eic-speech-nemo

NeMo-based speech pipeline with Wake Word detection, VAD, and ASR. It requires very bare minimum dependencies, and I like it this way. Completely useless outside our settings.

dependencies

• PortAudio C lib (sudo apt-get install portaudio19-dev)
• torch >= 2.7
• numpy
• sounddevice Clean API, callback-based, low-latency, better than pyaudio from my experience with less clunkier API

download weights

Put your weights under weights/ directory.

• nemotron_asr.pt
• silero_vad.pt
• wakeword_mel.pt
• wakeword_embed.pt
• wakeword_classifier.pt

demo

bash setup.sh
source .venv/bin/activate
eic-speech-nemo

Installation

(Install PortAudio beforehands)

# sudo apt-get install portaudio19-dev
pip install -e .

Usage

from eic_speech_nemo import ASRServer, State
from eic_speech_nemo.server import ASRConfig

server = ASRServer(config=ASRConfig(model_dir="./models"))
server.start()
server.run_forever()