Beyond Quantization: Self-Supervised Speech Representation Learning via Masked Hilbert Envelope Prediction

Abstract

Current self-supervised speech models typically avoid predicting raw time-domain waveforms directly because fast phase shifts cause instability. To bypass this, popular methods use vector-quantized codebooks, multi-stage clustering, or teacher networks, but these approaches can suffer from memory bloat, complex pipelines, and codebook collapse. This study introduces Masked Hilbert Envelope Prediction (MHEP), a streamlined, non-quantized pre-training approach that treats masked speech modeling as a continuous regression task based on instantaneous analytic envelopes. By filtering clean speech through a frozen Constant-Q Gabor filterbank, MHEP separates vocal-tract modulation from the carrier phase, generating a stable, phase-invariant target without relying on codebooks or auxiliary teacher models. In downstream speaker verification tasks, while standard waveform regression and STFT spectrogram reconstruction cause performance drops (10.60% and 9.80% Equal Error Rates, respectively) compared to a random baseline (8.00% EER), MHEP avoids this negative transfer and improves performance to 7.40% EER. When scaled to 100 hours of LibriSpeech data, MHEP pre-training prevents optimization failure in deep 8-layer Transformers and drastically reduces the downstream verification error from 34.90% to 10.90% EER, achieving a 68.8% relative error reduction.