Hermes Audio Server

Important information: I consider Hermes Audio Server deprecated now that Rhasspy has been modularized. I recommend to use rhasspy-speakers-cli-hermes and rhasspy-microphone-cli-hermes instead.

Hermes Audio server implements the audio server part of the Hermes protocol defined by Snips.

It's meant to be used with Rhasspy, an offline, multilingual voice assistant toolkit that works with Home Assistant and is completely open source.

With Hermes Audio Server, you can use the microphone and speaker of your computer (such as a Raspberry Pi) as remote audio input and output for a Rhasspy system.

System requirements

Hermes Audio Server requires Python 3. It has been tested on a Raspberry Pi running Raspbian 9.8 and an x86_64 laptop with Ubuntu 19.04, but in principle it should be cross-platform. Please open an issue on GitHub when you encounter problems or when the software exits with the message that your platform is not supported.

Installation

You can install Hermes Audio Server and its dependencies like this:

sudo apt install portaudio19-dev
sudo pip3 install hermes-audio-server

Note: this installs Hermes Audio Server globally. If you want to install Hermes Audio Server in a Python virtual environment, drop the sudo.

Configuration

Hermes Audio Server is configured in the JSON file /etc/hermes-audio-server.json, which has the following format:

{
    "site": "default",
    "mqtt": {
        "host": "localhost",
        "port": 1883,
        "authentication": {
            "username": "foobar",
            "password": "secretpassword"
        },
        "tls": {
            "ca_certificates": "",
            "client_certificate": "",
            "client_key": ""
        }
    },
    "vad": {
        "mode": 0,
        "silence": 2,
        "status_messages": true
    }
}

Note that this supposes that you're using authentication and TLS for the connection to your MQTT broker and that this enables the experimental Voice Activity Detection (see below).

All keys in the configuration file are optional. The default behaviour is to connect with localhost:1883 without authentication and TLS and to use default as the site ID and disable Voice Activity Detection. A configuration file for this situation would like like this:

{
    "site": "default",
    "mqtt": {
        "host": "localhost",
        "port": 1883
    }
}

Currently Hermes Audio Server uses the system's default microphone and speaker. In a future version this will be configurable.

Voice Activity Detection

Voice Activity Detection is an experimental feature in Hermes Audio Server, which is disabled by default. It is based on py-webrtcvad and tries to suppress sending audio frames when there's no speech. Note that the success of this attempt highly depends on your microphone, your environment and your configuration of the VAD feature. Voice Activity Detection in Hermes Audio Server should not be considered a privacy feature, but a feature to save network bandwidth. If you really don't want to send audio frames on your network except when giving voice commands, you should run a wake word service on your device and only then start streaming audio to your Rhasspy server until the end of the command.

If the vad key is not specified in the configuration file, Voice Activity Detection is not enabled and all recorded audio frames are streamed continuously on the network. If you don't want this, specify the vad key to only stream audio when voice activity is detected. You can configure the VAD feature with the following subkeys:

• mode: This should be an integer between 0 and 3. 0 is the least aggressive about filtering out non-speech, 3 is the most aggressive. Defaults to 0.
• silence: This defines how much silence (no speech detected) in seconds has to go by before Hermes Audio Recorder considers it the end of a voice message. Defaults to 2. Make sure that this value is higher than or equal to min_sec in the configuration of WebRTCVAD for the command listener of Rhasspy, otherwise the audio stream for the command listener could be aborted too soon.
• status_messages: This is a boolean: true or false. Specifies whether or not Hermes Audio Recorder sends messages on MQTT when it detects the start or end of a voice message. Defaults to false. This is useful for debugging, when you want to find the right values for mode and silence.

Running Hermes Audio Server

Hermes Audio Server consists of two commands: Hermes Audio Player that receives WAV files on MQTT and plays them on the speaker, and Hermes Audio Recorder that records WAV files from the microphone and sends them as audio frames on MQTT.

You can run the Hermes Audio Player like this:

hermes-audio-player

You can run the Hermes Audio Recorder like this:

hermes-audio-recorder

You can run both, or only one of them if you only want to use the speaker or microphone.

Usage

Both commands know the --help option that gives you more information about the recognized options. For instance:

usage: hermes-audio-player [-h] [-v] [-V] [-c CONFIG]

hermes-audio-player is an audio server implementing the playback part of
    the Hermes protocol.

optional arguments:
  -h, --help            show this help message and exit
  -v, --verbose         use verbose output
  -V, --version         print version information and exit
  -c CONFIG, --config CONFIG
                        configuration file [default: /etc/hermes-audio-
                        server.json]
  -d, --daemon          run as daemon

Running as a service

After you have verified that Hermes Audio Server works by running the player and recorder manually, possibly in verbose mode, it's better to run both commands as services.

It's recommended to run the Hermes Audio Server commands as a system user. Create this user without a login shell and without creating a home directory for the user:

sudo useradd -r -s /bin/false hermes-audio-server

This user also needs access to your audio devices, so add them to the audio group:

sudo usermod -a -G audio hermes-audio-server

Then create systemd service files for the hermes-audio-player and hermes-audio-recorder commands and copy them to /etc/systemd/system.

If you want to run the commands as another user, then cange the lines with User and Group.

After this, you can start the player and recorder as services:

sudo systemctl start hermes-audio-player.service
sudo systemctl start hermes-audio-recorder.service

If you want them to start automatically after booting the computer, enable the services with:

sudo systemctl enable hermes-audio-player.service
sudo systemctl enable hermes-audio-recorder.service

Known issues / TODO list

• You can't choose the audio devices yet: the commands use the system's default microphone and speaker.
• This project is really a minimal implementation of the audio server part of the Hermes protocol, meant to be used with Rhasspy. It's not a drop-in replacement for snips-audio-server, as it lacks additional metadata in the WAV frames.

Changelog

• 0.2.0 (2019-06-04): Added logging and a daemon mode.
• 0.1.1 (2019-05-30): Made the audio player more robust when receiving an incorrect WAV file.
• 0.1.0 (2019-05-16): Added Voice Activity Detection option.
• 0.0.2 (2019-05-11): First public version.

Other interesting projects

If you find Hermes Audio Server interesting, have a look at the following projects too:

• Rhasspy: An offline, multilingual voice assistant toolkit that works with Home Assistant and is completely open source.
• Snips Led Control: An easy way to control the leds of your Snips-compatible device, with led patterns when the hotword is detected, the device is listening, speaking, idle, ...
• Matrix-Voice-ESP32-MQTT-Audio-Streamer: The equivalent of Hermes Audio Server for a Matrix Voice ESP32 board, including LED control and OTA updates.
• OpenSnips: A collection of open source projects related to the Snips voice platform.

License

This project is provided by Koen Vervloesem as open source software with the MIT license. See the LICENSE file for more information.