SparkNet: Sparse Binarization for Fast Keyword Spotting (Interspeech 2024)

The paper is here.

Main features:

• Efficient (Tiny and Fast) Neural Network
• SOTA results in edge-device setup for KWS task
• Appropriate for micro-controllers

Updates: (23/03/2025)

• The code is released (training and inference)
  • The model is tested in Python 3.10.7
  • Install the requirements: pip install -r requirements.txt
  • Please notice, that the version of NeMo we use is different with our modifications, especially in file nemo/collections/asr/models/classification_models.py
• We release the checkpoints for configurations C=4,8,16,32 trained on Google Speech Commands v2 on 12 labels
• Test example: put your recordings in wavs directory and run the command:
  • python inference.py --model_path ckpt/kws_C_16.ckpt

Model:

Results:

Comparison of different methods trained on Google Speech Commands v1, v2 (SC1, SC2) datasets and tested on the official test set of keyword spotting task with 12 targets. While being with the same accuracy and number parameters our model has x4 less multiply-accumulate operations (MACs)

Model	Params	MACs	SC1	SC2
TinySpeech-X [Wong et al. 2020]	10.8K	10.9M	94.6 ± 0.00	-
res8-narrow [Tang et al. 2018]	19.9K	5.65M	90.1 ± 0.98	-
DS-ResNet10 [Xu et al. 2020]	10K	5.8M	95.2 ± 0.36	-
BC-ResNet-1 [Kim et al. 2021]	9,232	3.6M	96.6 ± 0.21	96.9 ± 0.30
SparkNet[C=32]	11,500	1.2M	96.2 ± 0.19	97.0 ± 0.18
TinySpeech-Z [Wong et al. 2020]	2.7K	2.6M	92.4 ± 0.00	-
BC-ResNet-0.625 [Kim et al. 2021]	4,585	1.9M	95.2 ± 0.37	95.4 ± 0.31
SparkNet[C=16]	4,636	454.5K	95.3 ± 0.33	95.7 ± 0.17

Reduced versions:

he accuracy of the proposed model as a function of number of parameters. The smallest model is still able to produce ~ 83% accuracy by using only 1.4K parameters and 105K MACs.

Model Version	Params	MACs	SC1	SC2
SparkNet[C=32]	11,500	1.2M	96.2 ± 0.19	97.1 ± 0.30
SparkNet[C=16]	4,636	454.5K	95.3 ± 0.33	95.7 ± 0.30
SparkNet[C=8]	2,292	190K	91.6 ± 0.76	92.1 ± 0.33
SparkNet[C=4]	1,416	105K	82.3 ± 1.91	83.5 ± 0.60

This markdown table retains the structure and formatting of your original LaTeX table, with columns for model version, parameters, MACs, and accuracy (SC1 and SC2).

Acknowledgments

This project builds upon the work and tools provided by the open-source community. Special thanks to:

• NeMo
• THOP

Cite this work:

@inproceedings{svirsky2024sparse,
  title={Sparse Binarization for Fast Keyword Spotting},
  author={Svirsky, Jonathan and Shaham, Uri and Lindenbaum, Ofir},
  booktitle={Proc. Interspeech 2024},
  pages={3010--3014},
  year={2024}
}