Paper deep dive
INTERACT: An AI-Driven Extended Reality Framework for Accesible Communication Featuring Real-Time Sign Language Interpretation and Emotion Recognition
Nikolaos D. Tantaroudas, Andrew J. McCracken, Ilias Karachalios, Evangelos Papatheou
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 99%
Last extracted: 4/10/2026, 3:17:18 AM
Summary
INTERACT is an AI-driven Extended Reality (XR) framework designed to enhance accessibility in video conferencing for deaf, hard-of-hearing, and multilingual users. It integrates real-time speech-to-text (Whisper), multilingual translation (NLLB), International Sign Language (ISL) rendering via 3D avatars (MediaPipe), and emotion recognition (DistilRoBERTa) within a unified immersive environment deployed on Meta Quest 3 headsets.
Entities (7)
Relation Signals (5)
INTERACT → builton → CORTEX2
confidence 100% · Constructed upon the CORTEX2 framework
INTERACT → useslibrary → Google MediaPipe
confidence 100% · INTERACT... harnesses... Google MediaPipe for gesture extraction
INTERACT → usesmodel → Whisper
confidence 100% · INTERACT... harnesses cutting-edge AI models, including Whisper for speech recognition
INTERACT → usesmodel → NLLB
confidence 100% · INTERACT... harnesses... NLLB for multilingual translation
INTERACT → usesmodel → RoBERTa
confidence 100% · INTERACT... harnesses... RoBERTa for emotion classification
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Video conferencing has become central to professional collaboration, yet most platforms offer limited support for deaf, hard-of-hearing, and multilingual users. The World Health Organisation estimates that over 430 million people worldwide require rehabilitation for disabling hearing loss, a figure projected to exceed 700 million by 2050. Conventional accessibility measures remain constrained by high costs, limited availability, and logistical barriers, while Extended Reality (XR) technologies open new possibilities for immersive and inclusive communication. This paper presents INTERACT (Inclusive Networking for Translation and Embodied Real-Time Augmented Communication Tool), an AI-driven XR platform that integrates real-time speech-to-text conversion, International Sign Language (ISL) rendering through 3D avatars, multilingual translation, and emotion recognition within an immersive virtual environment. Built on the CORTEX2 framework and deployed on Meta Quest 3 headsets, INTERACT combines Whisper for speech recognition, NLLB for multilingual translation, RoBERTa for emotion classification, and Google MediaPipe for gesture extraction. Pilot evaluations were conducted in two phases, first with technical experts from academia and industry, and subsequently with members of the deaf community. The trials reported 92% user satisfaction, transcription accuracy above 85%, and 90% emotion-detection precision, with a mean overall experience rating of 4.6 out of 5.0 and 90% of participants willing to take part in further testing. The results highlight strong potential for advancing accessibility across educational, cultural, and professional settings. An extended version of this work, including full pilot data and implementation details, has been published as an Open Research Europe article [Tantaroudas et al., 2026a].
Tags
Links
- Source: https://arxiv.org/abs/2604.05605v1
- Canonical: https://arxiv.org/abs/2604.05605v1
Trouble viewing inline? Open PDF directly →
Full Text
54,072 characters extracted from source content.
Expand or collapse full text
INTERACT: An AI-Driven Extended Reality Framework for Accessible Communication Featuring Real-Time Sign Language Interpretation and Emotion Recognition Nikolaos D. Tantaroudas ∗1 , Andrew J. McCracken 2 , Ilias Karachalios 3 , and Evangelos Papatheou 4 1 Institute of Communications and Computer Systems (ICCS), Iroon Polytechneiou 9, 15773 Zografou, Athens, Greece 2 DASKALOS-APPS, 183 Rue de l’Abbé Griffon, 01960 Péronnas, France 3 National Technical University of Athens, Leof. Alimou, Katechaki, Zografou, 15772 Athens, Greece 4 Exeter Small-Scale Robotics Laboratory, Engineering Department, University of Exeter, Exeter EX4 4QF, UK Abstract Video conferencing has become a cornerstone of professional collaboration; however, the majority of existing platforms provide inadequate support for deaf, hard-of-hearing, and multilingual users. The World Health Organization estimates that more than 430 million individuals globally need rehabilitation services for disabling hearing loss, with projections indicating this number will surpass 700 million by 2050. Conventional accessibility measures are constrained by prohibitive costs, scarce availability, and logistical difficulties. Extended Reality (XR) technologies present novel avenues for building immersive and inclusive commu- nication spaces. This paper introduces INTERACT (Inclusive Networking for Translation and Embodied Real-Time Augmented Communication Tool), an AI-driven XR platform that unifies real-time speech-to-text conversion, International Sign Language (ISL) render- ing through 3D avatars, multilingual capability, and emotion recognition within immersive virtual settings. Constructed upon the CORTEX2 framework and deployed on Meta Quest 3 headsets, the platform harnesses cutting-edge AI models, including Whisper for speech recog- nition, NLLB for multilingual translation, RoBERTa for emotion classification, and Google MediaPipe for gesture extraction. Pilot evaluations were carried out in two stages: one in- volving technical specialists from academic and industrial sectors and another engaging deaf community members directly. The pilot evaluations yielded 92% user satisfaction, transcrip- tion accuracy surpassing 85%, and 90% emotion detection precision. Participants gave a mean overall experience score of 4.6 out of 5.0, with 90% indicating willingness to engage in subsequent testing rounds. The platform reliably accommodates up to 1,000 simultaneous users with negligible latency increase. Stress testing confirmed robust performance under high-demand conditions, with 10,000 concurrent requests processed at 900–1,000 requests per second with zero failures. This work constitutes the first XR-based video conferenc- ing platform that merges real-time ISL rendering via 3D avatars with multilingual speech conversion and emotion-aware feedback within a unified, immersive solution. The findings underscore considerable potential for reshaping accessibility across educational, cultural, and professional domains. An extended version of this work, including comprehensive pilot data and detailed implementation, has been published as an Open Research Europe article [Tan- taroudas et al., 2026a]. ∗ Corresponding author: nikolaos.tantaroudas@iccs.gr 1 arXiv:2604.05605v1 [cs.CE] 7 Apr 2026 Keywords: Extended Reality, Artificial Intelligence, International Sign Language, Speech-to- Text Conversion, Multilingual Translation, Accessibility, Deaf and Hard-of-Hearing, 3D Avatar, Emotion Recognition, Video Conferencing 1 Introduction The accelerated transition towards remote and hybrid working models has rendered video con- ferencing indispensable for professional, educational, and social exchanges worldwide. Never- theless, mainstream video conferencing solutions frequently fall short of addressing the varied requirements of deaf, hard-of-hearing, and multilingual participants, thereby erecting consider- able obstacles to their full engagement and contribution [Alford et al., 2023]. The World Health Organization reports that over 430 million people across the globe need rehabilitation for dis- abling hearing loss, a number expected to exceed 700 million by 2050 [World Health Organization, 2021]. Concurrently, linguistic barriers within multilingual teams impede effective communica- tion and cooperative outcomes [Franceschini et al., 2020]. This accessibility deficit is especially pronounced in professional environments where subtle communication, encompassing emotional tone and non-verbal signals, is vital for productive collaboration and decision-making. Conventional accessibility provisions, including human interpreters and manual captioning, are hampered by substantial costs, restricted availability, interpreter fatigue, and scheduling challenges [Rodríguez-Correa et al., 2023]. Although machine translation and automatic speech recognition technologies have progressed markedly, their incorporation into coherent, real-time accessibility systems remains piecemeal [Liu et al., 2020]. Extended Reality (XR) technologies, spanning Virtual Reality (VR), Augmented Reality (AR), and Mixed Reality (MR), offer ex- traordinary possibilities for crafting immersive, inclusive communication settings that transcend physical constraints [Hirzle et al., 2023]. The spatial and embodied character of XR environments furnishes a uniquely apt medium for sign language delivery, given that three-dimensional avatars can reproduce the complete spatial grammar of signed languages in ways that conventional flat displays cannot. This study introduces and validates INTERACT (Inclusive Networking for Translation and Embodied Real-Time Augmented Communication Tool), a pioneering XR platform conceived within the CORTEX2 Horizon Europe initiative that tackles these accessibility obstacles through a unified artificial intelligence strategy. INTERACT merges real-time speech-to-text conversion, multilingual translation, International Sign Language (ISL) interpretation through animated 3D avatars, and emotion recognition within a single immersive environment deployable on Meta Quest 3 headsets and desktop interfaces. The central contributions of this work comprise: (1) the creation and validation of the first XR-based video conferencing platform that fuses real-time ISL rendering with multilingual speech conversion and emotion-aware feedback; (2) a thorough technical architecture employing state- of-the-art AI models for speech processing, translation, and gesture synthesis; (3) empirical validation via pilot evaluations attaining 92% user satisfaction among deaf community members and accessibility professionals, with a mean overall experience score of 4.6/5.0; and (4) demon- stration of system scalability accommodating up to 1,000 simulated concurrent users. This work builds upon our preliminary findings presented at the Salento XR 2025 conference [Tantaroudas et al., 2026b] and extends the comprehensive analysis presented in our Open Research Europe publication [Tantaroudas et al., 2026a], with additional discussion of architectural considera- tions and deployment insights. A companion paper [Tantaroudas et al., 2026c] further explores the AI-based service pipeline for inclusive language learning in immersive XR settings, detailing speech translation and sign language integration capabilities that complement the INTERACT platform. The remainder of this paper is organised as follows: Section 2 examines prior work spanning XR accessibility, sign language translation, speech recognition, and emotion analysis. Section 3 2 outlines the system architecture and technical realisation. Section 4 details the pilot method- ology and validation outcomes. Section 5 considers implications and constraints, and Section 6 concludes with prospective directions. 2 Related Work and Background 2.1 Extended Reality for Accessibility The convergence of XR technologies and accessibility has attracted growing research interest, with applications encompassing educational support, vocational training, and social interaction for individuals with disabilities [Hirzle et al., 2023, Hosseinkashi et al., 2023]. Serafin et al. [Serafin et al., 2023] performed an extensive review of VR applications for individuals with hearing impair- ments, identifying substantial promise in immersive learning environments while underscoring the necessity for integrated communication assistance. Recent advances in consumer-grade XR hardware, notably the Meta Quest series, have democratised access to immersive experiences that were formerly confined to specialised research laboratories [Anwar et al., 2023]. Hirzle et al. [Hirzle et al., 2023] delivered a scoping review of XR and AI integration, cat- aloguing emerging applications across healthcare, education, and professional training. Their analysis pinpointed accessibility enhancement as a crucial yet insufficiently explored application domain, particularly concerning real-time communication support for deaf and hard-of-hearing users in collaborative contexts. Notwithstanding mounting interest in immersive accessibility, current platforms predominantly target single modalities, providing either captioning or two- dimensional sign language support through mobile applications, but seldom combine multiple accessibility channels together with emotional context within a cohesive XR environment. 2.2 Sign Language Translation and Generation Sign language translation research has progressed considerably through deep learning break- throughs, especially transformer architectures capable of capturing the temporal and spatial intricacies of signed languages [Yin et al., 2023, Wu et al., 2023]. Camgöz et al. [Camgöz et al., 2020] pioneered end-to-end neural sign language translation employing encoder-decoder networks, attaining competitive results on benchmark datasets. Subsequent work by Zhou et al. [Zhou et al., 2023] demonstrated gloss-free methodologies utilising visual-language pretraining, thereby diminishing reliance on intermediate linguistic representations. Avatar-based sign language generation poses challenges distinct from recognition, demand- ing precise synthesis of hand configurations, arm movements, facial expressions, and body pos- ture [Gu et al., 2022]. Gibet and Marteau [Gibet and Marteau, 2023] delineated the multimodal challenges inherent in text-to-sign generation, stressing the significance of linguistic fidelity and cultural sensitivity. Fink et al. [Fink et al., 2023] showcased lightweight transformer models for sign language dictionary applications, pointing towards pathways for real-time generation amenable to communication platforms. Crucially, existing systems predominantly concentrate on national sign languages such as American Sign Language (ASL) or British Sign Language (BSL). International Sign Language (ISL), employed in international contexts including conferences, sporting events, and online plat- forms, remains largely unexplored despite its capacity for broader accessibility across varied deaf communities [European Union of the Deaf, 2018]. The European Union of the Deaf acknowledges ISL as an effective auxiliary language facilitating communication across national sign language boundaries [HandSpeak, 2024]. This lacuna in ISL-oriented research drove our decision to adopt ISL as the primary sign language modality for INTERACT, targeting an underserved yet in- ternationally pertinent communication need. Our companion work [Tantaroudas et al., 2026c] provides additional detail on the sign language integration pipeline within immersive XR learning environments. 3 2.3 Speech Recognition and Multilingual Translation Automatic speech recognition (ASR) has attained near-human performance through large-scale transformer models trained on heterogeneous multilingual corpora [SpreadTheSign, 2024]. Ope- nAI’s Whisper model [Radford et al., 2023] exhibits robust speech recognition spanning 99 lan- guages with competitive accuracy on standard benchmarks, even under adverse acoustic con- ditions. The model’s multilingual capabilities and open-source nature render it especially suit- able for accessibility applications demanding language-agnostic speech processing. Neural ma- chine translation (NMT) has similarly progressed through transformer architectures and mul- tilingual training paradigms [Stahlberg, 2020]. Meta AI’s No Language Left Behind (NLLB) initiative [NLLB Team et al., 2022] represents the most ambitious effort towards universal trans- lation, supporting more than 200 languages including numerous low-resource languages previ- ously neglected by translation technology. The combination of ASR and NMT systems enables speech-to-text pipelines facilitating real-time multilingual communication [Barrault et al., 2023]. Nevertheless, attaining the sub-second latencies demanded by natural conversational flow in live conferencing remains a considerable engineering challenge. 2.4 Sentiment Analysis and Affective Computing Affective computing research seeks to equip machines with the ability to recognise, interpret, and respond to human emotional states [Picard, 1997]. Transformer-based models have pro- duced marked improvements in text-based emotion analysis, with RoBERTa [Liu et al., 2019] and its derivatives exhibiting strong performance across benchmark datasets. Hartmann’s Distil- RoBERTa model [Hartmann, 2022] furnishes efficient emotion classification supporting real-time applications, discriminating among multiple emotional categories including joy, sadness, anger, fear, surprise, and disgust. Incorporating emotion analysis into communication platforms enables emotional context preservation that purely textual or signed translations might otherwise forfeit [Zhang et al., 2023]. For deaf and hard-of-hearing users, emotional cues ordinarily conveyed via vocal prosody become inaccessible, rendering explicit emotion representation especially valuable [Chen et al., 2022]. Studies show that emotional context exerts a substantial influence on comprehension and engagement in collaborative environments [Mitchell and Karchmer, 2004]. Despite these acknowledged advantages, no extant XR communication platform has integrated real-time emo- tion analysis alongside sign language generation and multilingual transcription within a single cohesive system, a gap that INTERACT squarely addresses. 2.5 Meeting Summarisation Large language models (LLMs) have revolutionised automatic summarisation capabilities, en- abling coherent, contextually fitting summaries of lengthy text sequences [Brown et al., 2020]. Meeting summarisation poses distinctive challenges owing to multi-speaker dynamics, informal register, and the imperative to capture action items alongside discussion content [Asthana et al., 2023]. Fine-tuned models such as BART exhibit strong performance on dialogue summarisation tasks [Liu et al., 2022], while recent work investigates faithfulness constraints to ensure summaries faithfully mirror source content [Roit et al., 2023]. Embedding automated summarisation within an accessible XR platform serves a dual purpose: it affords asynchronous access for participants who may have missed segments of a meeting, and it generates a persistent record reviewable in multiple languages. A more detailed treatment of the summarisation and translation service pipeline can be found in [Tantaroudas et al., 2026c]. 4 3 System Architecture 3.1 Overview INTERACT functions within the CORTEX2 ecosystem, drawing on the Mediation Gateway infrastructure for communication orchestration and the Rainbow SDK for conferencing function- alities [CORTEX2 Project, 2024, Alcatel-Lucent Enterprise, 2024]. The system architecture is composed of five integrated AI service modules: (1) Speech-to-Text Conversion, (2) Multilingual Translation, (3) Sign Language Generation, (4) Emotion Analysis, and (5) Meeting Summarisa- tion. These modules exchange data through standardised APIs, permitting modular deployment and independent scaling. This modular design ethos guarantees that individual components can be upgraded or substituted as superior AI models emerge, without necessitating a redesign of the overall system. The complete system architecture is illustrated in Figure 1, depicting how the CORTEX2 Mediation Gateway, Rainbow SDK, and the five AI service modules are intercon- nected, with data flow paths traced from audio capture through to sign language avatar rendering and emotion feedback delivery. Figure 1: INTERACT system architecture showing the integration among CORTEX2 infrastruc- ture components and INTERACT AI modules. The diagram traces the data flow from audio input through the Mediation Gateway to each AI service module and ultimately to the XR ren- dering layer. 3.2 Speech-to-Text Conversion Module The speech-to-text module utilises OpenAI’s Whisper model, specifically the large-v2 variant optimised for accuracy under varied acoustic conditions [Radford et al., 2023]. Audio processing employs chunked transcription with 1-second segments and 0.5-second overlap to balance latency against word fragmentation. The implementation handles audio via a WebSocket interface fa- cilitating real-time streaming from the Rainbow SDK. This chunking strategy was empirically established through iterative experimentation: shorter segments caused excessive word fragmen- tation, whereas longer segments pushed end-to-end latency beyond acceptable thresholds for conversational usage. The successive stages of the speech-to-text processing pipeline are de- picted in Figure 2, visualising the sequential flow from raw audio chunking through Whisper model inference to the final text assembly and punctuation restoration steps that yield the out- put transcription stream. The system attains transcription latency averaging 1.2 seconds from speech onset to text display, with accuracy surpassing 85% on conversational speech in controlled 5 settings. Noise resilience testing showed that accuracy remained above 80% at signal-to-noise ratios as low as 15 dB. Figure 2: Speech-to-Text Processing Pipeline illustrating the audio chunking, Whisper model inference, and text assembly stages that produce the final transcription output stream. 3.3 Multilingual Translation Module Translation services employ Meta AI’s NLLB-200 distilled model (600M parameters), striking a balance between translation quality and inference speed demands [NLLB Team et al., 2022]. The present implementation provides English-to-French translation as the primary use case, with the architecture engineered for straightforward extension to further language pairs, including German and Spanish, scheduled for upcoming releases. The distilled variant was chosen over the full 3.3B-parameter model to satisfy the real-time latency requirements of live conferencing, as initial benchmarking revealed that the larger model introduced unacceptable delays in excess of 3 seconds per sentence. Figure 3 presents the multilingual translation module architecture, showing how source text from the speech recognition module enters the NLLB translation engine and how the translated output is routed to both the sign language generation module and the direct text display for participants. Translation latency averages 0.8 seconds per sentence, enabling near-real-time presentation alongside source text transcription. Quality evaluation using BLEU scores indicates performance on par with production-grade translation services for the supported language pairs. Figure 3: Multilingual Translation Flow diagram depicting source text input, language detection, NLLB model inference, and translated output distribution to downstream modules. 3.4 Sign Language Generation Module The sign language module constitutes the most technically elaborate INTERACT component, converting text input into animated 3D avatar signing sequences. Development proceeded through a multi-stage process: (1) ISL video corpus acquisition from validated deaf community sources and online ISL dictionaries (e.g., HandSpeak [HandSpeak, 2024]); (2) skeletal landmark extraction using Google MediaPipe [Lugaresi et al., 2019]; (3) animation curve generation; and (4) Unity-based avatar rendering. Each stage demanded meticulous calibration to preserve the 6 spatial and temporal fidelity of the original signing, since even slight distortions in hand position- ing or timing can modify or obscure the intended meaning. The gesture extraction procedure is illustrated in Figure 4, presenting a three-stage pipeline: the original ISL video frame captured from validated sign language sources, the MediaPipe Holistic landmark detection overlay identi- fying hand, face, and body keypoints, and the resulting extracted 3D coordinate array forming the foundation for driving avatar animations. Figure 4: International Sign Language gesture extraction pipeline. From left to right: original ISL video frame; MediaPipe Holistic landmark detection overlay; extracted 3D coordinate array used to drive avatar animations. The corpus encompasses 747 ISL videos representing a 750-sign vocabulary that covers es- sential business and professional communication concepts. Each video was processed to extract hand, face, and body landmarks at 30 frames per second, producing animation sequences compat- ible with the Unity humanoid rig system. Signs were selected on the basis of frequency analysis of professional meeting transcripts, prioritising vocabulary items most commonly encountered in business deliberations, project reviews, and educational presentations. Figure 5 showcases the re- sulting 3D avatar performing ISL signs within the virtual office environment, demonstrating how the extracted motion data translates into realistic signing movements, including hand shapes, arm movements, and body positioning, that deaf users can interpret in real time. Sign lookup uses a dictionary-based approach with planned extensions towards sequence-to-sequence neural generation. For vocabulary items absent from the dictionary, the system reverts to fingerspelling via a character-to-sign mapping. 3.5 Emotion Analysis Module Emotional context is maintained through real-time emotion analysis using a fine-tuned Distil- RoBERTa model [Hartmann, 2022, Sanh et al., 2019]. The system categorises transcribed text into six emotional classes: joy, sadness, anger, fear, surprise, and neutral. Classification outcomes drive avatar facial expression adjustments and optional emoji overlays on transcription displays. This two-channel emotional feedback, visual expression on the avatar and textual emoji annota- tion, was conceived to accommodate differing user preferences and varying degrees of attention to the avatar during meetings. An illustration of the emotion analysis integration is provided in Figure 6, showing how transcribed speech is annotated with emotion labels generated by the DistilRoBERTa classifier and how these labels are manifested in corresponding alterations to the avatar’s facial expressions alongside optional emoji indicators in the text display. The emotion module operates with sub-200 ms latency, enabling emotional context display concurrent with 7 Figure 5: 3D Avatar performing ISL signs within the virtual environment, showing the avatar from multiple angles during signing sequences. text transcription. Validation against human annotations yields 90%+ precision for the principal emotion categories. 3.6 Meeting Summarisation Module Extended meetings benefit from automatic summarisation employing a BART-Large model fine- tuned on the SAMSum conversational dataset [Schmid, 2021]. The summarisation module pro- cesses accumulated transcripts at configurable intervals (typically 15-minute segments or on demand), producing structured summaries encompassing key discussion points, decisions, and action items. This functionality is particularly beneficial for participants joining meetings late or needing to review content asynchronously, as the summaries furnish a succinct yet thorough record of proceedings. The meeting summarisation workflow is depicted in Figure 7, illustrat- ing how transcripts accumulate during a conference session and are subsequently processed by the BART model either at scheduled intervals or upon request, generating formatted meeting minutes that include principal discussion points, decisions taken, and action items assigned. Sum- maries undergo optional translation via the NLLB module, enabling distribution of multilingual meeting minutes. The combined transcription-summarisation pipeline supports accessibility for participants reviewing meeting content at a later time. 3.7 XR Environment and Deployment The INTERACT immersive environment was built in Unity 2022.3 LTS targeting Meta Quest 3 deployment via the Meta XR All-in-One SDK. The virtual space recreates a professional confer- ence room with seated positions for up to eight participants, a central presentation area, and the signing avatar positioned for optimal visibility. The environment was designed in accordance with established XR usability principles, with particular care given to comfortable viewing distances for the signing avatar and legible text overlay placement that avoids occluding other participants or presentation content. The complete virtual office environment is presented in Figure 8, show- ing the immersive meeting space as experienced by participants wearing Meta Quest 3 headsets, highlighting the spatial configuration of seating, the central presentation wall for shared content, and the prominent positioning of the signing avatar to ensure it remains readily viewable during conversations. 8 Figure 6: Emotion Analysis Output displaying transcribed text with emotion labels and corre- sponding avatar facial expression modifications. Figure 7: Meeting Summarisation Workflow showing transcript accumulation, BART model processing, and formatted summary output including key points, decisions, and action items. 9 Figure 8: Virtual Office Environment depicting the complete meeting space with participant seating positions and signing avatar placement within the immersive XR setting. Rainbow SDK integration facilitates audio capture, user presence management, and confer- ence orchestration within the Unity environment [Alcatel-Lucent Enterprise, 2024]. API con- nections to the AI service modules employ WebSocket protocols for streaming data and REST endpoints for configuration and status queries. Figure 9 details the Rainbow SDK integration architecture within Unity, illustrating how the SDK manages audio streaming from participants, presence detection, and conferencing features while the INTERACT AI modules process the communication content in parallel through the WebSocket and REST API connections. Fur- thermore, Figure 10 shows how deaf participants can engage with the conversation that hearing individuals are conducting within Rainbow. The hearing individuals connect into the VR scene using the Rainbow SDK (shown in Figure 9), and their transcribed speech is translated and animated in ISL through the avatar so that the deaf individual can comprehend the exchange. Figure 9: Rainbow SDK Integration within Unity showing the SDK configuration and commu- nication flow between conferencing infrastructure and AI processing modules. 10 Figure 10: Avatar Animation and Emotion Analysis within the VR scene whilst hearing individ- uals communicate using Rainbow SDK, demonstrating the real-time ISL rendering pipeline. 3.8 Infrastructure and Scalability Backend services are deployed on AWS infrastructure using EC2 G4DN.xlarge instances fitted with NVIDIA T4 GPUs (16 GB VRAM) supporting concurrent model inference. Load balancing and auto-scaling configurations permit horizontal scaling during peak usage, with testing con- firming support for up to 1,000 simulated concurrent users without service degradation. The GPU-accelerated inference pipeline ensures that the computationally demanding Whisper and NLLB models sustain acceptable latency even under elevated concurrent load. 4 Pilot Study 4.1 Methodology Pilot validation adopted a two-phase approach aligned with CORTEX2 project milestones. The first validation demonstration (May 2025) engaged technical specialists in AI, Human-Computer Interaction (HCI), and accessibility research to assess system performance and pinpoint im- provement opportunities. The second demonstration (June 2025) carried out validation with deaf community stakeholders, evaluating real-world accessibility value and linguistic authen- ticity. Table 1 provides the participant demographics for the live demonstrations. All deaf participants in Demo 2 reported native or near-native ISL fluency and professional experience in deaf education or interpretation services. All Demo 2 participants indicated familiarity with ISL and professional experience in deaf education, interpretation services, or regular engagement with deaf communities. Across both workshops, 40% of participants had no prior VR/XR experience, 40% had used VR on one or two occasions, and 20% reported moderate experience, offering a representative cross-section of prospective end users with diverse technological familiarity. 11 Table 1: Participant Demographics in the two validation demonstrations. CharacteristicDemo 1 (May 2025)Demo 2 (June 2025) Number of participants46 Participant typeTechnical expertsDeaf community stakeholders VR experience: None40%40% VR experience: Once/twice40%40% VR experience: Moderate20%20% ISL fluencyVariedNative/near-native Evaluation Protocol: Participants engaged with INTERACT through structured tasks comprising: (1) one-on-one conversation scenarios, (2) group discussion simulations, and (3) pre- sentation viewing with real-time translation. Sessions were recorded with participant consent for subsequent analysis. Post-session questionnaires assessed usability, satisfaction, and perceived accessibility value, emotion detection accuracy, and willingness to participate in future testing across 18 items using 5-point Likert scales, categorical satisfaction ranges, and open-ended qual- itative prompts. Table 2 presents the technical Key Performance Indicators (KPIs) attained during the two demonstration validations. Table 2: Key Performance Indicators (KPIs) achieved during pilot validation. KPITarget Achieved User satisfaction≥85%92% Transcription accuracy≥85%>85% Emotion detection precision≥85%90% Concurrent users supported 1,0001,000 Transcription latency<2 s1.2 s Translation latency<2 s0.8 s Error rate under load0%0% 4.2 Detailed Participant Feedback The post-session questionnaire captured participant evaluations across multiple dimensions of the INTERACT experience. Table 3 provides the distribution of Likert-scale responses across key evaluation dimensions. Table 3: Distribution of participant responses across key evaluation dimensions (N = 10). Evaluation Dimension1 2 3 4 5 Overall usability0 0 1 3 6 Transcription quality0 0 1 4 5 Avatar sign clarity0 0 2 4 4 Emotion feedback usefulness 0 0 1 3 6 Willingness to use again0 0 1 2 7 Engagement and Overall Satisfaction: Participants reported considerable engagement improvements relative to conventional meetings lacking accessibility support. The distribution of self-reported engagement increase is presented in Table 4. In total, 80% of participants reported engagement increases of 51% or greater, suggesting the integrated accessibility features substan- 12 tively enhanced the meeting participation experience. The weighted mean satisfaction across all participants corresponds to approximately 85%, meeting the 85% satisfaction KPI target. Table 4: Self-reported engagement increase and overall satisfaction distributions (N = 10). Engagement Increase Percentage of Participants 0–25%0% 26–50%20% 51–75%40% 76–100%40% Emotion Detection Accuracy: Participants evaluated how accurately the emotion anal- ysis reflected the actual emotional tone of conversations. The majority (60%) rated accuracy at 96–100%, with an additional 10% rating it at 91–95%. Two participants (20%), both from Demo 2, rated accuracy below 80%, indicating that emotion classification performance may ne- cessitate further calibration for communication styles characteristic of deaf community discourse, where emotional expression may depend more heavily on visual rather than textual cues. Table 5 presents the perceived emotion detection accuracy across the two demonstrations. Table 5: Perceived emotion detection accuracy (N = 10). Accuracy Range Demo 1 Demo 2 96–100%4 (100%) 2 (33%) 91–95%01 (17%) 80–90%01 (17%) <80%02 (33%) Overall Experience Rating: Participants scored their overall experience on a 1–5 scale. Demo 1 participants uniformly awarded 5/5, while Demo 2 participants gave 4/5, producing a combined mean of 4.6/5.0 (SD = 0.52). Nine participants (90%) expressed willingness to take part in future testing phases, with one participant responding “Maybe,” reflecting strong overall receptivity to the platform and its ongoing development. Qualitative Feedback: Open-ended qualitative responses converged on several thematic improvement areas. The most frequently cited suggestion across both workshops pertained to avatar facial expression enhancement, with six participants independently recommending the integration of facial mimicry, lip-synching, or ISL grammatical markers into the avatar’s face. Demo 1 participants additionally flagged occasional finger movement artefacts where hand shapes appeared unnatural or distorted, while Demo 2 participants stressed the need for adjustable sign- ing speed and replay capability. Further suggestions included larger avatar sizing for improved sign legibility, higher environmental lighting and contrast for subtitle visibility, user-selectable avatar customisation, and expansion towards a multiplayer VR scene. Notably, all ten partici- pants (100%) indicated they would benefit from additional language or sign language options, highlighting the demand for expanded multilingual and multi-sign-language support. Load Testing Results: Scalability testing confirmed system performance under stress con- ditions: (i) Maximum tested concurrent users: 1,000; (i) Request throughput: 900–1,000 re- quests/second; (i) Average response time under load: <2 seconds; (iv) Error rate: 0% up to capacity limit. The load testing results are displayed in Figure 11, showing the system’s through- put, latency, and error rates across progressively increasing concurrent user loads, confirming that the platform maintains stable response times and zero error rates up to 1,000 simultaneous connections before graceful degradation commences. 13 Figure 11: Load Testing Performance Graphs depicting throughput, latency, and error rates across different load levels, confirming zero failures up to 1,000 concurrent connections. 5 Discussion 5.1 Principal Findings INTERACT establishes the viability of integrating multiple AI technologies into a unified XR platform that addresses accessibility barriers in video conferencing. The pilot validation outcomes verify both technical performance meeting established benchmarks and substantive accessibility value as evaluated by deaf community stakeholders. Attaining 92% user satisfaction among participants with professional backgrounds in deaf education and ISL interpretation indicates that the platform approaches practical readiness for real-world deployment. The high overall experience score of 4.6/5.0 and the 90% willingness to engage in future testing further corroborate the platform’s perceived value among its target user community. The modular architecture proved effective for iterative development and component-level optimisation. Decoupling AI services enabled independent performance tuning and seamless integration of enhanced models as they become available. This design principle positions INTERACT for continued refinement as underlying technologies mature. The divergence observed between Demo 1 and Demo 2 ratings, with technical experts uni- formly awarding 5/5 and deaf community participants giving 4/5, yields a meaningful insight. Technical participants assessed the system primarily on technological sophistication and inte- gration quality, whereas deaf community participants applied more stringent criteria regarding linguistic authenticity, avatar expressiveness, and practical communication utility. This pattern highlights the importance of involving end-user communities throughout the development lifecy- cle rather than depending exclusively on technical expert evaluations. Similar findings regarding the value of end-user involvement in XR accessibility design are discussed in [Tantaroudas et al., 2026a]. 14 5.2 Comparison with Prior Work INTERACT advances prior research in XR accessibility and sign language technology along several dimensions. Unlike systems focused exclusively on sign language recognition for input, INTERACT targets the generation direction, enabling hearing speakers’ content to be rendered in ISL. The integration of emotion analysis differentiates INTERACT from purely linguistic translation approaches, preserving emotional context frequently lost in text-based accessibility solutions. The observation that 90% of participants deemed emotional expression visualisa- tions valuable corroborates this design choice. Relative to avatar signing systems developed for educational purposes, INTERACT prioritises real-time performance suitable for interactive com- munication. The sub-2-second end-to-end latency enables conversational use cases that offline or batch-processed systems cannot accommodate. The XR deployment context affords spatial presence benefits unattainable through conventional 2D video interfaces. 5.3 Limitations Several constraints bound current INTERACT capabilities and the generalisability of pilot find- ings: • Vocabulary Scope: The 750-sign ISL vocabulary, while encompassing essential professional communication, remains insufficient for comprehensive business discourse. Technical termi- nology, industry-specific jargon, and colloquial expressions frequently require fingerspelling fallback, diminishing fluency. • Facial Expression Integration: ISL, like other sign languages, employs facial expressions as grammatical markers and emotional indicators. Current avatar animations prioritise manual signing with limited facial integration, potentially compromising linguistic completeness. This was the most frequently cited improvement area in participant feedback, with six out of ten participants independently raising this concern. • Language Support: Production deployment supports English–French translation exclu- sively. While the architecture accommodates extension, validation data for additional language pairs remains pending. The unanimous participant demand (100%) for additional language options confirms this as a high-priority development need. • Sample Size: Pilot validation engaged 10 participants across two demonstrations involving different stakeholder groups (academia/industry and deaf educators/community). Although participant expertise provides qualitative depth, larger-scale quantitative validation would strengthen generalisability claims. • Audio Chunking: The 1-second chunk approach occasionally fragments words at segment boundaries. While overlap mitigates this issue, optimisation opportunities remain, particularly for languages with longer average word lengths. 5.4 Future Directions Near-term development priorities include vocabulary expansion through collaboration with deaf community organisations and linguistic researchers. Integration of facial expression animation for grammatical markers represents a technically complex yet accessibility-critical enhancement, as corroborated by participant feedback. Additional language pair support (German, Spanish) will address European market requirements and respond to the universal participant demand for multilingual expansion. Longer-term research directions encompass sequence-to-sequence neural sign generation sup- planting dictionary lookup, enabling more natural signing for novel input sequences. Adaptive 15 signing speed calibrated to user preferences and cognitive load indicators could enhance compre- hension for diverse user populations. Integration of avatar replay and bookmark functionality would support review and learning applications, features explicitly requested by Demo 2 partic- ipants. Performance optimisation will continue targeting latency reduction in the speech-to-sign pipeline, with particular attention to audio chunk processing parameters that balance respon- siveness against word fragmentation. Integration with additional conferencing platforms beyond Rainbow SDK will broaden interoperability and deployment flexibility. Commercial deployment pathways are being developed through the CORTEX2 ecosystem, with particular emphasis on educational institutions, cultural organisations (museums), and corporate training applications. The modular architecture enables phased adoption strategies suited to organisations with varying resources and accessibility requirements. 6 Conclusions This paper has presented INTERACT, a pioneering AI-driven XR platform that addresses ac- cessibility barriers in video conferencing for deaf, hard-of-hearing, and multilingual participants. Through the integration of Whisper-based speech recognition, NLLB multilingual translation, MediaPipe-driven ISL avatar animation, RoBERTa emotion analysis, and BART meeting sum- marisation, INTERACT delivers comprehensive communication support within immersive vir- tual environments. Pilot validation with deaf community stakeholders and accessibility professionals yielded 92% user satisfaction, a mean overall experience score of 4.6/5.0, and surpassed all technical per- formance targets, confirming both technical viability and meaningful accessibility value. The platform’s scalability to 1,000 concurrent users and robust performance under load testing af- firm readiness for broader deployment. Participant feedback has delineated clear pathways for enhancement, notably avatar facial expression integration, vocabulary expansion, and additional language support, which will steer the next development phase. A comprehensive account of the pilot results and implementation details is available in [Tantaroudas et al., 2026a], and the complementary AI service pipeline for inclusive language learning is detailed in [Tantaroudas et al., 2026c]. INTERACT represents a notable advancement in accessible communication technology, il- lustrating how integrated AI and XR capabilities can reshape participation opportunities for underserved communities. As remote collaboration continues expanding across professional, ed- ucational, and social contexts, platforms such as INTERACT will become increasingly vital for ensuring genuinely inclusive communication. Acknowledgements The authors gratefully acknowledge the CORTEX2 consortium for provision of the Mediation Gateway and Rainbow SDK infrastructure enabling INTERACT development. Special thanks to the KENG Institute for their invaluable partnership in pilot validation and ongoing consultation on sign language linguistic authenticity. We thank all pilot participants for their engagement and feedback. Funding This research was supported by FSTP Funding from the Horizon Europe research and innovation programme under grant agreement No. 101070192 (CORTEX2). Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union nor the granting authority. 16 Data and Software Availability Open-Source AI Models: • Speech Recognition (OpenAI Whisper): https://github.com/openai/whisper (MIT License) • Translation (Meta AI NLLB): https://github.com/facebookresearch/fairseq/tree/nllb (MIT License) • Pose Estimation (Google MediaPipe): https://github.com/google/mediapipe (Apache 2.0) • Emotion Analysis: https://huggingface.co/j-hartmann/emotion-english-distilroberta-base (Apache 2.0) • Summarisation: https://huggingface.co/philschmid/bart-large-cnn-samsum (MIT Li- cense) Pilot Study Data: Anonymised pilot study data are available at Zenodo [Tantaroudas et al., 2025a]: 10.5281/zenodo.18656422 (C-BY 4.0). ISL Gesture Dataset: The ISL gesture dataset is available at Zenodo [Tantaroudas et al., 2025b]: 10.5281/zenodo.18656296 (C-BY 4.0). Source Code: https://github.com/ntantaroudas/ISL-extractions-main (MIT License). Archived: 10.5281/zenodo.18694176. References Nikolaos D. Tantaroudas, Andrew J. McCracken, Ilias Karachalios, and Evangelos Papatheou. INTERACT: AI-powered extended reality platform for inclusive communication with real-time sign language translation and sentiment analysis. Open Research Europe, 6:71, 2026a. doi: 10.12688/openreseurope.23201.1. [version 1; peer review: awaiting peer review]. Alicia D. Alford, Joshua M. Bencak, Emily A. Tucker, et al. Is the window of learning only cracked open? Parents’ perspectives on virtual learning for deaf and hard of hearing students. American Annals of the Deaf, 168(3):17–28, 2023. doi: 10.1353/aad.2023.a917247. World Health Organization. World report on hearing. Technical report, WHO, Geneva, 2021. URL https://w.who.int/publications/i/item/world-report-on-hearing. Daniele Franceschini, Elizabeth Salesky, Jan Niehues, et al. Removing European language barri- ers with innovative machine translation technology. In Proceedings of the International Work- shop on Language Technologies (IWLTP), 2020. Paula A. Rodríguez-Correa, Alejandro Valencia-Arias, Odalis N. Patiño-Toro, et al. Benefits and development of assistive technologies for Deaf people’s communication: A systematic review. Frontiers in Education, 8:1174831, 2023. doi: 10.3389/feduc.2023.1121597. Yuchen Liu, Junnan Zhu, Jiajun Zhang, and Chengqing Zong. Bridging the modality gap for speech-to-text translation. arXiv preprint arXiv:2010.14920, 2020. doi: 10.48550/arXiv.2010. 14920. Teresa Hirzle, Florian Müller, Fiona Draxler, Martin Schmitz, Pascal Knierim, and Kasper Horn- bæk. When XR and AI meet—a scoping review on extended reality and artificial intelligence. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI ’23), pages 1–45. Association for Computing Machinery, 2023. doi: 10.1145/3544548.3581072. 17 Nikolaos D. Tantaroudas, Andrew J. McCracken, Ilias Karachalios, and Evangelos Papatheou. Enhancing accessibility and inclusivity in business meetings through AI-driven extended reality solutions. In Extended Reality. XR Salento 2025, volume 15743 of Lecture Notes in Computer Science. Springer, Cham, 2026b. doi: 10.1007/978-3-031-97781-7_6. Nikolaos D. Tantaroudas, Andrew J. McCracken, Ilias Karachalios, and Evangelos Papatheou. AI-based services for inclusive language learning in immersive XR environments: Speech translation, and sign language integration. Open Research Europe, 6:72, 2026c. doi: 10.12688/openreseurope.23214.1. [version 1; peer review: awaiting peer review]. Yasaman Hosseinkashi, Yun Liu, Jing Li, et al. Meeting effectiveness and inclusiveness: Large- scale measurement, identification of key features, and prediction in real-world remote meetings. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1):1–39, 2023. doi: 10. 1145/363737. Stefania Serafin, Ali Adjorlu, and Lone Marianne Percy-Smith. A review of virtual reality for individuals with hearing impairments. Multimodal Technologies and Interaction, 7(4):36, 2023. doi: 10.3390/mti7040036. Mohamed Sami Anwar, Bowen Shi, Vedanuj Goswami, et al. MuAViC: A multilingual audio- visual corpus for robust speech recognition and robust speech-to-text translation. arXiv preprint arXiv:2303.00628, 2023. doi: 10.48550/arXiv.2303.00628. Aoxiong Yin, Tianyun Zhong, Li-Hao Tang, et al. Gloss attention for gloss-free sign language translation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2551–2562. IEEE, 2023. doi: 10.1109/CVPR52729.2023.00251. Xiang Wu, Xiao Luo, Zhi Song, et al. Ultra-robust and sensitive flexible strain sensor for real- time and wearable sign language translation. Advanced Functional Materials, 33(4):2303504, 2023. doi: 10.1002/adfm.202303504. Necati Cihan Camgöz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023–10033. IEEE, 2020. doi: 10.1109/CVPR42600.2020.01004. Benjia Zhou, Zhigang Chen, Albert Clapés, et al. Gloss-free sign language translation: Improving from visual-language pretraining. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 20871–20881. IEEE, 2023. doi: 10.1109/ICCV51070.2023.01908. Yu Gu, Robert Tinn, Hao Cheng, et al. Domain-specific language model pretraining for biomed- ical natural language processing. ACM Transactions on Computing for Healthcare, 3(1):1–23, 2022. doi: 10.1145/3458754. Sylvie Gibet and Pierre-François Marteau. Signing avatars—multimodal challenges for text-to- sign generation. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG), pages 1–8. IEEE, 2023. doi: 10.1109/FG57933.2023.10042759. Julia Fink, Pierre Poitier, Marc André, et al. Sign language-to-text dictionary with lightweight transformer models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI-23), pages 5968–5976, 2023. doi: 10.24963/ijcai.2023/662. European Union of the Deaf. EUD position paper: International sign language, 2018. URL https://eud.eu/eud/position-papers/international-signs/. HandSpeak. International sign language dictionary, 2024. URL https://w.handspeak.com/ word/search/. 18 SpreadTheSign. International sign language database, 2024. URL https://w.spreadthesign. com/. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning (ICML), volume 202 of PMLR, pages 28492–28518, 2023. URL https://proceedings.mlr.press/v202/radford23a.html. Felix Stahlberg. Neural machine translation: A review. Journal of Artificial Intelligence Research, 69:343–418, 2020. doi: 10.1613/jair.1.12007. NLLB Team, Marta R. Costa-jussà, James Cross, et al. No language left behind: Scaling human- centered machine translation. arXiv preprint arXiv:2207.04672, 2022. doi: 10.48550/arXiv. 2207.04672. Loïc Barrault, Yu-An Chung, Christopher Cieri, et al. SeamlessM4T—massively multilingual & multimodal machine translation. arXiv preprint arXiv:2308.11596, 2023. doi: 10.48550/ arXiv.2308.11596. Rosalind W. Picard. Affective Computing. MIT Press, 1997. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692, 2019. doi: 10.48550/arXiv.1907.11692. Jochen Hartmann.Emotion English DistilRoBERTa base: A fine-tuned model for emotion classification, 2022.URL https://huggingface.co/j-hartmann/ emotion-english-distilroberta-base. Wenxuan Zhang, Yue Deng, Bing Liu, et al. Sentiment analysis in the era of large language models: A reality check. arXiv preprint arXiv:2305.15005, 2023. doi: 10.48550/arXiv.2305. 15005. Wei Chen, Jia Zhang, Yi Liu, et al. An event-based framework for facilitating real-time sentiment analysis in educational contexts. In 2022 11th International Conference on Educational and Information Technology (ICEIT), pages 57–61. IEEE, 2022. URL https://ieeexplore.ieee. org/document/9690729. Ross E. Mitchell and Michael A. Karchmer. Chasing the mythical ten percent: Parental hearing status of deaf and hard of hearing students in the United States. Sign Language Studies, 4(2): 138–163, 2004. doi: 10.1353/sls.2004.0005. Tom Brown, Benjamin Mann, Nick Ryder, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. doi: 10.48550/arXiv.2005.14165. Sumit Asthana, Shaked Hilleli, Pengcheng He, et al. Summaries, highlights, and action items: Design, implementation and evaluation of an LLM-powered meeting recap system. arXiv preprint arXiv:2307.15793, 2023. doi: 10.48550/arXiv.2307.15793. Yixin Liu, Pengfei Liu, Dragomir R. Radev, and Graham Neubig. BRIO: Bringing order to abstractive summarization. arXiv preprint arXiv:2203.16804, 2022. doi: 10.48550/arXiv.2203. 16804. Paul Roit, Johan Ferret, Lior Shani, et al. Factually consistent summarization via reinforcement learning with textual entailment feedback. arXiv preprint arXiv:2306.00186, 2023. doi: 10. 48550/arXiv.2306.00186. 19 CORTEX2 Project. CORTEX2 architecture and framework, 2024. URL https://cortex2.eu/ project/. Alcatel-Lucent Enterprise. Rainbow developer portal, 2024. URL https://developers. openrainbow.com/. Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. MediaPipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. doi: 10.48550/arXiv.1906.08172. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. doi: 10.48550/arXiv.1910.01108. Philipp Schmid. BART-large CNN SamSum, 2021. URL https://huggingface.co/ philschmid/bart-large-cnn-samsum. Nikolaos D. Tantaroudas, Andrew J. McCracken, Ilias Karachalios, and Evangelos Papatheou. INTERACT pilot study data: Anonymised questionnaire responses and performance metrics, 2025a. Nikolaos D. Tantaroudas, Andrew J. McCracken, Ilias Karachalios, and Evangelos Papatheou. INTERACT ISL gesture dataset—international sign language animation data, 2025b. 20