RAVE uses Pseudo Quadrature Mirror Filters to split audio into sub-bands before encoding

Before encoding, RAVE decomposes the waveform into N sub-bands using a PQMF (Pseudo Quadrature Mirror Filter) bank. The encoder operates on these sub-bands rather than the raw waveform, which reduces temporal resolution requirements while preserving spectral detail. A prototype lowpass filter is designed using the Kaiser window method, then modulated into a cosine-modulated filter bank. During decoding, the inverse PQMF reconstructs the full-band signal from the decoded sub-bands. A polyphase implementation (requires n_band to be a power of 2) runs faster than the naive convolution approach. In streaming mode, a CachedPQMF variant uses cached convolutions for sample-by-sample inference.

Examples

The PQMF is applied before the encoder (raw waveform → PQMF → encoder) and inverted after the decoder (decoder output → inverse PQMF → waveform). The v2_nopqmf config removes this stage for bending purposes.

Assessment

Why does RAVE’s polyphase PQMF require n_band to be a power of 2? What tradeoff does the lazy v2_nopqmf config make?