Estimating Pitch Information From Simulated Cochlear Implant Signals With Deep Neural Networks.

Area of Science:

• Auditory Neuroscience
• Signal Processing
• Medical Devices

Background:

• Cochlear implant (CI) users often exhibit limited sensitivity to fundamental frequency (F0), impacting pitch perception.
• This insensitivity is typically attributed to spectral and temporal resolution limits in CI signals.
• However, individual variability suggests other physiological factors like neuroplasticity also play a role.

Purpose of the Study:

• To determine the maximum fundamental frequency (F0) information transferable via cochlear implant (CI) signals.
• To investigate the impact of spectral (electrode channels) and temporal (pulse rate) resolutions on F0 estimation.
• To compare CI signal F0 estimation with raw waveform processing, especially under noisy conditions.

Main Methods:

• Developed a deep neural network (CI model) to decode F0 from simulated CI signals.
• Varied the number of electrode channels and pulse rate to simulate different spectral and temporal resolutions.
• Compared CI model performance against a control waveform model using raw audio input.

Main Results:

• F0 estimation performance improved with increased electrode channels and pulse rate.
• Under quiet conditions, CI model performance was comparable to the waveform model.
• CI model performance degraded more than the waveform model in the presence of background noise.
• Pulse rate significantly impacted F0 estimation performance, particularly in noise.

Conclusions:

• CI signals contain sufficient F0 information for perception, especially in quiet environments.
• Temporal resolution, represented by pulse rate, is critical for pitch perception in noisy conditions for CI users.
• Individual differences in F0 sensitivity may stem from neural factors beyond signal processing limitations.