Skip to content

Keyword Spotting (KWS)

The Keyword Spotting (KWS) module in soundKIT enables accurate detection of predefined spoken keywords in diverse and noisy conditions. It is designed for real-time, low-power deployments using embedded platforms.

It supports:

Engineered for deployment on Ambiq's ultra-low power SoCs, KWS offers high accuracy even in constrained edge environments.

📘 Try it now: Explore the KWS Tutorial Notebook to get started.

Features

  • Keyword spotting with frame-level granularity
  • CRNN-based acoustic modeling
  • Support for long audio (15s) with extension frames
  • TFLite and C-array export for embedded use
  • PC-based and EVB-based real-time demos

Install soundKIT

Refer to the QuickStart Guide for installation and setup instructions.

KWS Task Modes

Use the soundkit CLI to run KWS tasks through different modes using a configuration file like kws.yaml.

kws.yaml 
name: galaxy
project: kws
job_dir: ./soundkit/tasks/${project}
...

KWS Task Mode Overview

Prepares training and validation samples by injecting keywords into a diverse background of speech and noise at different SNR levels.

soundkit -t kws -m data -c configs/kws/kws.yaml

See Data for dataset setup details.

Trains a keyword spotting model with CRNN architecture, focal loss, and SNR-augmented training.

soundkit -t kws -m train -c configs/kws/kws.yaml

Optional: run TensorBoard for live monitoring:

soundkit -t kws -m train --tensorboard -c configs/kws/kws.yaml

See Train for training methodology.

Evaluates model performance on sample audio files, producing keyword activity predictions.

soundkit -t kws -m evaluate -c configs/kws/kws.yaml

See Evaluate for usage details.

Converts the trained model into TFLite and C-format for deployment on Ambiq EVB or other embedded targets.

soundkit -t kws -m export -c configs/kws/kws.yaml

See Export for model format options.

Test the model in real-time either using PC or EVB hardware. We suggest to run on your PC first and try on EVB later:

soundkit -t kws -m demo -c configs/kws/kws.yaml demo.platform=pc
# or
soundkit -t kws -m demo -c configs/kws/kws.yaml demo.platform=evb

See Demo to test live detection.