segments in videos of varying lengths proves to be a challenging task. We propose a novel method for temporal action detection including statistical length and language modeling to represent temporal and contextual structure. Our approach aims at globally optimizing the joint probability of three components, a length and language model and a discriminative action model, without making intermediate decisions. The problem of finding the most likely action sequence and the corresponding segment boundaries in an exponentially large search space is addressed by dynamic programming. We provide an extensive evaluation of each model component on Thumos 14, a large action detection dataset, and report state-of-the-art results on three datasets.

Bag-of-words (BoW) models are widely used in the field of computer vision. A BoW model consists of a visual vocabulary that is generated by unsupervised clustering the features of the training data, e.g., by using kMeans. The clustering methods, however, struggle with large amounts of data, in particular, in the context of action recognition. In this paper, we propose a transformation of the standard BoW model into a neural network, enabling discriminative training of the visual vocabulary on large action recog-nition datasets. We show that our model is equivalent to the original BoW model but allows for the application of supervised neural network training. Our model outperforms the conventional BoW model and sparse coding methods on recent action recognition benchmarks. 1

This paper describes the new release of RASR — the open source version of the well-proven speech recognition toolkit developed and used at RWTH Aachen University. The focus is put on the implementation of the NN module for training neural network acoustic models. We describe code design, configuration, and features of the NN module. The key feature is a high flexibility regarding the network topology, choice of activation functions, training criteria, and optimization algorithm, as well as a built-in support for efficient GPU computing. The evaluation of run-time performance and recognition accuracy is performed exemplary with a deep neural network as acoustic model in a hybrid NN/HMM system. The results show that RASR achieves a state-of-the-art performance on a real-world large vocabulary task, while offering a complete pipeline for building and applying large scale speech recognition systems.

Deep neural networks are typically optimized with stochastic gradient descent (SGD). In this work, we propose a novel second-order stochastic optimization algorithm. The algorithm is based on analytic results showing that a non-zero mean of features is harmful for the optimization. We prove convergence of our algorithm in a convex setting. In our experiments we show that our proposed algorithm converges faster than SGD. Further, in contrast to earlier work, our algorithm allows for training models with a factorized structure from scratch. We found this structure to be very useful not only because it accelerates training and decoding, but also because it is a very effective means against overfitting. Combining our proposed optimization algorithm with this model structure, model size can be reduced by a factor of eight and still improvements in recognition error rate are obtained. Additional gains are obtained by improving the Newbob learning rate strategy.

Log-linear models find a wide range of applications in pattern recognition. The training of log-linear models is a convex optimization problem. In this work, we compare the performance of stochastic and batch optimization algorithms. Stochastic algorithms are fast on large data sets but can not be parallelized well. In our experiments on a broadcast conversations recognition task, stochastic methods yield competitive results after only a short training period, but when spending enough computational resources for parallelization, batch algorithms are competitive with stochastic algorithms. We obtained slight improvements by using a stochastic second order algorithm. Our best log-linear model outperforms the maximum likelihood trained Gaussian mixture model baseline although being ten times smaller.

Log-linear acoustic models have been shown to be competitive with Gaussian mixture models in speech recognition. Their high training time can be reduced by feature selection. We compare a simple univariate feature selection algorithm with ReliefF - an efficient multivariate algorithm. An alternative to feature selection is ℓ1-regularized training, which leads to sparse models. We observe that this gives no speedup when sparse features are used, hence feature selection methods are preferable. For dense features, ℓ1-regularization can reduce training and recognition time. We generalize the well known Rprop algorithm for the optimization of ℓ1-regularized functions. Experiments on the Wall Street Journal corpus showed that a large number of sparse features could be discarded without loss of performance. A strong regularization led to slight performance degradations, but can be useful on large tasks, where training the full model is not tractable.