Dataset and labelling plan

1. Use consented recordings from multiple speakers
2. Include unknown speech, silence and representative noise
3. Separate speakers between training and test where possible
4. Fix sample rate, clip length and amplitude policy
5. Do not retain unnecessary raw voice data

Training and validation pipeline

1. Verify audio clipping and DC offset
2. Create identical spectral preprocessing for training/device
3. Review per-class precision, recall and confusion
4. Quantise and measure model-size/accuracy change
5. Test quiet, near-field and noisy conditions
6. Set confidence, smoothing and refractory time

Model deployment procedure

1. Allocate ring buffers and tensor arena statically
2. Run feature extraction in bounded frames
3. Use only required TFLM operators
4. Measure worst-case inference and buffer headroom
5. Apply confidence smoothing and unknown class
6. Disable or indicate capture according to privacy policy

Memory and performance targets

Deployment checkpoints

audio_ring_push(new_samples);
while (feature_frame_ready()) {
  make_features(audio_ring, input_tensor);
  invoke();
  update_smoothed_scores(output_tensor);
  if (stable_command_above_threshold()) emit_command();
}

Troubleshooting

Safety, privacy and model limits

• Get consent before recording voices.
• Do not upload private conversations to training services.
• Provide a visible indication when audio capture is active.
• Never use a classroom keyword model for security authentication.

Interview and viva questions

1. Why should speakers be separated across dataset splits?
2. What is an unknown class?
3. Why use score smoothing and refractory time?
4. How do audio buffers contribute to peak RAM?
5. Which metrics matter when false activation is costly?

Lesson notes

Responsible TinyML workflow

Keep raw data, preprocessing, model version, compiler options and measured device results together. A desktop accuracy score does not prove embedded performance. Validate representative unseen samples on the actual target and define an explicit fallback for low confidence or out-of-distribution input.

Deployment evidence

Record the dataset split, confusion matrix, exported model hash, firmware build, board revision, peak memory, average and worst-case latency, power conditions and failures. Never treat an educational classifier as a safety-certified decision system.