Voice-Command Control of a ROS Wheelchair

The recognizer node captures audio from the laptop microphone through linuxrec.c, samples the signal and compares it with the stored command templates. A match is published as a string on /voiceWords. The voice_cmd node (sub_word in the node graph) subscribes to that topic, maps the string to an integer command and publishes cmd_vel, which drives the wheel joints in Gazebo through the state publishers.

The vocabulary is trained for Forward, Backward, Left, Right and Stop, and a "Go" keyword activates the subscribing node. An input that does not match a stored command is rejected and the loop returns to listening, so only exact matches produce motion.

1. Microphone: laptop mic, audio capture in C.
2. Recognizer: samples the signal and matches stored commands.
3. /voiceWords: the recognized string, published to ROS.
4. voice_cmd: subscriber that turns the string into an integer command.
5. cmd_vel: the velocity command to the wheelchair in Gazebo.

• Wrote the audio-capture code (linuxrec.c, C) that takes the voice input from the laptop microphone.
• Built the recognizer node that matches the input against the stored command set and publishes the result on /voiceWords.
• Built the voice_cmd subscriber node that converts the recognized string to an integer command and publishes cmd_vel.
• Tested the commands in Gazebo with different participants and measured accuracy and response time.

• Tested with ten inputs and different participants; a command is detected independently of the speaker if it is pronounced in the trained language.
• Latency from input to wheelchair motion is about 5 s.
• Left and right fail most often, probably because of accent and pronunciation. The results come from simulation and may differ on a real wheelchair.

A more natural speech API (Google Voice) or NLP, and more training data.