Quantifying Cochlear Implant Users’ Ability for Speaker Identification Using CI Auditory Stimuli

Nursadul Mamun, Ria Ghosh, John H.l. HansenInterspeech 2019

TL;DR

This paper investigates how well cochlear implant (CI) users can identify speakers using simulated auditory stimuli derived from their implants. The study finds that while CI users can identify speakers in quiet conditions, their performance significantly degrades with more speakers and in noisy environments, highlighting a critical limitation in current CI technology for robust speaker recognition.

What Problem Does This Paper Address?

Speaker identification (SID) is a crucial aspect of human auditory perception, allowing us to distinguish individuals based on their voice. While normal-hearing individuals perform this task with high accuracy, even in challenging conditions, people with sensorineural hearing loss who use cochlear implants (CIs) face significant difficulties. CIs process sound through a limited number of channels, which inherently restricts the richness of auditory information available to the user. This limitation is particularly problematic for tasks that rely on subtle spectral and temporal cues, such as speaker identification, which often depend on features like formant frequencies and pitch that are significantly attenuated or lost in CI processing. This paper aims to quantify the extent of this deficit by evaluating CI users' speaker identification capabilities using a simulated CI auditory stimulus.

Methods

The core of this research lies in simulating the auditory experience of a CI user and using that simulation for speaker identification. The authors employ the Advanced Combined Encoder (ACE) signal processing strategy, a common approach in CI technology, to convert incoming speech signals into electrodograms. Electrodograms are essentially time-electrode representations that mimic the electrical stimulation patterns delivered by a CI to the auditory nerve.

For speaker identification, two main approaches were used:

  • GMM-UBM (Gaussian Mixture Model - Universal Background Model): A standard technique in speaker recognition where a model for each speaker is adapted from a general background model.
  • I-Vector-PLDA (I-vector - Probabilistic Linear Discriminant Analysis): A more advanced technique that captures speaker and channel variability in a low-dimensional space, offering improved robustness.

The study trained speaker models using these methods on electrodograms and then tested their performance under various conditions.

Datasets

The study utilized a combination of databases for training and testing:

  • Text-Dependent Database: A small, custom database from the University of Malaya, consisting of 390 signals from 39 speakers. This was used for initial evaluation in quiet conditions.
  • Text-Independent Databases:
    • SRE-18: A dataset from the NIST Speaker Recognition Evaluations, known for its challenging telephone-quality speech.
  • YOHO: A large-scale, high-quality dataset collected by the ITT technical institute, often used for speaker identification and verification.

These text-independent databases were used to test the system's ability to identify speakers without relying on specific phrases, simulating real-world scenarios.

Key Insights and Ideas

  • Electrodograms as a Proxy for CI Auditory Stimuli: The paper successfully demonstrates the utility of electrodograms generated by the ACE strategy as a realistic representation of the auditory input for CI users, enabling the quantification of their SID capabilities.
  • The "Information Bottleneck" of CIs: The study strongly suggests that the limited number of stimulation channels in CIs (typically 8-22) severely restricts the spectral information available, which is crucial for distinguishing speakers. This is a fundamental limitation that current signal processing strategies struggle to overcome.
  • Quiet vs. Noisy Performance Discrepancy: The stark difference in performance between quiet and noisy conditions highlights the vulnerability of CI-based speaker identification to environmental noise, a common challenge for CI users in everyday listening.
  • Text-Dependent vs. Text-Independent Performance: CI users perform better when identifying speakers from a fixed set of utterances (text-dependent) compared to identifying speakers from arbitrary speech (text-independent). This indicates that while some speaker-specific cues might be preserved in limited contexts, they are insufficient for robust identification in more general scenarios.

Key Results

  • Quiet Conditions: In quiet environments, CI users demonstrated a reasonable ability to identify speakers. The highest accuracies achieved were 95% for the text-dependent database using GMM-UBM and 99% using i-Vector-PLDA. For text-independent scenarios, accuracies ranged from 84.74% (50 speakers, i-Vector-PLDA) down to 58.47% (50 speakers, GMM-UBM).
  • Noisy Conditions: Performance drastically deteriorated in noisy conditions (10dB AWGN/SSN). Accuracies dropped to as low as 2.00% for GMM-UBM and 5.20% for i-Vector-PLDA when tested with 50 speakers. This indicates that CI-based auditory stimuli are largely "confused and unable to predict speakers under noisy conditions."
  • Classifier Performance: The i-Vector-PLDA classifier generally outperformed the GMM-UBM classifier, especially in text-independent and noisy conditions, suggesting its greater capacity to model the limited information available.

Limitations and Future Work

  • Limited Number of Speakers: While the study tested up to 50 speakers, real-world scenarios might involve a larger pool of potential speakers, which could further challenge CI users.
  • Simulated Stimuli: The study relies on simulated electrodograms. While ACE is a representative strategy, individual CI users have unique device configurations and auditory processing capabilities that might influence results.
  • Future Work: The authors suggest a parallel investigation using subjective studies to validate their findings. They also propose exploring improvements to signal processing strategies within CI processors to enhance speaker characterization for CI listeners, ultimately aiming to improve their quality of life. Further research could also focus on developing more robust feature extraction methods or exploring multi-modal approaches (e.g., combining auditory cues with visual information) to aid speaker identification for CI users.

Summary updated

BibTeX
@inproceedings{mamun2019quantifying,
  author = {Nursadul Mamun and Ria Ghosh and John H.L. Hansen},
  title = {{Quantifying Cochlear Implant Users’ Ability for Speaker Identification Using CI Auditory Stimuli}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure