# Voice-controlled autonomous navigation for smart wheelchairs using ROS-based SLAM

## TL;DR

This paper introduces a voice-controlled smart wheelchair that improves mobility for people with motor impairments by combining inclusive speech recognition and robust navigation.

## Contribution

A low-cost, voice-controlled smart wheelchair with inclusive speech recognition, real-time navigation, and safety-aware behavior is developed and experimentally validated.

## Key findings

- The system achieves a 6.7% Word Error Rate in speech recognition under realistic noise conditions.
- The wheelchair demonstrates a 94% goal-completion rate and a mean localization error below 10 cm in dynamic environments.
- End-to-end voice-to-motion latency is 0.8 seconds, ensuring responsive and safe operation.

## Abstract

Smart wheelchairs have the potential to significantly improve autonomy for individuals with severe motor impairments, yet existing systems often exhibit limited speech robustness, insufficient handling of dynamic environments, and a lack of rigorously validated safety mechanisms. This work presents a fully integrated, voice-controlled smart wheelchair that advances assistive mobility through three main contributions. First, we introduce an inclusive speech-recognition module built from a fine-tuned deep learning model trained on a custom dataset that incorporates recordings from users with mild speech impairments. This adaptation improves robustness to non-standard pronunciation and maintains reliable command execution under realistic noise conditions (70–75 dB), achieving a Word Error Rate of 6.7% in quiet environments. Second, rather than proposing a new SLAM method, we develop a system-level navigation framework that optimally integrates 2D LiDAR-based SLAM (GMapping), AMCL localization, and a dual-level voice-command interface within a real-time coordination layer. This includes a quantitatively parameterized safety module featuring adaptive speed modulation and experimentally calibrated emergency-stop thresholds, ensuring reliable operation in dynamic indoor environments. Third, we conduct an extensive experimental campaign in both simulation and real-world conditions to provide a reproducible and quantitative evaluation of system performance. Tests involving dynamic obstacles (pedestrians, wheeled carts, small animals), constrained passages, and diverse acoustic settings demonstrate a mean localization error below 10 cm, a 94% goal-completion rate, and an end-to-end voice-to-motion latency of 0.8 s. Together, these contributions provide a low-cost, experimentally validated assistive-mobility platform that emphasizes inclusive voice interaction, robust real-time navigation, and safety-aware behavior. The proposed framework moves beyond component-level studies by offering a coherent, deployable, and reproducible solution for everyday indoor environments.