search

rhasspy / wyoming-openwakeword

Wyoming protocol server for openWakeWord wake word detection system
MIT License
97 stars 25 forks source link

CPU Usage maxed and no functionality on Raspberry Pi 0w #30

Open JamesOldigesOther opened 4 weeks ago

JamesOldigesOther commented 4 weeks ago

I've gone ahead and compiled TF-Lite for the Raspi Architecture without much issue. However, in running it in conjunction with Wyoming-Sattelite, OpenWakeWord is using 100% of the CPU.

As far as I can tell, it isn't functional/functioning

Without OpenWakeWord, I'm able to interface as expected with Wyoming Satellite only.

I am running it with the debug flag, though my console messages are largely lacking ExecStart=/home/james/wyoming-openwakeword/script/run --uri 'tcp://127.0.0.1:10400' --debug

Aug 15 19:22:27 Debra02 systemd[1]: Started wyoming-openwakeword.service - Wyoming openWakeWord.
Aug 15 19:22:35 Debra02 run[4544]: DEBUG:root:Namespace(uri='tcp://127.0.0.1:10400', models_dir=PosixPath('/home/james/wyoming-openwakeword/wyoming_openwakeword/models'), custom_model_dir=[], preload_model=[], threshold=0.5, trigger_level=1, output_dir=None, debug=True, log_format='%(levelname)s:%(name)s:%(message)s', debug_probability=False, version=False, model=[])
Aug 15 19:22:35 Debra02 run[4544]: DEBUG:root:Loading /home/james/wyoming-openwakeword/wyoming_openwakeword/models/melspectrogram.tflite
Aug 15 19:22:35 Debra02 run[4544]: DEBUG:root:Loading /home/james/wyoming-openwakeword/wyoming_openwakeword/models/embedding_model.tflite
Aug 15 19:22:35 Debra02 run[4544]: INFO:root:Ready
Aug 15 19:22:35 Debra02 run[4544]: DEBUG:wyoming_openwakeword.handler:Client connected: 110594151777903
Aug 15 19:22:35 Debra02 run[4544]: DEBUG:root:Loading ok_nabu_v0.1 from /home/james/wyoming-openwakeword/wyoming_openwakeword/models/ok_nabu_v0.1.tflite
Aug 15 19:22:35 Debra02 run[4544]: DEBUG:wyoming_openwakeword.handler:Started thread for ok_nabu_v0.1
Aug 15 19:22:36 Debra02 run[4544]: Error in cpuinfo: failed to parse file /sys/devices/system/cpu/cpu0/topology/physical_package_id: "-1
Aug 15 19:22:36 Debra02 run[4544]: " is not an unsigned number
Aug 15 19:22:43 Debra02 run[4544]: DEBUG:root:Loading ok_nabu_v0.1 from /home/james/wyoming-openwakeword/wyoming_openwakeword/models/ok_nabu_v0.1.tflite
Aug 15 19:22:43 Debra02 run[4544]: DEBUG:wyoming_openwakeword.handler:Started thread for ok_nabu_v0.1
Aug 15 19:22:43 Debra02 run[4544]: DEBUG:wyoming_openwakeword.handler:Sent info to client: 110594151777903
Joldiges commented 3 weeks ago

Consolidating this thread.

I have it somewhat "working"

  1. ONNX-runtime will build on-target (or qemu docker) on version v1.16.3 and latest numpy < 2.0 (For Wyoming Satellite)
  2. tflite-runtime was a pain. build on-target (or qemu docker) with the command ./tensorflow/lite/tools/pip_package/build_pip_package_with_cmake.sh native
    1. You'll need to add this to your service to link libatomic Environment="LD_PRELOAD=/usr/lib/arm-linux-gnueabihf/libatomic.so.1.2.0"
    2. you'll need to change the wyoming-openwakeword setup (or your pip config) to install the local instance of your newly compiled tflite-runtime.

For days and days of work, it resolved to be quite simple. A lot of wasted time was attempted cross-compiling the complex build systems. Lesson learned.

Unfortunately, the CPU is maxed out and the actual wake-word capturing is... not right.
I'm getting consistently low readings Aug 17 01:22:03 Debra02 run[2193]: DEBUG:root:client=3665034168370, wake_word=ok_nabu_v0.1, probability=0.0009... But when I say the keyword it does (consistently raise the probability slightly Aug 17 01:22:03 Debra02 run[2193]: DEBUG:root:client=3665034168370, wake_word=ok_nabu_v0.1, probability=0.0011...

My anticipation is that it's taking longer to process the audio clip than the time between them, cutting off the processing.

But my assumption is that I can't really change the sample rate, because the real metric is "processing faster than real life" Therefore, the solution would be to process a "simpler" model, which is outside of my scope. I hope someone may read this and pick up where I've left off, or guide me on how to proceed.