Research Foundations

Key Finding

MFA-SAT reached 86% accuracy on child speech (ages 3-6), the only system approaching human interrater agreement.

VoxKit Implementation

VoxKit defaults to MFA while supporting alternative engines for comparative research.

Critical Consideration

MFA-SAT was trained on adult speech; researchers should validate performance for their own use case.

Key Finding

Vowels showed 83% accuracy across systems. Fricative accuracy improved significantly with child age (OR = 1.29/year).

VoxKit Implementation

VoxKit tracks alignment metadata and speaker age, enabling age-stratified accuracy analysis.

Critical Consideration

These patterns emerged from elicited single-word productions and may not generalize to spontaneous speech.

Key Finding

Most clinical speech datasets contain only minutes to hours of audio with uncertain labels, leading to poor generalization.

VoxKit Implementation

VoxKit tracks metadata and versioned provenance for transparency.

Critical Consideration

VoxKit attempts rigorous documentation but cannot solve fundamental overfitting, only ensuring accidents happen less.

Key Finding

Clinically grounded measures outperform opaque embeddings.

VoxKit Implementation

VoxKit allows you to test and review this, as AI becomes more advanced, methodologies may advance as well.

Critical Consideration

Alignment errors can propagate downstream. Researchers must validate that phonetic boundaries are reliable.