doorbell_detector

ROS 2 package for detecting doorbell-like sounds using YAMNet.

This node was added for the RoboCup HRI task, where the robot waits for a doorbell-like sound before continuing the interaction. It can detect sounds such as bell, chime, ding-dong, buzzer, knock, and ding.

The node supports two audio sources:

alsa  -> read directly from a local microphone using sounddevice
ros   -> subscribe to a ROS audio topic

For our current RoboCup setup, the tested and recommended mode is:

--source alsa --device pulse

This uses the microphone selected in the laptop’s PulseAudio settings.

Node

doorbell_node

The node listens for doorbell-like sounds and publishes a ROS event when the aggregated YAMNet confidence passes the configured threshold.

Topics

Subscribed

/doorbell/listen

Type:

std_msgs/msg/Bool

Usage:

True  -> start listening / open microphone
False -> stop listening / release microphone

Example:

ros2 topic pub --once /doorbell/listen std_msgs/msg/Bool "{data: true}"

Published

/doorbell/event

Type:

std_msgs/msg/Bool

Usage:

True -> doorbell-like sound detected

Monitor detections:

ros2 topic echo /doorbell/event

Installation

Create and activate a virtual environment:

python3 -m venv --system-site-packages ~/venvs/doorbell_detector
source ~/venvs/doorbell_detector/bin/activate

Install dependencies:

pip install -r requirements.txt

If sounddevice fails:

sudo apt install portaudio19-dev
pip install -r requirements.txt

Useful audio tools:

sudo apt install alsa-utils pulseaudio-utils pavucontrol

Build:

cd ~/ros2_ws
colcon build --packages-select doorbell_detector
source install/setup.bash

When running the node, remember to activate the venv first:

source ~/venvs/doorbell_detector/bin/activate
source ~/ros2_ws/install/setup.bash

Setup on a new laptop

First list the available audio devices:

ros2 run doorbell_detector doorbell_node --source alsa --list-devices

Example:

0 HD-Audio Generic: HDMI 0 (hw:0,3), ALSA (0 in, 8 out)
1 hdmi, ALSA (0 in, 8 out)
2 pulse, ALSA (32 in, 32 out)
* 3 default, ALSA (32 in, 32 out)

Use a device with input channels. Avoid devices with 0 in.

For our setup, use:

pulse

Before testing the detector, verify that the microphone records sound:

arecord -D pulse -f S16_LE -c 1 -r 16000 -d 5 -vv /tmp/mic_test.wav
aplay /tmp/mic_test.wav

If the recording is silent, open:

pavucontrol

Then go to Input Devices, select the correct microphone, and check that the input level moves when speaking.

Recommended test command

Use this to quickly verify that the detector works:

ros2 run doorbell_detector doorbell_node \
  --source alsa \
  --device pulse \
  --top-k 10 \
  --threshold 0.2 \
  --consecutive 1 \
  --inference-hz 0.25 \
  --window-sec 1.5

In another terminal:

ros2 topic echo /doorbell/event

Play a doorbell sound close to the microphone.

A successful detection should print something like:

DOORBELL DETECTED

and /doorbell/event should publish:

data: true

Recommended RoboCup command

After the basic test works, use a slightly safer configuration:

ros2 run doorbell_detector doorbell_node \
  --source alsa \
  --device pulse \
  --top-k 5 \
  --threshold 0.2 \
  --consecutive 2 \
  --inference-hz 0.5 \
  --window-sec 1.5

This requires two positive inference windows before publishing the event, reducing false positives.

Parameters

--source

Audio source.

Options:

ros
alsa

Recommended:

--source alsa

--audio

ROS audio topic used only with:

--source ros

Default:

/audio/in

Example:

--source ros --audio /audio/in

For the current RoboCup setup, this mode was not the main tested path.

--device

Audio input device used with:

--source alsa

Examples:

--device pulse
--device default
--device 2

Recommended:

--device pulse

--list-devices

Lists available audio devices and exits:

ros2 run doorbell_detector doorbell_node --source alsa --list-devices

--out

Detection output topic.

Default:

/doorbell/event

--listen-cmd

Command topic used to enable or disable listening.

Default:

/doorbell/listen

--threshold

Doorbell confidence threshold.

Lower values are more sensitive but may create false positives. Higher values are safer but may miss quiet doorbells.

Tested value:

--threshold 0.2

--consecutive

Number of positive inference windows required before detection.

For testing:

--consecutive 1

For competition:

--consecutive 2

--inference-hz

Maximum YAMNet inference rate.

Lower values use less CPU but react slower.

Tested values:

--inference-hz 0.25
--inference-hz 0.5

--window-sec

Seconds of audio used per inference.

Tested value:

--window-sec 1.5

--top-k

Number of top YAMNet classes printed in the logs.

Useful for debugging:

--top-k 10

For normal use:

--top-k 5

--gain

Software gain applied before inference.

Default:

--gain 1.0

Use this only if the microphone is too quiet.

--cooldown-sec

Minimum time between detections.

Default:

--cooldown-sec 5.0

--release-on-fire / --no-release-on-fire

By default, the node releases the microphone after detection.

This is useful for HRI because the speech pipeline may need the microphone after the doorbell event.

Recommended:

--release-on-fire

HRI usage

The HRI state machine should:

1. Publish True to /doorbell/listen.
2. Wait for True on /doorbell/event.
3. Continue the task.
4. Let the detector release the microphone after detection.
5. Start or resume speech recognition if needed.

Expected flow:

WAIT_FOR_DOORBELL
  -> publish /doorbell/listen=True
  -> wait for /doorbell/event=True

DOORBELL_DETECTED
  -> detector publishes /doorbell/event=True
  -> detector releases microphone

CONTINUE_HRI_TASK

To listen again later:

ros2 topic pub --once /doorbell/listen std_msgs/msg/Bool "{data: true}"

Troubleshooting

/doorbell/event does not publish

First check the microphone:

arecord -D pulse -f S16_LE -c 1 -r 16000 -d 5 -vv /tmp/mic_test.wav
aplay /tmp/mic_test.wav

If the recording is silent, fix the input device in:

pavucontrol

The node prints agg=0.00 and bell={}

The node is running, but YAMNet is not hearing doorbell-like audio.

Check that:

• the microphone records sound;
• the selected input in pavucontrol is correct;
• the doorbell sound is loud enough;
• the node is using --device pulse.

Device or resource busy

Another process is using the microphone.

Check:

sudo fuser -v /dev/snd/*

If PulseAudio is using the device, prefer:

--device pulse

instead of opening raw hardware devices.

sounddevice query failed

The selected device is not available.

Re-list devices:

ros2 run doorbell_detector doorbell_node --source alsa --list-devices

Then use pulse, default, or another valid input device.

TensorFlow GPU warnings

Warnings about missing GPU/CUDA libraries are usually not fatal. TensorFlow falls back to CPU, which is acceptable for this detector.

First run may need internet

The node loads YAMNet from TensorFlow Hub. Run the node once before the competition to make sure the model is cached on the laptop.