Frequency-Adaptive Flow Matching for High-Fidelity Text-to-Speech Synthesis

Flow matching (FM) enables efficient acoustic modeling for text-to-speech (TTS), but fine spectral structures may remain imperfectly reconstructed. To address this issue, we propose frequency-adaptive flow matching (FAFM), which replaces the shared linear interpolation schedule with a frequency-dependent nonlinear schedule. By assigning different interpolation exponents to Mel-frequency bands, FAFM advances high-frequency components earlier along their conditional paths while preserving the endpoints of the linear baseline. The corresponding analytical target velocity modulates frequency-dependent evolution without introducing additional network modules. Evaluations across two TTS architectures and English and Chinese datasets show improvements in spectral reconstruction and several objective quality and intelligibility metrics. Results averaged over three independent training seeds, together with band-wise spectral analysis, support the effectiveness of moderate frequency modulation. Subjective listening tests show small numerical gains in perceived naturalness over standard CFM.