title: Submission for the sound detection emoji: 🔥 colorFrom: yellow colorTo: green sdk: docker pinned: false —

Random Baseline Model for Climate Disinformation Classification

Model Description

This is a quantified baseline model for sound classification on a rp2040, for the Frugal AI Challenge 2025. Why? Because the hardware runs real-time classification using 0.25W =)

Some under the hood peek:

• Refer to the implementation notebook
• Refer to the training notebook

Hardware description

A rp2040 microcontroler, with a PDM mic, and and OLED screen.

Check of consumptions

Consumption for realtime consumption

• without inference: 3.060 mWh/min –> 186mW
• with inference: 3.107 mWh/min –> 186.5mW

Intended Use

• Primary intended uses: Demonstrating that it’s possible to do real-time classification, with okay-ish accuracy, on a 2$ microcontroler.
• Primary intended users: Researchers and developers participating in the Frugal AI Challenge, with a focus on the EdgeAI cosiderations.
• Out-of-scope use cases: Not intended for production use or real-world classification tasks

Training Data

The model uses the rfcx/frugalai dataset:

• Size: ~50k samples of 3s
• Split: 80% train, 20% test
• 2 categories (chainsaw/no chainsaw)

Beware, there are some inconsistencies in the length.

Labels

1. Chainsaw
2. Environment

Performance

Refer to the implementation notebook

Metrics

• Accuracy: ~87.5%
• Environmental Impact:
  • Emissions tracked in gCO2eq : 0.287645456075
  • Energy consumption tracked in Wh : 0.779243206008

Result

{
  "username": "kelu124",
  "space_url": "https://huggingface.co/spaces/kelu124/pico-sound",
  "submission_timestamp": "2025-01-30T11:22:21.436878",
  "model_description": "rp2040 inference",
  "accuracy": 0.8713235294117647,
  "energy_consumed_wh": 0.7792432060080811,
  "emissions_gco2eq": 0.28764545607537884,
  "emissions_data": {
    "run_id": "74e7c14a-bcae-4f47-b010-e9c3c576d4f2",
    "duration": 22.433365739000237,
    "emissions": 0.00028764545607537885,
    "emissions_rate": 0.000012822262056714762,
    "cpu_power": 105,
    "gpu_power": 14.30419952643268,
    "ram_power": 5.74672794342041,
    "cpu_energy": 0.0006542979921624957,
    "gpu_energy": 0.00008913562686400055,
    "ram_energy": 0.0000358095869815849,
    "energy_consumed": 0.000779243206008081,
    "country_name": "United States",
    "country_iso_code": "USA",
    "region": "virginia",
    "cloud_provider": "",
    "cloud_region": "",
    "os": "Linux-5.10.230-223.885.amzn2.x86_64-x86_64-with-glibc2.36",
    "python_version": "3.11.11",
    "codecarbon_version": "2.8.3",
    "cpu_count": 4,
    "cpu_model": "Intel(R) Xeon(R) Platinum 8259CL CPU @ 2.50GHz",
    "gpu_count": 1,
    "gpu_model": "1 x Tesla T4",
    "ram_total_size": 15.324607849121094,
    "tracking_mode": "machine",
    "on_cloud": "N",
    "pue": 1
  },
  "api_route": "/audio",
  "dataset_config": {
    "dataset_name": "rfcx/frugalai",
    "test_size": 0.2,
    "test_seed": 42
  }
}

Model Architecture

The ETL:

• A spectrogram transformation (3s sample at 12kHz to a 35x129 image

Model layers include:

• Resize to 32x32 image
• 2D Convolution
• MaxPool2D
• Flatten
• Dropout
• Dense - 2 categories

Based on this work by ARM people

Environmental Impact

Environmental impact is tracked using CodeCarbon, measuring:

• Carbon emissions during inference
• Energy consumption during inference

This tracking helps establish a baseline for the environmental impact of model deployment and inference.

Limitations

• Not suitable for any real-world applications

Ethical Considerations

• Environmental impact is tracked to promote awareness of AI’s carbon footprint ```