Gabriel Skantze

Jessica Wenninger, Gabriel Skantze · deepmind

To enable meaningful human-robot interaction (HRI), a robot must continuously assess engagement by consistently tracking users over time. State-of-the-art computer vision models, however, are heavily optimized for surveillance or autonomous driving. A social robot faces distinct egocentric challenges, such as humans bouncing, obstructing each other, or leaving the frame. Frequent identity switches (IDSW) cause the robot to lose its footing mid-conversation. To address this, we introduce a novel, custom-annotated egocentric dataset collected via the Furhat robot to capture complex social dynamics. We present a systematic evaluation isolating detection errors from tracking logic, comparing face versus body tracking, and assessing the impact of extended spatial memory and appearance re-identification (ReID). Results indicate that increasing spatial memory mitigates prolonged occlusions but fails on complex dynamic events. Integrating ReID resolves complex switches but exhibits opposing effects: it substantially improves body tracking stability, yet causes facial IDSW to spike due to profile angle sensitivity. Ultimately, our optimized pipeline reduces IDSW by 49\%, mitigating interaction breakdowns. Because standard benchmarks lack dense, close-quarter occlusions, this work highlights the critical need for natively captured social dynamics to truly validate HRI perception models.

A. Seza Doğruöz, Gabriel Skantze

Open-domain chatbots are supposed to converse freely with humans without being restricted to a topic, task or domain. However, the boundaries and/or contents of open-domain conversations are not clear. To clarify the boundaries of "openness", we conduct two studies: First, we classify the types of "speech events" encountered in a chatbot evaluation data set (i.e., Meena by Google) and find that these conversations mainly cover the "small talk" category and exclude the other speech event categories encountered in real life human-human communication. Second, we conduct a small-scale pilot study to generate online conversations covering a wider range of speech event categories between two humans vs. a human and a state-of-the-art chatbot (i.e., Blender by Facebook). A human evaluation of these generated conversations indicates a preference for human-human conversations, since the human-chatbot conversations lack coherence in most speech event categories. Based on these results, we suggest (a) using the term "small talk" instead of "open-domain" for the current chatbots which are not that "open" in terms of conversational abilities yet, and (b) revising the evaluation methods to test the chatbot conversations against other speech events.

Bram Willemsen, Livia Qian, Gabriel Skantze

Vision-language models (VLMs) have shown to be effective at image retrieval based on simple text queries, but text-image retrieval based on conversational input remains a challenge. Consequently, if we want to use VLMs for reference resolution in visually-grounded dialogue, the discourse processing capabilities of these models need to be augmented. To address this issue, we propose fine-tuning a causal large language model (LLM) to generate definite descriptions that summarize coreferential information found in the linguistic context of references. We then use a pretrained VLM to identify referents based on the generated descriptions, zero-shot. We evaluate our approach on a manually annotated dataset of visually-grounded dialogues and achieve results that, on average, exceed the performance of the baselines we compare against. Furthermore, we find that using referent descriptions based on larger context windows has the potential to yield higher returns.

Agnes Axelsson, Gabriel Skantze

In any system that uses structured knowledge graph (KG) data as its underlying knowledge representation, KG-to-text generation is a useful tool for turning parts of the graph data into text that can be understood by humans. Recent work has shown that models that make use of pretraining on large amounts of text data can perform well on the KG-to-text task even with relatively small sets of training data on the specific graph-to-text task. In this paper, we build on this concept by using large language models to perform zero-shot generation based on nothing but the model's understanding of the triple structure from what it can read. We show that ChatGPT achieves near state-of-the-art performance on some measures of the WebNLG 2020 challenge, but falls behind on others. Additionally, we compare factual, counter-factual and fictional statements, and show that there is a significant connection between what the LLM already knows about the data it is parsing and the quality of the output text.

Livia Qian, Gabriel Skantze

Backchannels (e.g., `yeah', `mhm', and `right') are short, non-interruptive feedback signals whose lexical form and prosody jointly convey pragmatic meaning. While prior computational research has largely focused on predicting backchannel timing, the relationship between lexico-prosodic form and meaning remains underexplored. We propose a two-stage framework: first, fine-tuning large language models on dialogue transcripts to derive rich contextual representations; and second, learning a joint embedding space for dialogue contexts and backchannel realizations. We evaluate alignment with human perception via triadic similarity judgments (prosodic and cross-lexical) and a context-backchannel suitability task. Our results demonstrate that the learned projections substantially improve context-backchannel retrieval compared to previous methods. In addition, they reveal that backchannel form is highly sensitive to extended conversational context and that the learned embeddings align more closely with human judgments than raw WavLM features.

Nikita Torgashov, Gustav Eje Henter, Gabriel Skantze

Full-stream text-to-speech (TTS) for interactive systems must start speaking with minimal delay while remaining controllable as text arrives incrementally. We present VoXtream2, a zero-shot full-stream TTS model with dynamic speaking-rate control that can be updated mid-utterance on the fly. VoXtream2 combines a distribution matching mechanism over duration states with classifier-free guidance across conditioning signals to improve controllability and synthesis quality. Prompt-text masking enables textless audio prompting, removing the need for prompt transcription. Across standard zero-shot benchmarks and a dedicated speaking-rate test set, VoXtream2 achieves competitive objective and subjective results against public baselines despite a smaller model and less training data. In full-stream mode, it runs 4 times faster than real time with 74 ms first-packet latency on a consumer GPU.

Gabriel Skantze, A. Seza Doğruöz

There is a surge in interest in the development of open-domain chatbots, driven by the recent advancements of large language models. The "openness" of the dialogue is expected to be maximized by providing minimal information to the users about the common ground they can expect, including the presumed joint activity. However, evidence suggests that the effect is the opposite. Asking users to "just chat about anything" results in a very narrow form of dialogue, which we refer to as the "open-domain paradox". In this position paper, we explain this paradox through the theory of common ground as the basis for human-like communication. Furthermore, we question the assumptions behind open-domain chatbots and identify paths forward for enabling common ground in human-computer dialogue.

Luca Modica, Filip Landin, Mehrdad Farahani et al.

In recent years, several Speech Language Models (SLMs) that represent speech and written text jointly have been presented. The question then emerges about how model-internal mechanisms are similar and different when operating in the two modalities. We focus on how these systems encode, store, and retrieve factual knowledge, which has previously been investigated for text-only models. To investigate mechanisms behind the storage and recall of factual association in SLMs, we leverage Causal Mediation Analysis, a technique previously applied to text-based models. Initial results using SpiritLM, a multimodal model integrating discrete speech tokens reveal discrepancies between text-to-text and speech-to-text results, suggesting that the emergent mechanisms for factual recall are only partially carried over from the text to the speech modality. These results advance our understanding of how internal mechanisms encode factual associations in SLMs while contributing insights for improving speech-enabled AI systems.

Bram Willemsen, Dmytro Kalpakchi, Gabriel Skantze

An idealized, though simplistic, view of the referring expression production and grounding process in (situated) dialogue assumes that a speaker must merely appropriately specify their expression so that the target referent may be successfully identified by the addressee. However, referring in conversation is a collaborative process that cannot be aptly characterized as an exchange of minimally-specified referring expressions. Concerns have been raised regarding assumptions made by prior work on visually-grounded dialogue that reveal an oversimplified view of conversation and the referential process. We address these concerns by introducing a collaborative image ranking task, a grounded agreement game we call "A Game Of Sorts". In our game, players are tasked with reaching agreement on how to rank a set of images given some sorting criterion through a largely unrestricted, role-symmetric dialogue. By putting emphasis on the argumentation in this mixed-initiative interaction, we collect discussions that involve the collaborative referential process. We describe results of a small-scale data collection experiment with the proposed task. All discussed materials, which includes the collected data, the codebase, and a containerized version of the application, are publicly available.

Koji Inoue, Divesh Lala, Keiko Ochi et al.

This paper tackles the challenging task of evaluating socially situated conversational robots and presents a novel objective evaluation approach that relies on multimodal user behaviors. In this study, our main focus is on assessing the human-likeness of the robot as the primary evaluation metric. While previous research often relied on subjective evaluations from users, our approach aims to evaluate the robot's human-likeness based on observable user behaviors indirectly, thus enhancing objectivity and reproducibility. To begin, we created an annotated dataset of human-likeness scores, utilizing user behaviors found in an attentive listening dialogue corpus. We then conducted an analysis to determine the correlation between multimodal user behaviors and human-likeness scores, demonstrating the feasibility of our proposed behavior-based evaluation method.

Bram Willemsen, Gabriel Skantze

We propose an approach to referring expression generation (REG) in visually grounded dialogue that is meant to produce referring expressions (REs) that are both discriminative and discourse-appropriate. Our method constitutes a two-stage process. First, we model REG as a text- and image-conditioned next-token prediction task. REs are autoregressively generated based on their preceding linguistic context and a visual representation of the referent. Second, we propose the use of discourse-aware comprehension guiding as part of a generate-and-rerank strategy through which candidate REs generated with our REG model are reranked based on their discourse-dependent discriminatory power. Results from our human evaluation indicate that our proposed two-stage approach is effective in producing discriminative REs, with higher performance in terms of text-image retrieval accuracy for reranked REs compared to those generated using greedy decoding.

Koji Inoue, Bing'er Jiang, Erik Ekstedt et al.

A demonstration of a real-time and continuous turn-taking prediction system is presented. The system is based on a voice activity projection (VAP) model, which directly maps dialogue stereo audio to future voice activities. The VAP model includes contrastive predictive coding (CPC) and self-attention transformers, followed by a cross-attention transformer. We examine the effect of the input context audio length and demonstrate that the proposed system can operate in real-time with CPU settings, with minimal performance degradation.

Koji Inoue, Bing'er Jiang, Erik Ekstedt et al.

This paper investigates the application of voice activity projection (VAP), a predictive turn-taking model for spoken dialogue, on multilingual data, encompassing English, Mandarin, and Japanese. The VAP model continuously predicts the upcoming voice activities of participants in dyadic dialogue, leveraging a cross-attention Transformer to capture the dynamic interplay between participants. The results show that a monolingual VAP model trained on one language does not make good predictions when applied to other languages. However, a multilingual model, trained on all three languages, demonstrates predictive performance on par with monolingual models across all languages. Further analyses show that the multilingual model has learned to discern the language of the input signal. We also analyze the sensitivity to pitch, a prosodic cue that is thought to be important for turn-taking. Finally, we compare two different audio encoders, contrastive predictive coding (CPC) pre-trained on English, with a recent model based on multilingual wav2vec 2.0 (MMS).

Gabriel Skantze, Bahar Irfan

Turn-taking is a fundamental aspect of conversation, but current Human-Robot Interaction (HRI) systems often rely on simplistic, silence-based models, leading to unnatural pauses and interruptions. This paper investigates, for the first time, the application of general turn-taking models, specifically TurnGPT and Voice Activity Projection (VAP), to improve conversational dynamics in HRI. These models are trained on human-human dialogue data using self-supervised learning objectives, without requiring domain-specific fine-tuning. We propose methods for using these models in tandem to predict when a robot should begin preparing responses, take turns, and handle potential interruptions. We evaluated the proposed system in a within-subject study against a traditional baseline system, using the Furhat robot with 39 adults in a conversational setting, in combination with a large language model for autonomous response generation. The results show that participants significantly prefer the proposed system, and it significantly reduces response delays and interruptions.

Koji Inoue, Divesh Lala, Gabriel Skantze et al.

In human conversations, short backchannel utterances such as "yeah" and "oh" play a crucial role in facilitating smooth and engaging dialogue. These backchannels signal attentiveness and understanding without interrupting the speaker, making their accurate prediction essential for creating more natural conversational agents. This paper proposes a novel method for real-time, continuous backchannel prediction using a fine-tuned Voice Activity Projection (VAP) model. While existing approaches have relied on turn-based or artificially balanced datasets, our approach predicts both the timing and type of backchannels in a continuous and frame-wise manner on unbalanced, real-world datasets. We first pre-train the VAP model on a general dialogue corpus to capture conversational dynamics and then fine-tune it on a specialized dataset focused on backchannel behavior. Experimental results demonstrate that our model outperforms baseline methods in both timing and type prediction tasks, achieving robust performance in real-time environments. This research offers a promising step toward more responsive and human-like dialogue systems, with implications for interactive spoken dialogue applications such as virtual assistants and robots.

Koji Inoue, Divesh Lala, Keiko Ochi et al.

Establishing evaluation schemes for spoken dialogue systems is important, but it can also be challenging. While subjective evaluations are commonly used in user experiments, objective evaluations are necessary for research comparison and reproducibility. To address this issue, we propose a framework for indirectly but objectively evaluating systems based on users' behaviors. In this paper, to this end, we investigate the relationship between user behaviors and subjective evaluation scores in social dialogue tasks: attentive listening, job interview, and first-meeting conversation. The results reveal that in dialogue tasks where user utterances are primary, such as attentive listening and job interview, indicators like the number of utterances and words play a significant role in evaluation. Observing disfluency also can indicate the effectiveness of formal tasks, such as job interview. On the other hand, in dialogue tasks with high interactivity, such as first-meeting conversation, behaviors related to turn-taking, like average switch pause length, become more important. These findings suggest that selecting appropriate user behaviors can provide valuable insights for objective evaluation in each social dialogue task.

Yu Wang, Leyi Lao, Langchu Huang et al.

Backchannels and fillers are important linguistic expressions in dialogue, but are under-represented in modern transformer-based language models (LMs). Our work studies the representation of them in language models using three fine-tuning strategies. The models are trained on three dialogue corpora in English and Japanese, where backchannels and fillers are preserved and annotated, to investigate how fine-tuning can help LMs learn their representations. We first apply clustering analysis to the learnt representation of backchannels and fillers, and have found increased silhouette scores in representations from fine-tuned models, which suggests that fine-tuning enables LMs to distinguish the nuanced semantic variation in different backchannel and filler use. We also use natural language generation (NLG) metrics to confirm that the utterances generated by fine-tuned language models resemble human-produced utterances more closely. Our findings suggest the potentials of transforming general LMs into conversational LMs that are more capable of producing human-like languages adequately.

Nikita Torgashov, Gustav Eje Henter, Gabriel Skantze

We present VoXtream, a fully autoregressive, zero-shot streaming text-to-speech (TTS) system for real-time use that begins speaking from the first word. VoXtream directly maps incoming phonemes to audio tokens using a monotonic alignment scheme and a dynamic look-ahead that does not delay onset. Built around an incremental phoneme transformer, a temporal transformer predicting semantic and duration tokens, and a depth transformer producing acoustic tokens, VoXtream achieves, to our knowledge, the lowest initial delay among publicly available streaming TTS: 102 ms on GPU. Despite being trained on a mid-scale 9k-hour corpus, it matches or surpasses larger baselines on several metrics, while delivering competitive quality in both output- and full-streaming settings. Demo and code are available at https://herimor.github.io/voxtream.

Bram Willemsen, Gabriel Skantze

In this paper, we explore the use of a text-only, autoregressive language modeling approach for the extraction of referring expressions from visually grounded dialogue. More specifically, the aim is to investigate the extent to which the linguistic context alone can inform the detection of mentions that have a (visually perceivable) referent in the visual context of the conversation. To this end, we adapt a pretrained large language model (LLM) to perform a relatively course-grained annotation of mention spans in unfolding conversations by demarcating mention span boundaries in text via next-token prediction. Our findings indicate that even when using a moderately sized LLM, relatively small datasets, and parameter-efficient fine-tuning, a text-only approach can be effective, highlighting the relative importance of the linguistic context for this task. Nevertheless, we argue that the task represents an inherently multimodal problem and discuss limitations fundamental to unimodal approaches.

Julio Cesar Cavalcanti, Gabriel Skantze

Turn-taking in dialogue follows universal constraints but also varies significantly. This study examines how demographic (sex, age, education) and individual factors shape turn-taking using a large dataset of US English conversations (Fisher). We analyze Transition Floor Offset (TFO) and find notable interspeaker variation. Sex and age have small but significant effects female speakers and older individuals exhibit slightly shorter offsets - while education shows no effect. Lighter topics correlate with shorter TFOs. However, individual differences have a greater impact, driven by a strong idiosyncratic and an even stronger "dyadosyncratic" component - speakers in a dyad resemble each other more than they resemble themselves in different dyads. This suggests that the dyadic relationship and joint activity are the strongest determinants of TFO, outweighing demographic influences.

Livia Qian, Carol Figueroa, Gabriel Skantze

Vocal feedback (e.g., `mhm', `yeah', `okay') is an important component of spoken dialogue and is crucial to ensuring common ground in conversational systems. The exact meaning of such feedback is conveyed through both lexical and prosodic form. In this work, we investigate the perceived prosodic similarity of vocal feedback with the same lexical form, and to what extent existing speech representations reflect such similarities. A triadic comparison task with recruited participants is used to measure perceived similarity of feedback responses taken from two different datasets. We find that spectral and self-supervised speech representations encode prosody better than extracted pitch features, especially in the case of feedback from the same speaker. We also find that it is possible to further condense and align the representations to human perception through contrastive learning.

Livia Qian, Gabriel Skantze

Short feedback responses, such as backchannels, play an important role in spoken dialogue. So far, most of the modeling of feedback responses has focused on their timing, often neglecting how their lexical and prosodic form influence their contextual appropriateness and conversational function. In this paper, we investigate the possibility of embedding short dialogue contexts and feedback responses in the same representation space using a contrastive learning objective. In our evaluation, we primarily focus on how such embeddings can be used as a context-feedback appropriateness metric and thus for feedback response ranking in U.S. English dialogues. Our results show that the model outperforms humans given the same ranking task and that the learned embeddings carry information about the conversational function of feedback responses.

Bing'er Jiang, Erik Ekstedt, Gabriel Skantze

Filled pauses (or fillers), such as "uh" and "um", are frequent in spontaneous speech and can serve as a turn-holding cue for the listener, indicating that the current speaker is not done yet. In this paper, we use the recently proposed Voice Activity Projection (VAP) model, which is a deep learning model trained to predict the dynamics of conversation, to analyse the effects of filled pauses on the expected turn-hold probability. The results show that, while filled pauses do indeed have a turn-holding effect, it is perhaps not as strong as could be expected, probably due to the redundancy of other cues. We also find that the prosodic properties and position of the filler has a significant effect on the turn-hold probability. However, contrary to what has been suggested in previous work, there is no difference between "uh" and "um" in this regard.

Bing'er Jiang, Erik Ekstedt, Gabriel Skantze

Previous approaches to turn-taking and response generation in conversational systems have treated it as a two-stage process: First, the end of a turn is detected (based on conversation history), then the system generates an appropriate response. Humans, however, do not take the turn just because it is likely, but also consider whether what they want to say fits the position. In this paper, we present a model (an extension of TurnGPT) that conditions the end-of-turn prediction on both conversation history and what the next speaker wants to say. We found that our model consistently outperforms the baseline model in a variety of metrics. The improvement is most prominent in two scenarios where turn predictions can be ambiguous solely from the conversation history: 1) when the current utterance contains a statement followed by a question; 2) when the end of the current utterance semantically matches the response. Treating the turn-prediction and response-ranking as a one-stage process, our findings suggest that our model can be used as an incremental response ranker, which can be applied in various settings.

Gabriel Skantze, Bram Willemsen

This paper presents CoLLIE: a simple, yet effective model for continual learning of how language is grounded in vision. Given a pre-trained multimodal embedding model, where language and images are projected in the same semantic space (in this case CLIP by OpenAI), CoLLIE learns a transformation function that adjusts the language embeddings when needed to accommodate new language use. This is done by predicting the difference vector that needs to be applied, as well as a scaling factor for this vector, so that the adjustment is only applied when needed. Unlike traditional few-shot learning, the model does not just learn new classes and labels, but can also generalize to similar language use and leverage semantic compositionality. We verify the model's performance on two different tasks of identifying the targets of referring expressions, where it has to learn new language use. The results show that the model can efficiently learn and generalize from only a few examples, with little interference with the model's original zero-shot performance.

Erik Ekstedt, Gabriel Skantze

Syntactic and pragmatic completeness is known to be important for turn-taking prediction, but so far machine learning models of turn-taking have used such linguistic information in a limited way. In this paper, we introduce TurnGPT, a transformer-based language model for predicting turn-shifts in spoken dialog. The model has been trained and evaluated on a variety of written and spoken dialog datasets. We show that the model outperforms two baselines used in prior work. We also report on an ablation study, as well as attention and gradient analyses, which show that the model is able to utilize the dialog context and pragmatic completeness for turn-taking prediction. Finally, we explore the model's potential in not only detecting, but also projecting, turn-completions.

Matthew Roddy, Gabriel Skantze, Naomi Harte

In human conversational interactions, turn-taking exchanges can be coordinated using cues from multiple modalities. To design spoken dialog systems that can conduct fluid interactions it is desirable to incorporate cues from separate modalities into turn-taking models. We propose that there is an appropriate temporal granularity at which modalities should be modeled. We design a multiscale RNN architecture to model modalities at separate timescales in a continuous manner. Our results show that modeling linguistic and acoustic features at separate temporal rates can be beneficial for turn-taking modeling. We also show that our approach can be used to incorporate gaze features into turn-taking models.

Matthew Roddy, Gabriel Skantze, Naomi Harte

For spoken dialog systems to conduct fluid conversational interactions with users, the systems must be sensitive to turn-taking cues produced by a user. Models should be designed so that effective decisions can be made as to when it is appropriate, or not, for the system to speak. Traditional end-of-turn models, where decisions are made at utterance end-points, are limited in their ability to model fast turn-switches and overlap. A more flexible approach is to model turn-taking in a continuous manner using RNNs, where the system predicts speech probability scores for discrete frames within a future window. The continuous predictions represent generalized turn-taking behaviors observed in the training data and can be applied to make decisions that are not just limited to end-of-turn detection. In this paper, we investigate optimal speech-related feature sets for making predictions at pauses and overlaps in conversation. We find that while traditional acoustic features perform well, part-of-speech features generally perform worse than word features. We show that our current models outperform previously reported baselines.