- Consumer sends WhatsApp voice note - records 10 seconds of the product running with the fault audible
- Audio preprocessing - noise reduction, ambient filtering, and extraction of the product's core sound signature
- Spectrogram analysis - AI converts audio to frequency-time representation and matches against known fault patterns for that product category
- Diagnosis returned - fault type, confidence level (high/medium/low), severity (urgent/moderate/minor), and plain-language explanation
- Action recommendation - self-fix with video guide, or technician booking with fault and required parts pre-identified
- Feedback loop - technician confirms or corrects diagnosis after visit, improving the model over time
Technical Foundation
Audio ML Model: Convolutional neural network trained on spectrogram images of product sounds. Transfer learning from industrial predictive maintenance models, fine-tuned on Havells product-specific audio data.
Training Data: Initial dataset built from service center recordings - technicians record products before and after repair. Augmented with synthetic fault audio generated from known acoustic models of bearing, blade, and motor faults.
Accuracy Target: 80%+ diagnosis accuracy within 6 months of deployment, improving to 90%+ as the feedback loop from technician confirmations grows the training dataset.
Ambient Noise Handling: Indian homes have background noise (TV, traffic, other appliances). The preprocessing pipeline isolates the target product's sound using frequency band filtering specific to each product category.