Paper deep dive
Voice-Driven Semantic Perception for UAV-Assisted Emergency Networks
Nuno Saavedra, Pedro Ribeiro, André Coelho, Rui Campos
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 89%
Last extracted: 7/20/2026, 11:58:08 PM
Summary
The paper introduces SIREN, an AI-driven framework that converts unstructured emergency voice communications into structured, machine-readable semantic information for UAV-assisted networks. By integrating Automatic Speech Recognition (ASR), Large Language Models (LLM), and Natural Language Processing (NLP), SIREN extracts entities such as responding units, locations, emergency severity, and Quality-of-Service (QoS) requirements. The system is evaluated using synthetic emergency scenarios, demonstrating robust transcription and semantic extraction, though speaker diarization and geographic ambiguity remain limiting factors.
Entities (10)
Relation Signals (11)
SIREN → uses → Automatic Speech Recognition
confidence 95% · By integrating Automatic Speech Recognition (ASR) with Large Language Model (LLM)-based semantic extraction
SIREN → uses → Large Language Model
confidence 95% · integrating Automatic Speech Recognition (ASR) with Large Language Model (LLM)-based semantic extraction
SIREN → uses → Natural Language Processing
confidence 95% · integrating Automatic Speech Recognition (ASR) with Large Language Model (LLM)-based semantic extraction and Natural Language Processing (NLP) validation
SIREN → supports → UAV-assisted networks
confidence 90% · SIREN, an AI-driven framework that enables voice-driven perception for UAV-assisted networks.
SIREN → extracts → responding units
confidence 88% · SIREN converts emergency voice traffic into structured, machine-readable information, including responding units
SIREN → extracts → location references
confidence 88% · SIREN converts emergency voice traffic into structured, machine-readable information, including ... location references
SIREN → extracts → emergency severity
confidence 88% · SIREN converts emergency voice traffic into structured, machine-readable information, including ... emergency severity
SIREN → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Unmanned Aerial Vehicle (UAV)-assisted networks are increasingly foreseen as a promising approach for emergency response, providing rapid, flexible, and resilient communications in environments where terrestrial infrastructure is degraded or unavailable. In such scenarios, voice radio communications remain essential for first responders due to their robustness; however, their unstructured nature prevents direct integration with automated UAV-assisted network management. This paper proposes SIREN, an AI-driven framework that enables voice-driven perception for UAV-assisted networks. By integrating Automatic Speech Recognition (ASR) with Large Language Model (LLM)-based semantic extraction and Natural Language Processing (NLP) validation, SIREN converts emergency voice traffic into structured, machine-readable information, including responding units, location references, emergency severity, and Quality-of-Service (QoS) requirements. SIREN is evaluated using synthetic emergency scenarios with controlled variations in language, speaker count, background noise, and message complexity. The results demonstrate robust transcription and reliable semantic extraction across diverse operating conditions, while highlighting speaker diarization and geographic ambiguity as the main limiting factors. These findings establish the feasibility of voice-driven situational awareness for UAV-assisted networks and show a practical foundation for human-in-the-loop decision support and adaptive network management in emergency response operations.
Tags
Links
- Source: https://arxiv.org/abs/2602.17394v1
- Canonical: https://arxiv.org/abs/2602.17394v1
Trouble viewing inline? Open PDF directly →
Full Text
38,512 characters extracted from source content.
Expand or collapse full text
Voice-Driven Semantic Perception for UAV-Assisted Emergency Networks Nuno Saavedra, Pedro Ribeiro, Andr ́ e Coelho, Rui Campos INESC TEC and Faculdade de Engenharia, Universidade do Porto, Portugal nuno.m.carvalho, pedro.m.ribeiro, andre.f.coelho, rui.l.campos@inesctec.pt Abstract—Unmanned Aerial Vehicle (UAV)-assisted networks are increasingly foreseen as a promising approach for emergency response, providing rapid, flexible, and resilient communications in environments where terrestrial infrastructure is degraded or unavailable. In such scenarios, voice radio communications remain essential for first responders due to their robustness; however, their unstructured nature prevents direct integration with automated UAV-assisted network management. This paper proposes SIREN, an AI-driven framework that enables voice- driven perception for UAV-assisted networks. By integrating Automatic Speech Recognition (ASR) with Large Language Model (LLM)–based semantic extraction and Natural Language Processing (NLP) validation, SIREN converts emergency voice traffic into structured, machine-readable information, includ- ing responding units, location references, emergency severity, and Quality-of-Service (QoS) requirements. SIREN is evaluated using synthetic emergency scenarios with controlled variations in language, speaker count, background noise, and message complexity. The results demonstrate robust transcription and reliable semantic extraction across diverse operating conditions, while highlighting speaker diarization and geographic ambiguity as the main limiting factors. These findings establish the fea- sibility of voice-driven situational awareness for UAV-assisted networks and show a practical foundation for human-in-the- loop decision support and adaptive network management in emergency response operations. Index Terms—Artificial Intelligence, automatic speech recogni- tion, emergency communications, Large Language Models, UAV networks. I. INTRODUCTION Unmanned Aerial Vehicle (UAV)-assisted networks are in- creasingly foreseen as a promising approach for emergency response, driven by the need for resilient and adaptable com- munications in environments where terrestrial infrastructure is damaged or unavailable. In such scenarios, first responders continue to rely on voice radio transmissions due to their robustness under harsh conditions. However, while effective for basic coordination, voice communications are inherently unstructured, limiting their integration with automated systems and constraining coordinated decision-making and situational awareness. UAVs, acting as airborne communication relays, enable the rapid deployment of flying networks that restore connectivity in affected areas. Beyond supporting voice traffic, these net- works can deliver rich-media data such as images, video, and This work is financed by National Funds through the FCT – Fundac ̧ ̃ ao para a Ci ˆ encia e a Tecnologia, I.P. (Portuguese Foundation for Science and Technology) within the project FALCON, with reference 2023.15645.PEX (https://doi.org/10.54499/2023.15645.PEX). sensor measurements, thereby enhancing situational awareness and enabling more informed operational decisions. Their in- herent flexibility makes UAV-assisted networks particularly suitable for temporary and highly dynamic emergency oper- ations, where communication demands and network topology must continuously adapt. When integrated into UAV-assisted network management, semantic perception extracted from field communications can inform adaptive decisions such as UAV positioning, trajectory adjustment, and communication resource allocation, including the prioritization of links, bandwidth, and mission-critical traf- fic. In emergency operations, many of these operational needs are conveyed implicitly or explicitly through voice exchanges among first responders. Consequently, voice communications constitute a rich yet largely underutilized source of semantic perception for UAV-assisted network management. Recent advances in Automatic Speech Recognition (ASR) and Large Language Models (LLMs) [1], [2], have enabled the transformation of unstructured voice communications into actionable information. By analyzing voice communications, these technologies can extract semantic entities such as lo- cation references, emergency severity levels, and Quality-of- Service (QoS) requirements. This semantic inference comple- ments traditional telemetry by providing contextual awareness in scenarios where other sensing information, such as imagery, is limited, degraded, or unavailable. Although NLP-enhanced ASR has already demonstrated its value for situational aware- ness in domains such as air traffic control [3], the exploitation of voice-derived semantic perception as a direct input for adaptive UAV-assisted network management remains largely unexplored, specifically for UAV positioning and resource allocation in UAV-assisted emergency response scenarios. This paper proposes SIREN, an AI-driven framework de- signed to function as a voice-driven semantic perception layer by extracting structured information from emergency voice communications. To the best of our knowledge, SIREN is the first approach to jointly integrate ASR, LLMs, and NLP techniques, including Named Entity Recognition (NER) [4] for location validation, speaker diarization for unit attribution, and sentiment analysis for emergency severity analysis, in order to interpret emergency audio streams. It is worth noting that SIREN does not perform network management or UAV deployment decisions itself; instead, it produces structured se- mantic outputs to support human-in-the-loop decision-making and existing network planning and control algorithms. This arXiv:2602.17394v1 [cs.NI] 19 Feb 2026 Fig. 1.Conceptual illustration of a UAV-assisted emergency networking scenario in which SIREN extracts semantic information from voice communi- cations to support situational awareness and network management for ground response units. capability enables its direct use as a perception input for adaptive UAV-assisted network management, as illustrated in Fig. 1. The main contributions of this paper are two-fold: 1) SIREN, a modular AI-based framework that transforms unstructured emergency voice communications into struc- tured, machine-readable perception inputs suitable for UAV-assisted network management tasks, such as UAV positioning and communication resource allocation. 2) A comprehensive performance evaluation of SIREN across five synthetic emergency communication scenarios with varying language, number of speakers, message complexity, and background noise, demonstrating the feasibility of the proposed solution while identifying key limitations, including speaker diarization errors and geographic ambiguity. I. RELATED WORK UAV-assisted networks have been widely explored as a solution for resilient and rapidly deployable communications in disaster response and search-and-rescue scenarios, where terrestrial infrastructure may be damaged or unavailable [5], [6]. In such scenarios, UAVs can act as airborne relays and access points, where performance is strongly affected by UAV mobility and traffic demand, requiring continuous adaptation of UAV positioning and communications resources [7]. Many UAV placement solutions still rely on assumptions about user locations, traffic demand, and channel conditions [5], [7], which may not hold in dynamic emergency environments. Perception is a core capability in emerging UAV-assisted networks, typically driven by computer vision for detection and tracking [8]. While computer vision provides rich spa- tial information, it degrades under occlusions, illumination changes, and adverse conditions common in disasters. Audio- based perception is a promising complement to vision by reducing reliance on line-of-sight and lighting [9]. In parallel, LLMs are increasingly used as cognitive layers that operate on structured semantic representations to bridge perception and decision-making, enabling contextual interpre- tation and intent inference in UAV systems [1]. Representative frameworks show LLM-based spatial reasoning and multi- modal interpretation when operating on structured outputs from perception modules [2]. Despite these advances, most LLM-enabled UAV perception frameworks remain vision- centric and focus on sensor-derived data. The use of unstructured emergency voice communications as a semantic perception source, capturing operational intent, urgency, and contextual requirements from first responders, remains largely unexplored. Although agentic architectures for UAV-assisted networks have considered separating percep- tion from decision-making via structured semantic represen- tations [10], such frameworks remain largely conceptual. The integration of voice-based operational information is proposed only as a potential future direction, lacking concrete imple- mentations. I. SIREN FRAMEWORK This section introduces SIREN, an AI-driven framework that converts unstructured emergency voice communications into structured, machine-readable semantic representations. We describe the end-to-end pipeline from raw audio to structured outputs, designed for integration with UAV-assisted network management functions such as UAV positioning and communications resource allocation. A. System design The SIREN framework is designed as a modular, multi-stage pipeline (cf. Fig. 2) that processes emergency communications into structured outputs suitable for integration with UAV- assisted network decision-making. Each stage of the pipeline is described in the following. 1) Automatic Speech Recognition: The ASR stage performs speech-to-text conversion and is designed to accommodate different approaches, depending on the operational context. The pipeline supports the integration of lightweight local models for offline processing in connectivity-constrained en- vironments, as well as cloud-based APIs that allow for high- accuracy transcription, especially under noisy conditions, and enable advanced features such as speaker diarization and sentiment analysis. 2) Information extraction: The information extraction stage processes the transcribed text through a modular LLM compo- nent, which can be deployed either locally or via cloud-based services. To ensure reliability, the system employs schema- constrained prompting to define specific output fields, such as location, units, emergency level, and QoS expectations. The LLM output is validated against this schema to mitigate hallucinations and to ensure interoperability with UAV control and network management systems (cf. Listing 1). To further enhance reliability, the extracted information is cross-validated using complementary NLP techniques. NER is applied to verify geographic entities and, when a cloud-based API is available, speaker diarization and sentiment analysis may be employed to improve unit attribution and to refine emergency severity estimation, respectively. AudiosASR Information Extraction Structured Output LLM NLP Validation Transcribed Text Transcribed Text LLM Output Sentiment Analysis | Speaker Diarization Fig. 2. Overview of the SIREN pipeline, comprising ASR, Information Extraction, integrating probabilistic LLM inference with deterministic NLP validation, and generation of structured outputs. 3) Output information: The SIREN framework produces a structured representation (e.g., a JSON object) designed for integration with state-of-the-art solutions for network management, such as UAV positioning, network planning, and communications resource allocation. Listing 1 provides a reference example of this representation and its main semantic categories. This structured output may also serve as support for human-in-the-loop decision-making. "locations": [ "Stow Lake", "Gold Star Mother’s Rock" ], "emergency_level": "Critical", "units": [ "name": "Unit Alpha", "location": "Gold Star Mother’s Rock", "video_support": "needed": true, "issue": "Poor signal; need uplink for fire assessment.", "priority": "high", "requirements": "10 Mbit/s" , "image_support": "needed": true, "issue": "Thermal upload failed; coverage minimal." , "priority": "medium", "requirements": "5 Mbit/s" , "times_intervened": 6 ] Listing 1. Example of structured semantic output generated by SIREN from emergency voice communications. The structured output includes three main semantic cate- gories. First, the locations field aggregates geographic ref- erences mentioned in the communication (e.g., street names or landmarks). Second, the emergency level field indicates the severity of the situation (e.g., critical, minor). These emergency level categories are predefined in the prompt, and the model selects one of them based only on the content of the communication. Third, the units structure describes the active responders and their associated attributes. It is worth mentioning that only information explicitly stated in the communication is included; no additional inference or geocoding-based enrichment is performed. For each identified unit, SIREN associates the reported location reference with the extracted locations and populates Fig. 3.Interactive map interface depicting a geo-referenced unit and its specific communication requirements extracted by SIREN. the timesintervened field, which reflects the number of inter- actions in the communication and indicates operational rele- vance. The videosupport and imagesupport fields capture QoS expectations, including the requested support type, a brief justification, a priority level, and an indicative throughput requirement (in Mbit/s). These throughput values are intended as expectations rather than precise measurements. Requiring explicit justifications further acts as a grounding mechanism, improving transparency, reducing ambiguity in support re- quests, and mitigating the risk of hallucinated requirements. B. System implementation In its current implementation, SIREN’s pipeline relies on a local instance of OpenAI Whisper [11] for offline ASR or on the Asssembly API [12], a cloud-based ASR API. Information extraction is handled by the LLaMA 3.2 model [13] deployed locally via the Ollama framework [14], which provides the inference engine to identify location references, unit identi- fiers, emergency levels, and QoS expectations. To enforce the design’s requirement for structured data, a strict JSON schema is applied at the prompt level. Given the probabilistic nature of the LLM outputs, they are audited by a validation layer comprising three deterministic NLP-based modules: 1) Named Entity Recognition (NER): A NER model implemented using the SpaCy library [15] is used to cross-check geographic entities extracted by the LLM, reducing the probability of spurious or incorrect location references. 2) Speaker diarization: When available, speaker diarization results provided by the transcription Asssembly API are cross-referenced with the units extracted by the LLM. Discrepancies, such as a mismatch between the number of detected speakers and extracted operational units, are flagged as potential issues or hallucinations, and trigger the validation layer to rectify the LLM output. 3) Sentiment analysis: To calibrate the extracted emergency severity level and improve robustness against misclassi- fication, SIREN uses the mean sentiment score across transcript segments, provided by the Asssembly API. If the LLM classifies an event as moderate while senti- ment indicates a strong negative affect, such as panic or distress, the emergency level is escalated to severe or critical. Conversely, neutral or positive sentiment may trigger a downward adjustment. Finally, validated location entities are mapped to geographic coordinates using Geopy [16] and rendered through an inter- active web-based interface using Folium [17]. As depicted in Fig. 3, this interactive map displays the identified units and their specific communications requirements, providing opera- tors with quick situational awareness. While the structured out- put is designed to support autonomous network management, this visualization ensures that human operators can monitor and validate the semantic context of voice communications. IV. PERFORMANCE EVALUATION AND RESULTS This section presents the performance evaluation of SIREN in terms of transcription accuracy, output quality, and execu- tion time. The goal is to assess the feasibility and behavior of SIREN under diverse operating conditions. A. Audio dataset Authentic emergency communication datasets are scarce due to privacy, legal, and sensitivity constraints. A reference state-of-the-art dataset, RescueSpeech [18], contains annotated German search-and-rescue (SAR) recordings; however, it lacks continuous, multi-speaker conversational structure, making it unsuitable for the objectives of this work. To address these limitations, a synthetic audio dataset 1 , was generated follow- ing a two-stage process. First, coherent emergency dialogues were generated using LLMs, ensuring contextual continuity across speakers and message exchanges. These dialogues were then converted into speech using ElevenLabs [19], a text-to-speech (TTS) system. Multiple distinct voices were employed to emulate different first responders for increased realism. Additionally, noisy variants of the original audio were generated using Python post-processing techniques to assess robustness under degraded acoustic conditions. The dataset is organized into five scenarios grouped into three levels of complexity, designed to assess the SIREN framework across linguistic and operational dimensions: • Low-complexity scenario – Scenario 1: It is the simplest scenario. It consists of short, well-structured emergency dialogues in English, involving four distinct speakers and three location 1 https://github.com/Slyfenon/SIREN-audio files references. Both clean and noisy audio versions are included. Duration: 1 min 10 s. • Medium-complexity scenarios – Scenario 2: This scenario increases difficulty by intro- ducing longer dialogues and richer operational context. It maintains four speakers but includes a greater num- ber of location references. Duration: 1 min 28 s. – Scenario 3: Closely mirrors the structure of Scenario 2 to validate consistency. It features similar linguistic patterns, speaker and location counts as Scenario 2. Duration: 1 min 30 s. • High-complexity scenarios – Scenario 4: This scenario introduces audio in Por- tuguese language, referencing geographical locations in Portugal. It contains four speakers, language-specific characteristics, such as pronunciation differences, syn- tactic variation, and ambiguous place names. These introduce further challenges, particularly for models optimized for English. Duration: 1 min 54 s. – Scenario 5: The most complex scenario. It includes six distinct speakers, an increased number of locations, and dialogue that is less explicit regarding speaker identi- fication and QoS expectations. Duration: 1 min 41 s. B. Hardware specification All experiments were conducted on a single machine with the following specifications: Ubuntu 22.04 operating system, an AMD Ryzen 5 5600H processor (12 threads at 3.3 GHz), and an NVIDIA GeForce RTX 3050 laptop GPU with 4 GB of VRAM. To ensure measurement accuracy and consistency, no significant background processes were active during testing. C. Transcription quality assessment Transcription quality was evaluated using the Word Error Rate (WER), a standard performance metric for ASR systems, as defined in Eq. 1, where S, D, and I represent the number of substitutions, deletions, and insertions, respectively, and N is the number of words in the reference transcript, as defined in [20]. WER = S + D + I N ,(1) Tables I and I compare the performance of the Assembly API-based and local Whisper models under clean and noisy audio conditions. Results are reported as WER (%) with the corresponding substitution, deletion, and insertion counts shown in parentheses as (S/D/I). In clean audio conditions (Table I), both models achieve relatively low WER values; however, the API-based model consistently outperforms the local model, primarily due to fewer substitution errors. Under noisy conditions (Table I), WER increases for both models; however, the degradation is substantially greater for the local Whisper model. This effect is most evident in Scenario 2, where the local Whisper WER rises from 15.31% to 39.80%, while the API-based model shows a more moderate increase from 12.76% to 17.86% (mean over 10 iterations). TABLE I WER AND ERROR BREAKDOWN ON CLEAN AUDIO ScenarioAssembly APILocal Whisper 17.84 (6/0/6)12.42 (14/0/5) 212.76 (15/0/10)15.31 (19/1/10) 310.64 (13/2/10)10.64 (9/8/8) 46.25 (9/0/6)10.42 (15/0/10) 518.67 (23/3/16)17.78 (22/0/18) Total11.34 (66/5/48)13.25 (79/9/51) TABLE I WER AND ERROR BREAKDOWN ON NOISY AUDIO ScenarioAssembly APILocal Whisper 18.50 (7/0/6)14.38 (15/2/5) 217.86 (22/3/10)39.80 (52/10/16) 310.64 (13/2/10)13.19 (20/2/9) 47.50 (12/0/6)12.08 (19/0/10) 516.89 (22/1/15)18.67 (24/1/17) Total12.30 (76/6/47)19.26 (130/15/57) The higher WER observed for the local model is primarily driven by an increase in substitution and deletion errors, indicating reduced robustness in noisy environments. The observed differences can be justified by model size. The local Whisper implementation uses a smaller model variant due to hardware constraints, limiting its ability to generalize under noisy conditions. These results indicate that transcription quality is primarily influenced by model capacity and training data. As such, the external API-based model is better suited for deployment in settings with variable or degraded audio quality, as it offers higher transcription accuracy and robustness. D. Output quality assessment Output quality was evaluated across the same five scenarios that comprise the created dataset, as described in Section IV-A. Each scenario included five independent runs and featured both clean and noisy audio conditions. All runs were pro- cessed using the external transcription API, as the transcription quality assessment indicated higher transcription accuracy and stability compared to the local Whisper model. 1) Evaluation metrics: We assessed output quality along four dimensions: i) location identification; i) unit extraction; i) speaker count estimation; and iv) QoS expectation extrac- tion. • Location identification: This measures the ability to extract geographic references explicitly mentioned in the audio, such as street names or landmarks, which are central to situational awareness. Performance relies on the accuracy of the transcription. Since geocoding is performed by an external service, any geocoding errors are not attributed to SIREN. • Unit extraction: This measures the ability to identify operational units and associate them with their corre- sponding location references. Accurate unit identification is crucial for emergency coordination, as mislabeling a unit or incorrectly attributing its location can lead to suboptimal resource allocation. • Speaker count estimation: This measures the diarization component’s ability to estimate the number of distinct speakers present in the audio, even in cases with similar vocal characteristics. Errors in speaker differentiation can reduce the interpretability of the extracted semantics. • QoS expectation extraction: This assesses the ability to infer operational intent related to communications needs, such as multimedia support or additional resources, based on semantic understanding. All dimensions were assessed through manual inspection of the synthetic ground-truth dialogues. Tables I and IV report success rates across the five scenarios under clean and noisy conditions, where each value represents the proportion of successful runs (out of five) for a given scenario and metric. 2) Results: Results are reported on a per-scenario basis to analyze how increasing linguistic and operational complexity affects the performance of the SIREN framework. a) Scenario 1: In the baseline scenario, SIREN achieved 100% success in location identification, unit extraction, and speaker count estimation under both clean and noisy condi- tions (Table IV). The primary limitation lay in QoS expectation extraction, which achieved only 40% success. This indicates that SIREN misinterpreted the relationships between units and requested actions. Therefore, QoS extraction was more sensitive to semantic ambiguity than entity extraction, even in dialogues of low complexity. b) Scenario 2: SIREN achieved high performance in location identification and unit extraction. However, speaker count estimation dropped to 0%. Although the audio contained four speakers, two TTS voices exhibited highly similar timbre and prosodic characteristics, causing the diarization model to treat them as a single speaker across all runs. This behavior was consistent in both clean and noisy conditions, indicat- ing that voice similarity, rather than noise or transcription quality, was the dominant factor. Meanwhile, QoS expectation extraction remained at 100%, as this task depends on textual semantic cues rather than precise speaker attribution. c) Scenario 3: It resulted in stable performance, achiev- ing 100% success in both unit extraction and QoS expectation extraction. However, speaker count estimation recorded 0% success due to recurring similarities among synthetic voices, as noted in Scenario 2. A minor mismatch was observed in the geocoding stage, where one extracted location was resolved to an incorrect coordinate by the external service; nevertheless, the textual entity was correctly extracted by SIREN. Aside from this external dependency, the results remained consistent under both clean and noisy conditions. d) Scenario 4: SIREN achieved 100% success across all evaluated dimensions under both clean and noisy conditions. While occasional incorrect coordinate resolutions occurred due to the external geocoding service when place names had mul- tiple plausible matches (an aspect common in Portugal), this does not indicate an extraction failure. Under noisy conditions, SIREN also demonstrates successful semantic extraction. e) Scenario 5: The highest-complexity scenario revealed limitations primarily related to external ambiguity and diariza- Fig. 4. Example of geocoding ambiguity in complex scenarios. TABLE I OUTPUT QUALITY ACROSS SCENARIOS (CLEAN AUDIO) ScenarioLocationUnitSpeaker CountQoS 1100%100%100%40% 2100%100%0%100% 3100%100%0%100% 4100%100%100%100% 50%60%0%100% tion. For this scenario, no run achieved complete location identification, with the number of correctly identified loca- tions varying across runs. Similar to previous scenarios, the geocoding service often resolved them to incorrect coordinates when name ambiguity existed. Additionally, speaker count es- timation degraded, as several of the six TTS-generated voices had similar acoustic characteristics. In contrast, unit extraction remained robust (three out of five), and QoS expectation extraction was fully successful (five out of five). These results highlighted speaker diarization and geocoding ambiguity (cf. Fig. 4) as the primary bottlenecks in high-complexity audio conditions. E. Execution time assessment SIREN execution time was evaluated across scenarios by breaking down the total processing time into two primary components: transcription and LLM analysis. The transcription stage comprises ASR and speaker diarization. Because this stage relies on an external API, its latency fluctuates due to network conditions and server-side variations. LLM analysis involves the semantic processing of the transcribed text, in- cluding information extraction and structural mapping, with its runtime affected by the semantic complexity of the content and the hardware used. Each scenario was executed ten times to ensure representative statistical average measurements. Results in Table V follow the expected trend with respect to scenario complexity. Scenario 1, which represents the lowest- complexity case, leads to the shortest execution time, as both TABLE IV OUTPUT QUALITY ACROSS SCENARIOS (NOISY AUDIO) ScenarioLocationUnitSpeaker CountQoS 1100%100%100%40% 280%80%0%100% 3100%100%0%100% 4100%100%100%100% 50%60%0%100% TABLE V AVERAGE EXECUTION TIME PER SCENARIO AND SIREN’S PIPELINE STAGE. ScenarioTranscription (s)LLM (s)Total (s) 110.5914.2024.87 229.4822.1851.67 323.1823.4846.67 413.9527.4541.40 515.5548.9264.47 Mean18.5527.2545.82 transcription and LLM analysis complete quickly due to the limited audio duration and semantic content. Scenarios 2 and 3, characterized by increased linguistic and interaction complexity, yielded higher execution times, with more bal- anced contributions from transcription and LLM analysis. Scenario 4 (Portuguese audio) presents a moderate total ex- ecution time: while transcription remains relatively fast due to shorter segments, LLM analysis becomes the dominant component, reflecting the additional processing required for multilingual semantic interpretation. As expected, Scenario 5 is the most computationally demanding. The combination of multiple speakers and higher semantic complexity results in the longest overall execution time, with LLM analysis accounting for most of the runtime. This behavior reflects both the increased reasoning load and the limitations of the available local hardware. The per-scenario assessment shows that execution time in- creases with both audio and semantic complexity. Additionally, the results indicate that LLM processing is the dominant factor as scenario complexity rises. Thus, while transcription latency may vary due to external factors, the overall execution time of the SIREN’s pipeline is primarily influenced by the computational demands of LLM processing. V. DISCUSSION The evaluation results indicate that using LLMs to process voice communications and extract perception-level informa- tion relevant to UAV-assisted network management is feasi- ble, while also highlighting limitations that affect end-to-end reliability. A primary design trade-off concerns the choice of ASR backend. The comparison between the local Whisper model and the Assembly API-based transcription shows that, while the local configuration enables offline operation, its error rate increases significantly under noise (from approximately 15% to nearly 40%), whereas the API-based model presents a more moderate degradation. These results indicate that offline deployment may require more capable hardware, to support larger and more robust ASR models, or alternative noise- robust front-ends, to maintain comparable transcription quality in degraded acoustic conditions. The results also highlight a clear separation between se- mantic extraction and geographic resolution. SIREN is able to extract location references and QoS expectations from the transcript; however, mapping extracted entities to coordinates introduces errors when place names are ambiguous. In par- ticular, the external geocoding service may resolve common street names to incorrect places. This indicates that the main limitation lies in geographic disambiguation rather than in SIREN’s entity extraction. Such findings motivate context- aware disambiguation strategies (e.g., bounding regions, op- erational maps, or gazetteers) when converting extracted to- ponyms into actionable coordinates. Moreover, SIREN shows limitations in speaker differen- tiation in scenarios with acoustically similar voices. In the conducted experiments, this effect is exacerbated by synthetic text-to-speech voices with similar timbre and prosody, which can lead the diarization model to cluster distinct speakers. While human speakers may exhibit greater variability, real- world radio communications often involve compression and channel artifacts that attenuate speaker-specific cues. Conse- quently, diarization may be a potential source of attribution errors, especially in multi-speaker settings or under channel degradation. Geographic ambiguity can be reduced when communica- tions reference distinctive landmarks (e.g., Fort Point National Historic Site) instead of generic street names that allow for multiple interpretations. Similarly, the extraction of QoS expectations is more reliable when requirements are expressed explicitly (e.g., video support, high bandwidth) rather than indirectly. Finally, diarization errors are mitigated in communi- cations with clearer turn-taking and reduced speaker overlap, which improves unit attribution (e.g., Unit Alpha and Unit Charlie). Although SIREN was developed for emergency communi- cations, its perception pipeline is domain-agnostic and can be applied to scenarios that rely on voice-based coordination and situational awareness. Its structured outputs can support human-in-the-loop decision-making and provide perception input for autonomous or semi-autonomous systems, such as in surveillance, monitoring, or search-and-rescue operations where other sensing or communication channels are limited. VI. CONCLUSIONS This paper proposed SIREN and demonstrated it can convert unstructured emergency voice communications into structured, machine-readable information, achieving high transcription accuracy and reliably extracting locations and responding units in both clean and noisy conditions. The results confirm the feasibility of voice-driven semantic extraction as a decision- support mechanism for UAV-assisted emergency networks. The performance evaluation demonstrated that SIREN offers near-real-time situational awareness to network operators and serves as a viable proof of concept for decision support. How- ever, the framework has limitations such as speaker diarization failures in scenarios with similar voices, ambiguous geocoding due to place-name resolution, and scalability constraints as audio duration and semantic complexity increase. Future work will focus on enhancing geographic disam- biguation by utilizing constrained search spaces and domain- specific location dictionaries. Additionally, we aim to validate the framework with authentic emergency recordings to better capture real-world radio phenomena, such as interference, signal dropouts, and ambient noise. REFERENCES [1] Y. Tian, F. Lin, Y. Li, T. Zhang, Q. Zhang, X. Fu, J. Huang, X. Dai, Y. Wang, C. Tian, B. Li, Y. Lv, L. Kov ́ acs, and F.-Y. Wang, “Uavs meet llms: Overviews and perspectives towards agentic low-altitude mobility,” Information Fusion, vol. 122, p. 103158, 2025. [Online]. Available: https://w.sciencedirect.com/science/article/pii/S1566253525002313 [2] F. Lin, Y. Tian, Y. Wang, T. Zhang, X. Zhang, and F.-Y. Wang, “AirVista: Empowering UAVs with 3D Spatial Reasoning Abilities Through a Multimodal Large Language Model Agent,” in 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), 2024, p. 476–481. [3] A. Sarhan, R. Fathy, and H. Ali, “Intelligent air traffic control using NLP-enhanced speech recognition and natural language generation,” Journal of Electrical Systems and Information Technology, vol. 12, 07 2025. [4] S. Naseer, M. Ghafoor, S. Alvi, A. Kiran, G. Shafique Ur Rahmand, and G. Murtazae, “Named entity recognition (NER) in NLP techniques, tools accuracy and performance,” Pakistan Journal of Multidisciplinary Research, vol. 2, no. 2, p. 293–308, Jan. 2022. [5] M. J. Sobouti, A. Mohajerzadeh, H. Y. Adarbah, Z. Rahimi, and H. Ahmadi, “Utilizing UAVs in Wireless Networks: Advantages, Challenges, Objectives, and Solution Methods,” Vehicles, vol. 6, no. 4, p. 1769–1800, 2024. [Online]. Available: https://w.mdpi.com/2624- 8921/6/4/86 [6] S. A. Owaid, A. H. Miry, and T. M. Salman, “A Survey on UAV-Assisted Wireless Communications: Challenges, Technologies, and Application,” in 2024 11th International Conference on Electrical and Electronics Engineering (ICEEE), 2024, p. 333–340. [7] P. Ribeiro, A. Coelho, and R. Campos, “SUPPLY: Sustainable Multi- UAV Performance-Aware Placement Algorithm for Flying Networks,” IEEE Access, vol. 12, p. 159 445–159 461, 2024. [8] J. Xiao, R. Zhang, Y. Zhang, and M. Feroskhan, “Vision-Based Learning for Drones: A Survey,” IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 9, p. 15 601–15 621, 2025. [9] I. Alla, H. B. Olou, V. Loscri, and M. Levorato, “From Sound to Sight: Audio-Visual Fusion and Deep Learning for Drone Detection,” in Proceedings of the 17th ACM Conference on Security and Privacy in Wireless and Mobile Networks, ser. WiSec ’24, 2024, p. 123–133. [Online]. Available: https://doi.org/10.1145/3643833.3656133 [10] A. Coelho, P. Ribeiro, H. Fontes, and R. Campos, “A4FN: an Agentic AI Architecture for Autonomous Flying Networks,” in 2025 IEEE 36th International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), 2025, p. 1–6. [11] OpenAI,“Whisper:OpenAI’sSpeechRecognitionModel,” https://github.com/openai/whisper, 2022, Accessed: Jan. 16, 2026. [12] AssemblyAI,“AssemblyAIDocumentation,” https://w.assemblyai.com/docs, 2025, Accessed: Jan. 16, 2026. [13] H. Touvron, T. Lavril, G. Izacard, X. Martinet et al., “LLaMA 3: Open Foundation and Instruction Models,” 2024, meta AI. [Online]. Available: https://ai.meta.com/llama [14] Ollama, “Ollama: Get up and running with large language models.” https://ollama.com, 2024, Accessed: Jan. 16, 2026. [15] M. Honnibal, I. Montani, S. Van Landeghem, and A. Boyd, “spaCy: Industrial-strength Natural Language Processing in Python,” 2020. [Online]. Available: https://doi.org/10.5281/zenodo.1212303 [16] GeoPy, “GeoPy: Geocoding library for Python,” Accessed: 2026-1-16. [Online]. Available: https://geopy.readthedocs.io/en/stable/ [17] Folium, “Folium,” Accessed: Jan. 16, 2026. [Online]. Available: https://python-visualization.github.io/folium/latest/ [18] S.Sagar,M.Ravanelli,B.Kiefer,I.K.Korbayova,and J. van Genabith, “RescueSpeech: A German Corpus for Speech Recognition in Search and Rescue Domain,” 2023. [Online]. Available: https://arxiv.org/abs/2306.04054 [19] ElevenLabs,“TexttoSpeech—ElevenLabsDocumenta- tion,”2025,Accessed:Jan.16,2026.[Online].Available: https://elevenlabs.io/docs/capabilities/text-to-speech [20] K. Zechner and A. Waibel, “Minimizing Word Error Rate in Tex- tual Summaries of Spoken Language,” Language Technologies Insti- tute, Carnegie Mellon University, 2000, https://aclanthology.org/A00- 2025.pdf.