Refining algorithmic estimation of relative fundamental frequency: Accounting for sample characteristics and fundamental frequency estimation method

• Main Authors: Buckley, D.P (Author), Kolin, K.R (Author), Noordzij, J.P (Author), Segina, R.K (Author), Stepp, C.E (Author), Tardif, M.C (Author), Vojtech, J.M (Author)
• Format: Article
• Language: English
• Published: Acoustical Society of America 2019
• Subjects: acoustics; Acoustics; algorithm; Algorithm threshold; Algorithms; Autocorrelation methods; Automation; Category specifics; Estimation techniques; Frequency estimation; Fundamental frequencies; Fundamental frequency estimation; human; Humans; Mean bias errors; Mean square error; Natural frequencies; pathophysiology; phonation; Phonation; procedures; Refining; Root mean square errors; speech analysis; Speech Production Measurement; voice; voice disorder; Voice Disorders; Voice Quality
• Online Access: View Fulltext in Publisher

 Relative fundamental frequency (RFF) is a promising acoustic measure for evaluating voice disorders. Yet, the accuracy of the current RFF algorithm varies across a broad range of vocal signals. The authors investigated how fundamental frequency (fo) estimation and sample characteristics impact the relationship between manual and semi-automated RFF estimates. Acoustic recordings were collected from 227 individuals with and 256 individuals without voice disorders. Common fo estimation techniques were compared to the autocorrelation method currently implemented in the RFF algorithm. Pitch strength-based categories were constructed using a training set (1158 samples), and algorithm thresholds were tuned to each category. RFF was then computed on an independent test set (291 samples) using category-specific thresholds and compared against manual RFF via mean bias error (MBE) and root-mean-square error (RMSE). Auditory-SWIPE′ for fo estimation led to the greatest correspondence with manual RFF and was implemented in concert with category-specific thresholds. Refining fo estimation and accounting for sample characteristics led to increased correspondence with manual RFF [MBE = 0.01 semitones (ST), RMSE = 0.28 ST] compared to the unmodified algorithm (MBE = 0.90 ST, RMSE = 0.34 ST), reducing the MBE and RMSE of semi-automated RFF estimates by 88.4% and 17.3%, respectively. © 2019 Acoustical Society of America.