VQ/RVQ comparison at matched bitrate

[awais-de]

Parent: none -- standalone extension (Thesis Goal 2)

Key finding

Pending evaluation

Hypothesis

If the R-D gap vs EnCodec is a scalar-quantizer artifact, VQ/RVQ at matched bitrate should close it. If the gap persists, the ceiling finding generalises across quantizer type.

Setup

Parameter	This variant	Baseline (Phase G)
Quantizer	VQ/RVQ (TBD)	3-bit SQ + zlib
Target bitrate	~5.9 kbps	5.89 kbps

Execution checklist

• Select VQ/RVQ implementation + codebook size
• Train, checkpoint, evaluate on n=40 (same protocol as Phase G)

Results

Metric	This variant	32-dim SQ baseline
kbps	TBD	5.89
PESQ-WB	TBD	1.258
STOI	TBD	0.792

Artifacts

TBD

Open

• Decide VQ vs RVQ + codebook size before starting

Baseline corrected 2026-08-10 (n=40 genuine — see #10).

[awais-de]

Deferred — out of scope for the current thesis timeline. The VQ vs. split-VQ design decision and codebook sizing remain open; revisit if the scope reopens.