Trabalho em anais de eventoOpen Access

A Modular Methodology for On-Device Audio Transcription from Live Media

SOUZA, T.; OLIVEIRA, S. B.; KIMURA , P. A. S.; ALVES, O.; Lima Filho, E. B.; CORDEIRO, L. C.; FABRICIO, R

2026 • 2026 IEEE 2nd International Conference on Consumer Technology

Resumo

Speech-to-text processing is a key feature for a richer experience in audiovisual content consumption, even in digital TV (DTV) environments, as it can provide information regarding content context. In that sense, cloud-based solutions are a possibility, but limitations related to latency, operational costs, and privacy concerns hinder their use on consumer electronics. To address this gap, this paper proposes a modular offline methodology for an on-device audio transcription pipeline, organized into three stages: Media Source Acquisition, Media Processing, and Transcription Engine. To evaluate it, we devised a practical implementation, experimented with it using different transcription engines, and computed metrics such as Word Error Rate, Real-Time Factor, and Random Access Memory (RAM) usage. Indeed, cloud APIs deliver the highest accuracy but remain impractical for embedded DTV use due to connectivity and privacy constraints. However, we were able to identify a viable on-device setup for resource-constrained hardware. Overall, the proposed methodology offers a feasible, lightweight, and privacypreserving foundation for enabling on-device intelligent services.

palavras-chave

Palavras-chave

speech-to-textdigital televisionembedded systemsresource-constrained devicesprivacy preservationword error ratereal-time factormemory usageaudiovisual consumption

Status

Published

links