README.md

metadata

library_name: pytorch
license: apache-2.0

DS-CNN Keyword Spotting (Ambarella N1-655 CVflow v3 NVP)

Depthwise Separable Convolutional Neural Network (DS-CNN) for 12-class Keyword Spotting (KWS) on 1-second 16 kHz audio clips, trained on the Google Speech Commands dataset.

Original paper: Hello Edge: Keyword Spotting on Microcontrollers

Model Configuration:

• Reference implementation: https://github.com/ARM-software/ML-KWS-for-MCU
• Original Weight: https://huggingface.co/EclipseAidge/ds_cnn
• Dataset: Google Speech Commands v0.02 (CC-BY-4.0)
• License: Apache-2.0 (LICENSE)
• Support Cooper version:
  • CNNGen: [3.9.0 / n1-655x] (cavalry_gen 3.0.4)

Upstream Attribution & Modification Notice

• Upstream Copyright: Copyright (c) Arm Limited and Contributors (ARM-software/ML-KWS-for-MCU), MLCommons Association (mlcommons/tiny), and Eclipse Aidge (EclipseAidge/ds_cnn), licensed under the Apache License, Version 2.0. Trained on Google Speech Commands v0.02 (Creative Commons Attribution 4.0 International, Copyright Google LLC).
• Modification Notice (Apache-2.0 Section 4(b)): Modified by Ambarella International LLC (2026). Transposed input layout to NCHW [1, 1, 49, 10] IEEE FP16 and compiled via Ambarella Cooper SDK CNNGen 3.9.0 (cavalry_gen 3.0.4) into CVflow v3 N1-655 (vdg_arch = 0xB) Cavalry DAG binary format (n1-655_kws_dscnn_amba_optimized.bin).

NVP Port & Runtime Specification

• Target SoC: Ambarella N1-655 (vdg_arch = 0xB, CVflow v3 cavalry_gen 3.0.4)
• Input port: StatefulPartitionedCall_functional_1_conv2d_BiasAdd_in0 [1, 1, 49, 10] IEEE FP16 (1 s @ 16 kHz, 40 ms window, 20 ms hop, 40 HTK Mel bins [20..4000 Hz], DCT-II top 10 MFCCs)
• Output port: dense [1, 12] IEEE FP16 (12 Speech Commands classes: silence, unknown, yes, no, up, down, left, right, on, off, stop, go)

Model	Device	Compression	Model Link
DSCNN	N1-655	Amba_optimized	Model_Link