The solution – real time

Getting this to work in real time is tough

• 16 kHz, 16-bit speech samples
• Single channel
• Floating point implementation needs single-cycle FPU
• 400MHz ARM Cortex-M
• Approx. 1.6 Mbytes model size
• Fixed point potential
• Model parameters are 32-bit, can reduce to 16-bit
• Currently trained with speech, domestic noise and reverberation – needs fine tuning for different domain noises