This research presents the design and implementation of an AI-driven wireless real-time voice translation system incorporating directional acoustic output, designed to facilitate seamless multilingual communication in dynamic environments. The proposed architecture integrates real-time speech recognition, Neural Machine Translation (NMT), and spatially controlled audio synthesis within a unified framework. Voice input is captured via a Frequency Modulated (FM) wireless microphone and transmitted to a Python-based desktop platform. The signal undergoes Automatic Speech Recognition (ASR) using a deep learning-based Speech-To-Text (STT) engine, followed by semantic translation via Google’s NMT-API, leveraging transformer-based models for high contextual fidelity. The translated linguistic output is rendered into naturalistic human like speech through a neural Text-To-Speech (TTS) engine and delivered via a parametric speaker array utilising ultrasonic transducers with "40 kHz" Pulse Width Modulation (PWM). This enables highly directional audio propagation with minimal ambient leakage, ensuring privacy and intelligibility for the intended listener without the need for wearable audio devices. A resource constrained ESP32 microcontroller orchestrates real-time data acquisition, translation synchronisation, and modulation control for the parametric output. Empirical evaluation demonstrates low end-to-end latency (1.5 - 2.5 seconds) and high ASR accuracy (90–95%), validating the system’s viability for deployment in multilingual conferences, educational domains, and public communication interfaces.
Jeyaram et al. (Sat,) studied this question.