Aborted Error

[theMARVguy]

I have been trying to get RealtimeSTT to work on my 1GB RAM Raspberry Pi 3, but I keep getting this one error whenever I try to launch a python program that imports RealtimeSTT. I have tried increasing my swap size from 100 to 1024 to 2048, but that has not worked. Can I have some help fixing this?

This is what it looks like when I do "free -h"

[KoljaB]

RealtimeSTT on a 1 GB Pi 3 might be not realistic, esp if it's 32 bit OS. I think I'd just use the Pi only as audio client and serve the realtime translation from a stronger machine.

The error looks like a native ML dependency is aborting in C++ before RealtimeSTT can handle it as a python exception.

If you still want to give it a try, I would first check every component separately:

python3 -c "import onnxruntime; print('onnxruntime ok')"
python3 -c "from openwakeword.model import Model; print('openwakeword ok')"
python3 -c "import ctranslate2; print('ctranslate2 ok')"
python3 -c "import faster_whisper; print('faster_whisper ok')"
python3 -c "import torch; model, utils = torch.hub.load(repo_or_dir='snakers4/silero-vad', model='silero_vad', verbose=False); print('silero ok')"
python3 -c "import webrtcvad; vad = webrtcvad.Vad(); vad.set_mode(3); print('webrtcvad ok')"

If one of these commands aborts, then that dependency is the source of the crash. Can't do much then - would require huge changes to the code.

If all checks pass, try low resource RealtimeSTT settings:

from RealtimeSTT import AudioToTextRecorder

recorder = AudioToTextRecorder(
    model="tiny.en", # or "tiny"
    device="cpu",
    compute_type="int8",
    enable_realtime_transcription=False,
    beam_size=1,
    faster_whisper_vad_filter=False,
    wake_words="",
    wakeword_backend="",
    spinner=False,
)

print("Say something...")
text = recorder.text()
print(text)

You might want to try some lower batch_size and realtime_batch_size values too.

It might also help to limit native thread usage before starting Python:

export OMP_NUM_THREADS=1
export OPENBLAS_NUM_THREADS=1
export MKL_NUM_THREADS=1
export NUMEXPR_NUM_THREADS=1

Maybe 64-bit Raspberry Pi OS will be needed, I guess some dependencies might have better support on aarch64 vs 32-bit ARM.

[theMARVguy]

Would it be feasible to use a Raspberry Pi 4 with 2 GB RAM or is more needed (4 GB)?