env_sound_esc10: Verified on the ECS-DoT EVB

Small 2-D CNN (conv + ReLU + max-pool stages, one linear classifier) that classifies a 5-second clip into one of 10 everyday sounds: dog, rooster, rain, sea waves, crackling fire, crying baby, sneezing, clock tick, helicopter, chainsaw.

Compiled with ECS-Compiler in ECS-Studio 1.140.0 and verified on a physical ECS-DoT EVB.

Getting Started

This repository is a ready-to-use model folder for ECS-Compiler, the AI pipeline of ECS-Studio. It holds the ONNX model and its calibration and evaluation data. The verified result was built with the default ECS-Compiler settings; keep them unchanged to reproduce it.

Compile the model

  1. Download the model folder, either from the Files tab of this page or with the Hugging Face CLI:

    hf download emass-ai/env_sound_esc10 --local-dir env_sound_esc10
    
  2. In ECS-Studio (1.140.0 or later), open a project for the ECS-DoT EVB, open the ECS-Compiler panel and choose esc10_cnn.onnx as the model file. The panel loads esc10_cnn_train.npz (calibration) and esc10_cnn_test.npz (evaluation) from the same folder automatically.

  3. Click Start Pipeline, then Build and Upload to run the model on the board. The ECS-Compiler User Guide describes each step.

Model Details

Model type environmental_sound_classification
Architecture Conv 3x3 1->16 +BN -> ReLU -> MaxPool 2x2 -> Conv 3x3 16->24 +BN -> ReLU -> MaxPool 2x2 -> Conv 3x3 24->64 +BN -> ReLU -> Flatten (64x12x3) -> Linear 10 -> Softmax
Input input [1, 1, 50, 12] float32 (Log-mel spectrogram image of a 5 s clip: 50 frames (100 ms hop) x 12 mel bands, one channel)
Preprocessing 5 s of audio resampled to 16 kHz mono, zero-padded/cropped to 80000 samples. STFT: n_fft 2048, hop 1600, Hann window, centered, first 50 frames; power spectrum; 12 mel bands 50-8000 Hz (librosa.filters.mel, sr 16000, n_fft 2048); log(mel + 1e-5) -> [50, 12] (time, mel). Per-band standardization (x - mean) / std with mean = [-2.737, -3.013, -3.277, -3.652, -3.946, -4.208, -4.574, -5.009, -5.454, -5.892, -6.204, -6.657] and std = [5.026, 4.967, 4.966, 4.867, 4.813, 4.735, 4.613, 4.461, 4.292, 4.121, 4.020, 3.830] (computed on all 400 ESC-10 clips).
Output output [1, 10] float32
Labels dog, rooster, rain, sea_waves, crackling_fire, crying_baby, sneezing, clock_tick, helicopter, chainsaw
Model size (float) 160.2 KB
Parameters 40,578

Performance Summary

Model Precision Board Inference Time Library Size Compute Unit
env_sound_esc10 int8 ECS-DoT EVB 6.01 ms 68.0 KB DLA + CPU

Inference time is the measured time of one dla_infer() call on the board (median of 204 runs, ECS-DoT at 50 MHz). Library size is the compiled model library libdla.a (9 layers on the DLA, 1 on the CPU).

On-board verification: Run on 3 real samples. The board's output matched the compiler's expected output exactly, and its prediction matched ONNX Runtime on every sample (cosine similarity 0.9999 or higher). On the 200-sample evaluation set the board (int8) scored 75.00% against 75.00% for the float model on the same samples, and agreed with ONNX Runtime on 197 of 200 samples.

Accuracy

77.50% top-1 accuracy on ESC-10, fold 5 (80 clips), trained on folds 1-4; 80.0% mean over the 5 official folds (5-fold cross-validation; the ESC-50 paper's CNN reaches 80.5% on ESC-10). Measured with ONNX Runtime on the float model. The 200-sample evaluation file holds the 80 fold-5 clips plus 120 time-shifted copies of them.

Training Data

  • ESC-10 (subset of ESC-50), ESC-50 master (2015 release), esc10 == True clips only: https://github.com/karolpiczak/ESC-50 (licence: CC BY 3.0 Unported for the ESC-10 subset (the full ESC-50 is CC BY-NC and is NOT used). README: 'A smaller subset (clips tagged as ESC-10) is distributed under CC BY (Attribution).' LICENSE file: 'The ESC-10 subset is licensed as a Creative Commons Attribution 3.0 Unported (https://creativecommons.org/licenses/by/3.0/) dataset.' Individual clips are CC0 or CC BY (Freesound).).
  • Citation: Karol J. Piczak. ESC: Dataset for Environmental Sound Classification. Proceedings of the 23rd ACM International Conference on Multimedia, 2015. https://doi.org/10.1145/2733373.2806390

Files

File Contents Download
esc10_cnn.onnx The ONNX model Download
esc10_cnn_train.npz Calibration set the board build quantizes with Download
esc10_cnn_test.npz Evaluation set Download
sample_input.npz The real samples run on the board (x, with labels y) Download
LICENSE License Download
README.md This model card Download

Intended Use

  • This model is provided to evaluate and validate ECS-Compiler and the ECS-DoT. It is not a product and has not been tested for any particular end use.

License

  • This model is released under apache-2.0; see LICENSE. The training data keeps its own licence (see Training Data).

Provenance

  • Verified with ECS-Studio 1.140.0 (ECS-Compiler engine 0.6.0) on the ECS-DoT EVB, 2026-11-17. Build fingerprint 3f7df3b32741.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support