Anirudh Sundar, Sinead Williamson, Katherine Metcalf, Barry-John Theobald, Skyler Seto, Masha Fedzechkina. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Iterative self-training, or iterative pseudo-labeling (IPL)—using an improved model from the current iteration to provide pseudo-labels for the next iteration—has proven to be a powerful approach to enhance the quality of speaker representations. Recent applications of IPL in unsupervised speaker recognition start with representations extracted from very elaborate self-supervised methods (e.g., DINO). However, training such strong self-supervised models is not straightforward (they require hyper-parameter tuning and may not generalize to out-of-domain data) and, moreover, may not be needed at all. To this end, we show that the simple, well-studied, and established i-vector generative model is enough to bootstrap the IPL process for the unsupervised learning of speaker representations. We also systematically study the impact of other components on the IPL process, which includes the initial model, the encoder, augmentations, the number of clusters, and the clustering algorithm. Remarkably, we find that even with a simple and significantly weaker initial model like i-vector, IPL can still achieve speaker verification performance that rivals state-of-the-art methods.

This study explores using embedding rank as an unsupervised evaluation metric for general-purpose speech encoders trained via self-supervised learning (SSL). Traditionally, assessing the performance of these encoders is resource-intensive and requires labeled data from the downstream tasks. Inspired by the vision domain, where embedding rank has shown promise for evaluating image encoders without tuning on labeled downstream data, this work examines its applicability in the speech domain, considering the temporal nature of the signals. The findings indicate rank correlates with downstream performance within encoder layers across various downstream tasks and for in- and out-of-domain scenarios. However, rank does not reliably predict the best-performing layer for specific downstream tasks, as lower-ranked layers can outperform higher-ranked ones. Despite this limitation, the results suggest that embedding rank can be a valuable tool for monitoring training progress in SSL speech models, offering a less resource-demanding alternative to traditional evaluation methods.

Speech foundation models, such as HuBERT and its variants, are pre-trained on large amounts of unlabeled speech data and then used for a range of downstream tasks. These models use a masked prediction objective, where the model learns to predict information about masked input segments from the unmasked context. The choice of prediction targets in this framework impacts their performance on downstream tasks. For instance, models pre-trained with targets that capture prosody learn representations suited for speaker-related tasks, while those pre-trained with targets that capture phonetics learn representations suited for content-related tasks. Moreover, prediction targets can differ in the level of detail they capture. Models pre-trained with targets that encode fine-grained acoustic features perform better on tasks like denoising, while those pre-trained with targets focused on higher-level abstractions are more effective for content-related tasks. Despite the importance of prediction targets, the design choices that affect them have not been thoroughly studied. This work explores the design choices and their impact on downstream task performance. Our results indicate that the commonly used design choices for HuBERT can be suboptimal. We propose approaches to create more informative prediction targets and demonstrate their effectiveness through improvements across various downstream tasks.

Vision foundation models pre-trained on massive data encode rich representations of real-world concepts, which can be adapted to downstream tasks by fine-tuning. However, fine-tuning foundation models on one task often leads to the issue of *concept forgetting* on other tasks. Recent methods of robust fine-tuning aim to mitigate forgetting of prior knowledge without affecting the fine-tuning performance. Knowledge is often preserved by matching the original and fine-tuned model weights or feature pairs. However, such point-wise matching can be too strong, without explicit awareness of the feature neighborhood structures that encode rich knowledge as well. We propose a novel regularization method **Proxy-FDA** that explicitly preserves the structural knowledge in feature space. Proxy-FDA performs Feature Distribution Alignment (using nearest neighbor graphs) between the pre-trained and fine-tuned feature spaces, and the alignment is further improved by informative proxies that are generated dynamically to increase data diversity. Experiments show that Proxy-FDA significantly reduces concept forgetting during fine-tuning, and we find a strong correlation between forgetting and a distributional distance metric (in comparison to L2 distance). We further demonstrate Proxy-FDA's benefits in various fine-tuning settings (end-to-end, few-shot and continual tuning) and across different tasks like image classification, captioning and VQA.

Accommodating human preferences is essential for creating aligned LLM agents that deliver personalized and effective interactions. Recent work has shown the potential for LLMs acting as writing agents to infer a description of user preferences. Agent alignment then comes from conditioning on the inferred preference description. However, existing methods often produce generic preference descriptions that fail to capture the unique and individualized nature of human preferences. This paper introduces PROSE, a method designed to enhance the precision of preference descriptions inferred from user writing samples. PROSE incorporates two key elements: (1) iterative refinement of inferred preferences, and (2) verification of inferred preferences across multiple user writing samples. We evaluate PROSE with several LLMs (i.e., Qwen2.5 7B and 72B Instruct, GPT-mini, and GPT-4o) on a summarization and an email writing task. We find that PROSE more accurately infers nuanced human preferences, improving the quality of the writing agent’s generations over CIPHER (a state-of-the-art method for inferring preferences) by 33%. Lastly, we demonstrate that ICL and PROSE are complementary methods, and combining them provides up to a 9% improvement over ICL alone. Code: https://github.com/apple/ml-predict

The recent rapid adoption of large language models (LLMs) highlights the critical need for benchmarking their fairness. Conventional fairness metrics, which focus on discrete accuracy-based evaluations (i.e., prediction correctness), fail to capture the implicit impact of model uncertainty (e.g., higher model confidence about one group over another despite similar accuracy). To address this limitation, we propose an uncertainty-aware fairness metric, UCerf, to enable a fine-grained evaluation of model fairness that is more reflective of the internal bias in model decisions. Furthermore, observing data size, diversity, and clarity issues in current datasets, we introduce a new gender-occupation fairness evaluation dataset with 31,756 samples for co-reference resolution, offering a more diverse and suitable benchmark for modern LLMs. Combining our metric and dataset, we provide insightful comparisons of eight open-source LLMs. For example, Mistral-8B exhibits suboptimal fairness due to high confidence in incorrect predictions, a detail overlooked by Equalized Odds but captured by UCerF. Overall, this work provides a holistic framework for LLM evaluation by jointly assessing fairness and uncertainty, enabling the development of more transparent and accountable AI systems.

Preference-based Reinforcement Learning (PbRL) enables non-experts to train Reinforcement Learning models using preference feedback. However, the effort required to collect preference labels from real humans means that PbRL research primarily relies on synthetic labellers. We validate the most common synthetic labelling strategy by comparing against labels collected from a crowd of humans on three Deep Mind Control (DMC) suite tasks: stand, walk, and run. We find that: (1) the synthetic labels are a good proxy for real humans under some circumstances, (2) strong preference label agreement between human and synthetic labels is not necessary for similar policy performance, (3) policy performance is higher at the start of training from human feedback and is higher at the end of training from synthetic feedback, and (4) training on only examples with high levels of inter-annotator agreement does not meaningfully improve policy performance. Our results justify the use of synthetic labellers to develop and ablate PbRL methods, and provide insight into how human labelling changes over the course of policy training.

Humans can picture a sound scene given an imprecise natural language description. For example, it is easy to imagine an acoustic environment given a phrase like "the lion roar came from right behind me!". For a machine to have the same degree of comprehension, the machine must know what a lion is (semantic attribute), what the concept of "behind" is (spatial attribute) and how these pieces of linguistic information align with the semantic and spatial attributes of the sound (what a roar sounds like when its coming from behind). State-of-the-art audio foundation models, such as CLAP, which learn to map between audio scenes and natural textual descriptions, are trained on non-spatial audio and text pairs, and hence lack spatial awareness. In contrast, sound event localization and detection models are limited to recognizing sounds from a fixed number of classes, and they localize the source to absolute position (e.g., 0.2m) rather than a position described using natural language (e.g., "next to me"). To address these gaps, we present ELSA (Embeddings for Language and Spatial Audio), a spatially aware-audio and text embedding model trained using multimodal contrastive learning. ELSA supports non-spatial audio, spatial audio, and open vocabulary text captions describing both the spatial and semantic components of sound. To train ELSA: (a) we spatially augment the audio and captions of three open-source audio datasets totaling 4,738 hours and 890,038 samples of audio comprised from 8,972 simulated spatial configurations, and (b) we design an encoder to capture the semantics of non-spatial audio, and the semantics and spatial attributes of spatial audio using contrastive learning. ELSA is a single model that is competitive with state-of-the-art for both semantic retrieval and 3D source localization. In particular, ELSA achieves +2.8\% mean audio-to-text and text-to-audio R@1 above the LAION-CLAP baseline, and outperforms by -11.6° mean-absolute-error in 3D source localization over the SeldNET baseline on the TUT Sound Events 2018 benchmark. Moreover, we show that the representation-space of ELSA is structured, enabling swapping of direction of audio via vector arithmetic of two directional text embeddings.

Fully-test-time adaptation (F-TTA) can mitigate performance loss due to distribution shifts between train and test data (1) without access to the training data, and (2) without knowledge of the model training procedure. In online F-TTA, a pre-trained model is adapted using a stream of test samples by minimizing a self-supervised objective, such as entropy minimization. However, models adapted with online using entropy minimization, are unstable especially in single sample settings, leading to degenerate solutions, and limiting the adoption of TTA inference strategies. Prior works identify noisy, or unreliable, samples as a cause of failure in online F-TTA. One solution is to ignore these samples, which can lead to bias in the update procedure, slow adaptation, and poor generalization. In this work, we present a general framework for improving robustness of F-TTA to these noisy samples, inspired by self-paced learning and robust loss functions. Our proposed approach, Robust Entropy Adaptive Loss Minimization (REALM), achieves better adaptation accuracy than previous approaches throughout the adaptation process on corruptions of CIFAR-10 and ImageNet-1K, demonstrating its effectiveness.

Generating realistic lip motion from audio to simulate speech production is critical for driving natural character animation. Previous research has shown that traditional metrics used to optimize and assess models for generating lip motion from speech are not a good indicator of subjective opinion of animation quality. Devising metrics that align with subjective opinion first requires understanding what impacts human perception of quality. In this work, we focus on the degree of articulation and run a series of experiments to study how articulation strength impacts human perception of lip motion accompanying speech. Specifically, we study how increasing under-articulated (dampened) and over-articulated (exaggerated) lip motion affects human perception of quality. We examine the impact of articulation strength on human perception when considering only lip motion, where viewers are presented with talking faces represented by landmarks, and in the context of embodied characters, where viewers are presented with photo-realistic videos. Our results show that viewers prefer over-articulated lip motion consistently more than under-articulated lip motion and that this preference generalizes across different speakers and embodiments.

Synthesizing natural head motion to accompany speech for an embodied conversational agent is necessary for pro-viding a rich interactive experience. Most prior works assess the quality of generated head motion by comparing them against a single ground-truth using an objective metric. Yet there are many plausible head motion sequences to accompany a speech utterance. In this work, we study the variation in the perceptual quality of head motions sampled from a generative model. We show that, despite providing more di-verse head motions, the generative model produces motions with varying degrees of perceptual quality. We finally show that objective metrics commonly used in previous research do not accurately reflect the perceptual quality of generated head motions. These results open an interesting avenue for future work to investigate better objective metrics that correlate with human perception of quality.

We present Spatial LibriSpeech, a spatial audio dataset with over 650 hours of 19-channel audio, first-order ambisonics, and optional distractor noise. Spatial LibriSpeech is designed for machine learning model training, and it includes labels for source position, speaking direction, room acoustics and geometry. Spatial LibriSpeech is generated by augmenting LibriSpeech samples with 200k+ simulated acoustic conditions across 8k+ synthetic rooms. To demonstrate the utility of our dataset, we train models on four spatial audio tasks, resulting in a median absolute error of 6.60{\deg} on 3D source localization, 0.43m on distance, 90.66ms on T30, and 2.74dB on DRR estimation. We show that the same models generalize well to widely-used evaluation datasets, e.g., obtaining a median absolute error of 12.43{\deg} on 3D source localization on TUT Sound Events 2018, and 157.32ms on T30 estimation on ACE Challenge.

We present an introspection of an audiovisual speech enhancement model. In particular, we focus on interpreting how a neural audiovisual speech enhancement model uses visual cues to improve the quality of the target speech signal. We show that visual cues provide not only high-level information about speech activity, i.e., speech/silence, but also fine-grained visual information about the place of articulation. One byproduct of this finding is that the learned visual embeddings can be used as features for other visual speech applications. We demonstrate the effectiveness of the learned visual embeddings for classifying visemes (the visual analogy to phonemes). Our results provide insight into important aspects of audiovisual speech enhancement and demonstrate how such models can be used for self-supervision tasks for visual speech applications.

Automatic speech recognition (ASR) is widely used in consumer electronics. ASR greatly improves the utility and accessibility of technology, but usually the output is only word sequences without punctuation. This can result in ambiguity in inferring user-intent. We first present a transformer-based approach for punctuation prediction that achieves 8% improvement on the IWSLT 2012 TED Task, beating the previous state of the art [1]. We next describe our multimodal model that learns from both text and audio, which achieves 8% improvement over the text-only algorithm on an internal dataset for which we have both the audio and transcriptions. Finally, we present an approach to learning a model using contextual dropout that allows us to handle variable amounts of future context at test time.

We describe our novel deep learning approach for driving animated faces using both acoustic and visual information. In particular, speech-related facial movements are generated using audiovisual information, and non-verbal facial movements are generated using only visual information. To ensure that our model exploits both modalities during training, batches are generated that contain audio-only, video-only, and audiovisual input features. The probability of dropping a modality allows control over the degree to which the model exploits audio and visual information during training. Our trained model runs in real-time on resource limited hardware (e.g. a smart phone), it is user agnostic, and it is not dependent on a potentially error-prone transcription of the speech. We use subjective testing to demonstrate: 1) the improvement of audiovisual-driven animation over the equivalent video-only approach, and 2) the improvement in the animation of speech-related facial movements after introducing modality dropout. Without modality dropout, viewers prefer audiovisual-driven animation in 51% of the test sequences compared with only 18% for video-driven. After introducing dropout viewer preference for audiovisual-driven animation increases to 74%, but decreases to 8% for video-only.

Speech-driven visual speech synthesis involves mapping acoustic speech features to the corresponding lip animation controls for a face model. This mapping can take many forms, but a powerful approach is to use deep neural networks (DNNs). The lack of synchronized audio, video, and depth data is a limitation to reliably train DNNs, especially for speaker-independent models. In this paper, we investigate adapting an automatic speech recognition (ASR) acoustic model (AM) for the visual speech synthesis problem. We train the ASR-AM on ten thousand hours of audio-only transcribed speech. The ASR-AM is then adapted to the visual speech synthesis domain using ninety hours of synchronized audio-visual speech. Using a subjective assessment test, we compared the performance of the AM-initialized DNN to a randomly initialized model. The results show that viewers significantly prefer animations generated from the AM-initialized DNN than the ones generated using the randomly initialized model. We conclude that visual speech synthesis can significantly benefit from the powerful representation of speech in the ASR acoustic models.

We describe experiments towards building a conversational digital assistant that considers the preferred conversational style of the user. In particular, these experiments are designed to measure whether users prefer and trust an assistant whose conversational style matches their own. To this end we conducted a user study where subjects interacted with a digital assistant that responded in a way that either matched their conversational style, or did not. Using self-reported personality attributes and subjects' feedback on the interactions, we built models that can reliably predict a user's preferred conversational style.

This paper introduces a method for automatic redubbing of video that exploits the many-to-many mapping of phoneme sequences to lip movements modelled as dynamic visemes (Taylor et al., 2012). For a given utterance, the corresponding dynamic viseme sequence is sampled to construct a graph of possible phoneme sequences that synchronize with the video. When composed with a pronunciation dictionary and language model, this produces a vast number of word sequences that are in sync with the original video, literally putting plausible words into the mouth of the speaker. We demonstrate that traditional, many-to-one, static visemes lack flexibility for this application as they produce significantly fewer word sequences. This work explores the natural ambiguity in visual speech and offers insight for automatic speech recognition and the importance of language modeling.

The speed that an utterance is spoken affects both the duration of the speech and the position of the articulators. Consequently, the sounds that are produced are modified, as are the position and appearance of the lips, teeth, tongue and other visible articulators. We describe an experiment designed to measure the effect of variable speaking rate on audio and visual speech by comparing sequences of phonemes and dynamic visemes appearing in the same sentences spoken at different speeds. We find that both audio and visual speech production are affected by varying the rate of speech, however, the effect is significantly more prominent in visual speech.

Visual-only speech recognition is dependent upon a number of factors that can be difficult to control, such as: lighting; identity; motion; emotion and expression. But some factors, such as video resolution are controllable, so it is surprising that there is not yet a systematic study of the effect of resolution on lip-reading. Here we use a new data set, the Rosetta Raven data, to train and test recognizers so we can measure the affect of video resolution on recognition accuracy. We conclude that, contrary to common practice, resolution need not be that great for automatic lip-reading. However it is highly unlikely that automatic lip-reading can work reliably when the distance between the bottom of the lower lip and the top of the upper lip is less than four pixels at rest.

Computer lip-reading is one of the great signal processing challenges. Not only is the signal noisy, it is variable. However it is almost unknown to compare the performance with human lip-readers. Partly this is because of the paucity of human lip-readers and partly because most automatic systems only handle data that are trivial and therefore not representative of human speech. Here we generate a multiview dataset using connected words that can be analysed by an automatic system, based on linear predictive trackers and active appearance models, and human lip-readers. The automatic system we devise has a viseme accuracy of ≈ 46% which is comparable to poor professional human lip-readers. However, unlike human lip-readers our system is good at guessing its fallibility.

Computer lip-reading systems are usually designed to work using a full-frontal view of the face. However, many human experts tend to prefer to lip-read using an angled view. In this paper we consider issues related to the best viewing angle for an automated lip-reading system. In particular, we seek answers to the following questions: (1) Do computers lip-read better using a frontal or a non-frontal view of the face? (2) What is the best viewing angle for a computer lip-reading system? (3) How can a computer lip-reading system be made to work independently of viewing angle? We investigate these issues using a purpose built audio-visual dataset that contains simultaneous recordings of a speaker reciting continuous speech at five angles. We find that the system performs best on a non-frontal view, perhaps because lip gestures, such as lip-protrusion and lip-rounding, are more pronounced when viewing from an angle. We also describe a simple linear mapping that allows us to map any view of the face to the view that we find to be optimal. Hence we present a view-independent lip-reading system.

This research is part of a wider project to build predictive models of bone age using hand radiograph images. We examine ways of finding the outline of a hand from an X-ray as the first stage in segmenting the image into constituent bones. We assess a variety of algorithms including contouring, which has not previously been used in this context. We introduce a novel ensemble algorithm for combining outlines using two voting schemes, a likelihood ratio test and dynamic time warping (DTW). Our goal is to minimize the human intervention required, hence we investigate alternative ways of training a classifier to determine whether an outline is in fact correct or not. We evaluate outlining and classification on a set of 1370 images. We conclude that ensembling with DTW improves performance of all outlining algorithms, that the contouring algorithm used with the DTW ensemble performs the best of those assessed, and that the most effective classifier of hand outlines assessed is a random forest applied to outlines transformed into principal components.

We compare two approaches for synthesizing visual speech using active appearance models (AAMs): one that utilizes acoustic features as input, and one that utilizes a phonetic transcription as input. Both synthesizers are trained using the same data and the performance is measured using both objective and subjective testing. We investigate the impact of likely sources of error in the synthesized visual speech by introducing typical errors into real visual speech sequences and subjectively measuring the perceived degradation. When only a small region (e.g., a single syllable) of ground-truth visual speech is incorrect we find that the subjective score for the entire sequence is subjectively lower than sequences generated by our synthesizers. This observation motivates further consideration of an often ignored issue, which is to what extent are subjective measures correlated with objective measures of performance? Significantly, we find that the most commonly used objective measures of performance are not necessarily the best indicator of viewer perception of quality. We empirically evaluate alternatives and show that the cost of a dynamic time warp of synthesized visual speech parameters to the respective ground-truth parameters is a better indicator of subjective quality.

This paper proposes a learnt data-driven approach for accurate, real-time tracking of facial features using only intensity information. Constraints such as a-priori shape models or temporal models for dynamics are not required or used. Tracking facial features simply becomes the independent tracking of a set of points on the face. This allows us to cope with facial configurations not present in the training data. Tracking is achieved via linear predictors which provide a fast and effective method for mapping pixel-level information to tracked feature position displacements. To improve on this, a novel and robust biased linear predictor is proposed in this paper. Multiple linear predictors are grouped into a rigid flock to increase robustness. To further improve tracking accuracy, a novel probabilistic selection method is used to identify relevant visual areas for tracking a feature point. These selected flocks are then combined into a hierarchical multi-resolution LP model. Experimental results also show that this method performs more robustly and accurately than AAMs, without any a priori shape information and with minimal training examples.

In this paper we present an overview of LIPS2008: Visual Speech Synthesis Challenge.The aim of this challenge is to bring together researchers in the field of visual speech synthesis to firstly evaluate their systems within a common framework, and secondly to identify the needs of the wider community in terms of evaluation.In doing so we hope to better understand the differences between the various approaches and to identify the strengths/weaknesses of the competing approaches.In this paper we firstly motivate the need for the challenge, before describing the capture and preparation of the training data, the evaluation framework, and conclude with an outline of possible directions for standardising the evaluation of talking heads.

Pain is typically assessed by patient self-report. Self-reported pain, however, is difficult to interpret and may be impaired or not even possible, as in young children or the severely ill. Behavioral scientists have identified reliable and valid facial indicators of pain. Until now they required manual measurement by highly skilled observers. We developed an approach that automatically recognizes acute pain. Adult patients with rotator cuff injury were video-recorded while a physiotherapist manipulated their affected and unaffected shoulder. Skilled observers rated pain expression from the video on a 5-point Likert-type scale. From these ratings, sequences were categorized as no-pain (rating of 0), pain (rating of 3, 4, or 5), and indeterminate (rating of 1 or 2). We explored machine learning approaches for pain-no pain classification. Active Appearance Models (AAM) were used to decouple shape and appearance parameters from the digitized face images. Support vector machines (SVM) were used with several representations from the AAM. Using a leave-one-out procedure, we achieved an equal error rate of 19% (hit rate = 81%) using canonical appearance and shape features. These findings suggest the feasibility of automatic pain detection from video.

Active Appearance Models (AAMs) are generative parametric models commonly used to track, recognise and synthesise faces in images and video sequences. In this paper we describe a method for transferring dynamic facial gestures between subjects in real-time. The main advantages of our approach are that: 1) the mapping is computed automatically and does not require high-level semantic information describing facial expressions or visual speech gestures. 2) The mapping is simple and intuitive, allowing expressions to be transferred and rendered in real-time. 3) The mapped expression can be constrained to have the appearance of the target producing the expression, rather than the source expression imposed onto the target face. 4) Near-videorealistic talking faces for new subjects can be created without the cost of recording and processing a complete training corpus for each. Our system enables face-to-face interaction with an avatar driven by an AAM of an actual person in real-time and we show examples of arbitrary expressive speech frames cloned across different subjects.

Two dimensional (2D) shape and appearance models are applied to the problem of creating a near-videorealistic talking head. A speech corpus of a talker uttering a set of phonetically balanced training sentences is analysed using a generative model of the human face. Segments of original parameter trajectories, corresponding to the synthesis unit (e.g.~triphone), are extracted from a codebook, then normalised, blended, concatenated and smoothed before being applied to the model to give natural, realistic animations of novel utterances. The system provides a 2D image sequence corresponding to the face of a talker. It is also used to animate the face of a 3D avatar by displacing the mesh according to movements of points in the shape model and dynamically texturing the face polygons using the appearance model.

We present work towards videorealistic synthetic visual speech using non-rigid appearance models. These models are used to track a talking face enunciating a set of training sentences. The resultant parameter trajectories are used in a concatenative synthesis scheme, where samples of original data are extracted from a corpus and concatenated to form new unseen sequences. Here we explore the effect on the synthesiser output of blending several synthesis units considered similar to the desired unit. We present preliminary subjective and objective results used to judge the realism of the system.

In this paper we present initial work towards a video-realistic visual speech synthesiser based on statistical models of shape and appearance. A synthesised image sequence corresponding to an utterance is formed by concatenation of synthesis units (in this case phonemes) from a pre-recorded corpus of training data. A smoothing spline is applied to the concatenated parameters to ensure smooth transitions between frames and the resultant parameters applied to the model—early results look promising.