This paper presents an offline speech-driven semantic input mod- ule born to support an indoor navigation application in the con- text of Sapienza University of Rome. The system addresses the difficulty international students face when searching for rooms, as Italian names are often mispronounced and incorrectly tran- scribed by standard speech-to-text (STT) systems. To operate in buildings with limited connectivity, the solution relies entirely on on-device processing and does not require cloud services or dedi- cated speech recognition models. Spoken queries are normalized through domain-specific vocabulary correction and spoken num- ber extraction before being matched against a local repository of 243 rooms across four campus buildings. The matching engine combines Levenshtein similarity with an Italian-specific phonetic representation to improve robustness against transcription errors and pronunciation variation. Evaluation on 150 spoken queries achieved 92% top-candidate accuracy, with an additional 6% cor- rectly resolved through disambiguation. The complete resolution pipeline executes in 8.0 ms on consumer mobile hardware, demon- strating the feasibility of lightweight, real-time, offline voice search to support indoor navigation applications
A Multilingual Speech-Driven Semantic Input Module for University Indoor Wayfinding / Ceylan, G., Trasciatti, G., Panizzi, E.. - (2026), pp. 106-110. (International Conference on Multimodal Interaction (was International Conference on Multimodal Interfaces) Napoli ) [10.1145/3776591.3832512].
A Multilingual Speech-Driven Semantic Input Module for University Indoor Wayfinding
Gabriella Trasciatti
Primo
;Emanuele PanizziUltimo
2026
Abstract
This paper presents an offline speech-driven semantic input mod- ule born to support an indoor navigation application in the con- text of Sapienza University of Rome. The system addresses the difficulty international students face when searching for rooms, as Italian names are often mispronounced and incorrectly tran- scribed by standard speech-to-text (STT) systems. To operate in buildings with limited connectivity, the solution relies entirely on on-device processing and does not require cloud services or dedi- cated speech recognition models. Spoken queries are normalized through domain-specific vocabulary correction and spoken num- ber extraction before being matched against a local repository of 243 rooms across four campus buildings. The matching engine combines Levenshtein similarity with an Italian-specific phonetic representation to improve robustness against transcription errors and pronunciation variation. Evaluation on 150 spoken queries achieved 92% top-candidate accuracy, with an additional 6% cor- rectly resolved through disambiguation. The complete resolution pipeline executes in 8.0 ms on consumer mobile hardware, demon- strating the feasibility of lightweight, real-time, offline voice search to support indoor navigation applicationsI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


