Snips running in Docker using Pulseaudio

If you are using snips in a desktop environment that uses pulseaudio to share/mix your sound devices, then you'll probably want the snips beeps/tts on pulse as well. Especially if dmix is not supported on your system.

This repo is an example of how to extend snipsdocker/platform with pulseaudio support. It is assumed you have setup pulseaudio with your default sink and source devices and those are the ones you want to use with snips.

Running the docker image

Build the docker image

docker build -t snips-pulse-docker .

and run it

./run-snips.sh

Then test it out in another terminal

docker exec -it snips aplay /usr/share/snips/assistant/custom_dialogue/sound/start_of_input.wav
docker exec -it snips mosquitto_pub -h localhost -p 1883 -t hermes/tts/say -m '{"siteId":"default","text":"hello how are you"}'

Issues

With default pulse configuration, sound coming out of snips-audio-server comes out crackling. Both TTS and the hotword beeps. However I did finally fix that by tuning my pulse config, details below. Note this is specifically for my Odroid C2 + Hifi Shield 2 (card 1) setup using PS3 eye mic (card 2) and ignoring the HDMI sound (card 0).

First up /etc/asound.conf. Specifically the rate / format / channels combo was necessary for me to get sound working from the docker container, but not from snips-audio-server.

pcm.!default {
  type asym
  playback.pcm {
    type plug
    slave {
      pcm "hw:1,0"
      rate 48000
      format "S16_LE"
      channels 2
    }
  }
  capture.pcm {
    type plug
    slave.pcm "hw:2,0"
  }
}

ctl.!default {
  type hw
  card 1
}

Next up was modifying pulse configuration. I had noticed logs in pulseaudio along the lines of

(   9.856|   0.000) D: [pulseaudio] protocol-native.c: Early requests mode enabled, configuring sink latency to minreq.
(   9.856|   0.000) D: [pulseaudio] protocol-native.c: Requested latency=20.00 ms, Received latency=92.88ms

and so decided to go mess with buffer sizes.

This guide https://wiki.archlinux.org/index.php/PulseAudio/Troubleshooting#Setting_the_default_fragment_number_and_buffer_size_in_PulseAudio was very useful.

I modified /etc/pulse/default.pa to add tsched=0

load-module module-udev-detect tsched=0

then modified /etc/pulse/daemon.conf

default-sample-format = s16le
default-sample-rate = 48000
default-sample-channels = 2

default-fragments = 5
default-fragment-size-msec = 4

The first three were to match my asound.conf and the latter two were to match the requested 20ms latency. And that's it, I could now hear sound clearly from snips-audio-server.