The emergence of large language models (LLMs) has transformed spoken dialog systems, yet the optimal architecture for real-time on-device voice agents remains an open question. While end-to-end approaches promise theoretical advantages, cascaded systems (CSs) continue to outperform them in language understanding tasks, despite being constrained by sequential processing latency. In this work, we introduce ChipChat, a novel low-latency CS that overcomes traditional bottlenecks through architectural innovations and streaming optimizations. Our system integrates streaming (a) conversational speech recognition with mixture-of-experts, (b) state-action augmented LLM, (c) text-to-speech synthesis, (d) neural vocoder, and (e) speaker modeling. Implemented using MLX, ChipChat achieves subsecond response latency on a Mac Studio without dedicated GPUs, while preserving user privacy through complete on-device processing. Our work shows that strategically redesigned CSs can overcome their historical latency limitations, offering a promising path forward for practical voice-based AI agents.

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.

The success of large language models in text processing has inspired their adaptation to speech modeling. However, since speech is continuous and complex, it is often discretized for autoregressive modeling. Speech tokens derived from self-supervised models (known as semantic tokens) typically focus on the linguistic aspects of speech but neglect prosodic information. As a result, models trained on these tokens can generate speech with reduced naturalness. Existing approaches try to fix this by adding pitch features to the semantic tokens. However, pitch alone cannot fully represent the range of paralinguistic attributes, and selecting the right features requires careful hand-engineering. To overcome this, we propose an end-to-end variational approach that automatically learns to encode these continuous speech attributes to enhance the semantic tokens. Our approach eliminates the need for manual extraction and selection of paralinguistic features. Moreover, it produces preferred speech continuations according to human raters. Code, samples and models are available at https://github.com/b04901014/vae-gslm.

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.

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.

The robustness of an acoustic emotion recognition system hinges on first having access to features that represent an acoustic input signal. These representations should abstract extraneous low-level variations present in acoustic signals and only capture speaker characteristics relevant for emotion recognition. Previous research has demonstrated that, in other classification tasks, when large labeled datasets are available, neural networks trained on these data learn to extract robust features from the input signal. However, the datasets used for developing emotion recognition systems remain significantly smaller than those used for developing other speech systems. Thus, acoustic emotion recognition systems remain in need of robust feature representations. In this article, we study the utility of speaker embeddings, representations extracted from a trained speaker recognition network, as robust features for detecting emotions. We first study the relationship between emotions and speaker embeddings and demonstrate how speaker embeddings highlight the differences that exist between neutral speech and emotionally expressive speech. We quantify the modulations that variations in emotional expression incur on speaker embeddings and show how these modulations are greater than those incurred from lexical variations in an utterance. Finally, we demonstrate how speaker embeddings can be used as a replacement for traditional low-level acoustic features for emotion recognition.

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.

Zakaria Aldeneh, Matthew Perez, Emily Mower Provost. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Robust speech recognition is a key prerequisite for semantic feature extraction in automatic aphasic speech analysis. However, standard one-size-fits-all automatic speech recognition models perform poorly when applied to aphasic speech. One reason for this is the wide range of speech intelligibility due to different levels of severity (i.e., higher severity lends itself to less intelligible speech). To address this, we propose a novel acoustic model based on a mixture of experts (MoE), which handles the varying intelligibility stages present in aphasic speech by explicitly defining severity-based experts. At test time, the contribution of each expert is decided by estimating speech intelligibility with a speech intelligibility detector (SID). We show that our proposed approach significantly reduces phone error rates across all severity stages in aphasic speech compared to a baseline approach that does not incorporate severity information into the modeling process.

Emotion recognition algorithms rely on data annotated with high quality labels. However, emotion expression and perception are inherently subjective. There is generally not a single annotation that can be unambiguously declared "correct." As a result, annotations are colored by the manner in which they were collected. In this paper, we conduct crowdsourcing experiments to investigate this impact on both the annotations themselves and on the performance of these algorithms. We focus on one critical question: the effect of context. We present a new emotion dataset, Multimodal Stressed Emotion (MuSE), and annotate the dataset using two conditions: randomized, in which annotators are presented with clips in random order, and contextualized, in which annotators are presented with clips in order. We find that contextual labeling schemes result in annotations that are more similar to a speaker's own self-reported labels and that labels generated from randomized schemes are most easily predictable by automated systems.

Various psychological factors affect how individuals express emotions. Yet, when we collect data intended for use in building emotion recognition systems, we often try to do so by creating paradigms that are designed just with a focus on eliciting emotional behavior. Algorithms trained with these types of data are unlikely to function outside of controlled environments because our emotions naturally change as a function of these other factors. In this work, we study how the multimodal expressions of emotion change when an individual is under varying levels of stress. We hypothesize that stress produces modulations that can hide the true underlying emotions of individuals and that we can make emotion recognition algorithms more generalizable by controlling for variations in stress. To this end, we use adversarial networks to decorrelate stress modulations from emotion representations. We study how stress alters acoustic and lexical emotional predictions, paying special attention to how modulations due to stress affect the transferability of learned emotion recognition models across domains. Our results show that stress is indeed encoded in trained emotion classifiers and that this encoding varies across levels of emotions and across the lexical and acoustic modalities. Our results also show that emotion recognition models that control for stress during training have better generalizability when applied to new domains, compared to models that do not control for stress during training. We conclude that is is necessary to consider the effect of extraneous psychological factors when building and testing emotion recognition models.

Bipolar disorder, a severe chronic mental illness characterized by pathological mood swings from depression to mania, requires ongoing symptom severity tracking to both guide and measure treatments that are critical for maintaining long-term health. Mental health professionals assess symptom severity through semi-structured clinical interviews. During these interviews, they observe their patients' spoken behaviors, including both what the patients say and how they say it. In this work, we move beyond acoustic and lexical information, investigating how higher-level interactive patterns also change during mood episodes. We then perform a secondary analysis, asking if these interactive patterns, measured through dialogue features, can be used in conjunction with acoustic features to automatically recognize mood episodes. Our results show that it is beneficial to consider dialogue features when analyzing and building automated systems for predicting and monitoring mood.

This work focuses on the use of acoustic cues for modeling turn-taking in dyadic spoken dialogues. Previous work has shown that speaker intentions (e.g., asking a question, uttering a backchannel, etc.) can influence turn-taking behavior and are good predictors of turn-transitions in spoken dialogues. However, speaker intentions are not readily available for use by automated systems at run-time; making it difficult to use this information to anticipate a turn-transition. To this end, we propose a multi-task neural approach for predicting turn-transitions and speaker intentions simultaneously. Our results show that adding the auxiliary task of speaker intention prediction improves the performance of turn-transition prediction in spoken dialogues, without relying on additional input features during run-time.

In this paper, we show that convolutional neural networks can be directly applied to temporal low-level acoustic features to identify emotionally salient regions without the need for defining or applying utterance-level statistics. We show how a convolutional neural network can be applied to minimally hand-engineered features to obtain competitive results on the IEMOCAP and MSP-IMPROV datasets. In addition, we demonstrate that, despite their common use across most categories of acoustic features, utterance-level statistics may obfuscate emotional information. Our results suggest that convolutional neural networks with Mel Filterbanks (MFBs) can be used as a replacement for classifiers that rely on features obtained from applying utterance-level statistics.

In this paper, we present an analysis of different multimodal fusion approaches in the context of deep learning, focusing on pooling intermediate representations learned for the acoustic and lexical modalities. Traditional approaches to multimodal feature pooling include: concatenation, element-wise addition, and element-wise multiplication. We compare these traditional methods to outer-product and compact bilinear pooling approaches, which consider more comprehensive interactions between features from the two modalities. We also study the influence of each modality on the overall performance of a multimodal system. Our experiments on the IEMOCAP dataset suggest that: (1) multimodal methods that combine acoustic and lexical features outperform their unimodal counterparts; (2) the lexical modality is better for predicting valence than the acoustic modality; (3) outer-product-based pooling strategies outperform other pooling strategies.

Many paralinguistic tasks are closely related and thus representations learned in one domain can be leveraged for another. In this paper, we investigate how knowledge can be transferred between three paralinguistic tasks: speaker, emotion, and gender recognition. Further, we extend this problem to cross-dataset tasks, asking how knowledge captured in one emotion dataset can be transferred to another. We focus on progressive neural networks and compare these networks to the conventional deep learning method of pre-training and fine-tuning. Progressive neural networks provide a way to transfer knowledge and avoid the forgetting effect present when pre-training neural networks on different tasks. Our experiments demonstrate that: (1) emotion recognition can benefit from using representations originally learned for different paralinguistic tasks and (2) transfer learning can effectively leverage additional datasets to improve the performance of emotion recognition systems.

The goal of continuous emotion recognition is to assign an emotion value to every frame in a sequence of acoustic features. We show that incorporating long-term temporal dependencies is critical for continuous emotion recognition tasks. To this end, we first investigate architectures that use dilated convolutions. We show that even though such architectures outperform previously reported systems, the output signals produced from such architectures undergo erratic changes between consecutive time steps. This is inconsistent with the slow moving ground-truth emotion labels that are obtained from human annotators. To deal with this problem, we model a downsampled version of the input signal and then generate the output signal through upsampling. Not only does the resulting downsampling/upsampling network achieve good performance, it also generates smooth output trajectories. Our method yields the best known audio-only performance on the RECOLA dataset.

Automatic emotion recognition from audio-visual data is a topic that has been broadly explored using data captured in the laboratory. However, these data are not necessarily representative of how emotion is manifested in the real-world. In this paper, we describe our system for the 2016 Emotion Recognition in the Wild challenge. We use the Acted Facial Expressions in the Wild database 6.0 (AFEW 6.0), which contains short clips of popular TV shows and movies and has more variability in the data compared to laboratory recordings. We explore a set of features that incorporate information from facial expressions and speech, in addition to cues from the background music and overall scene. In particular, we propose the use of a feature set composed of dimensional emotion estimates trained from outside acoustic corpora. We design sets of multiclass and pairwise (one-versus-one) classifiers and fuse the resulting systems. Our fusion increases the performance from a baseline of 38.81% to 43.86% and from 40.47% to 46.88%, for validation and test sets, respectively. While the video features perform better than audio features alone, a combination of the two modalities achieves the greatest performance, with gains of 4.4% and 1.4%, with and without information gain, respectively. Because of the flexible design of the fusion, it is easily adaptable to other multimodal learning problems.