DawsonPrince/iHear-SAFE-v4.7.1

iHear-SAFE v4.7.1

iHear-SAFE v4.7.1 is an experimental embedded audio model developed for the iHear hearing-assistance research project. It combines speech-enhancement masking with hierarchical detection of selected safety-critical sounds.

Supported safety sounds

The v4.7.1 subtype head supports:

1. siren
2. car_horn
3. fire_alarm

doorbell and warning_shout are treated as salient but unsupported negative classes during the current development stage.

Model interface

Input

• Sample rate: 16 kHz
• Mono audio
• STFT:
  • n_fft = 256
  • win_length = 256
  • Hann window
  • hop_length = 128
  • center = True
• Feature: log(|STFT| + 1e-8)
• Context: 5 frames
• TFLite input shape: [1, 5, 129, 1]
• INT8 input quantization:
  • scale: 0.08686613291501999
  • zero point: 83

Outputs

The model has three outputs, identified by shape rather than output order:

• Enhancement mask: [1, 129]
• Supported-alert gate: [1, 1]
• Alert subtype: [1, 3]
  • index 0: siren
  • index 1: car horn
  • index 2: fire alarm

All three INT8 outputs use scale 0.00390625 and zero point -128.

Selected deployment policy

The policy selected on the fixed development validation set is:

Parameter	Value
Gate threshold	0.70
Subtype threshold	0.75
Combined threshold	0.65
Minimum consecutive frames	12
Minimum active-frame ratio	0.04
Safety-preservation mask floor	0.85

The combined score is gate × subtype.

Development validation results

The selected INT8 policy achieved:

Metric	Result
Siren recall	84%
Car-horn recall	80%
Fire-alarm recall	92%
Overall supported recall	85.3%
Doorbell rejection	96%
Warning-shout rejection	68%
Unsupported rejection	82%
Ordinary-noise rejection	99%
Hard-negative rejection	89%
Deployment-negative rejection	90%
Deployment-balanced score	87.7%

The fixed validation set contained 450 deterministic mixtures: 50 each for siren, car horn, fire alarm, doorbell, and warning shout; 100 ordinary no-alert mixtures; and 100 hard-negative mixtures.

Important: the 50 fire-alarm validation mixtures were generated from only 8 independent fire-alarm validation recordings. The 92% fire-alarm result is therefore a development-validation result, not a final independent estimate of real-world fire-alarm performance.

Quantization

Balanced post-training INT8 calibration used representative examples from seven groups:

• siren
• car horn
• fire alarm
• doorbell
• warning shout
• ordinary no-alert
• hard negative

Float32-to-INT8 mean absolute error:

Output	MAE
Mask	0.0111
Gate	0.0102
Subtype	0.0094

The gate range was preserved well:

• Float32: approximately 0.009 – 0.996
• INT8: approximately 0.008 – 0.996

Intended use

This release is intended for:

• research on embedded hearing-assistance systems,
• experimentation with safety-sound preservation,
• TFLite/ESP32-S3 deployment studies,
• reproduction and further development of the iHear prototype.

Limitations and safety notice

This model is a research prototype.

• It has not been clinically validated as a hearing aid.
• It has not been certified as a fire alarm, emergency alert, or life-safety device.
• It must not be relied on as the sole method for detecting hazards or emergencies.
• It currently recognizes only three supported alert classes.
• Warning-shout rejection remains a known weakness.
• Fire-alarm evaluation currently uses a small number of independent source recordings.
• Real-world performance can vary with microphones, acoustics, signal level, noise, distance, hardware, and preprocessing.

Demo

A public Gradio demo can be deployed from the companion Hugging Face Space:

https://huggingface.co/spaces/DawsonPrince/iHear-SAFE-v4.7.1-demo

Files

The main deployment model is:

ihear_safe_v4_7_1_int8.tflite

Additional JSON/CSV files in this repository document the selected policy, class mapping, DSP configuration, and quantization validation.

Citation

If you use this model in research, please cite the iHear project/relevant publication when the project team provides the final citation.

License

No explicit model license is declared in this initial release template. The project owners should select an appropriate license after confirming institutional ownership and the terms of the training datasets.