GBNF-VAE: A Pathological Voice Enhancement Model Based on Gold Section for Bottleneck Feature With Variational...

Area of Science:

• Speech processing
• Biomedical engineering
• Signal processing

Background:

• Speech enhancement aims to improve degraded speech signals.
• Existing methods often overlook pathological speech quality issues caused by anomalous glottis flow.
• Effective enhancement requires separating high-dimensional timbre and speech features to suppress low-dimensional noise.

Purpose of the Study:

• To propose an effective enhancement model for pathological speech.
• To address the challenge of anomalous glottis flow affecting speech quality.
• To extract and combine high-dimensional timbre and semantic features for improved speech synthesis.

Main Methods:

• Proposed the GBNF-VAE model for efficient timbre extraction and reduction of airflow noise interference.
• Utilized the Golden Section method to control bottleneck features for efficient timbre characterization.
• Employed a variational autoencoder to extract semantic features, combined with timbre features for enhanced speech synthesis.

Main Results:

• The GBNF-VAE model demonstrated outstanding performance in pathological speech quality enhancement.
• Evaluations included spectrum observation, objective indicators, and subjective assessments.
• The proposed method effectively suppressed anomalous airflow noise and improved speech quality.

Conclusions:

• The GBNF-VAE model offers a significant advancement in pathological speech enhancement.
• Separating timbre and semantic features is crucial for improving speech quality in pathological cases.
• The model's efficiency and effectiveness are validated by comprehensive performance evaluations.