This page summarizes how compression ratio (CR) trades off against signal- and physiology-level fidelity for the v1 RVQ goldens. The Effective CR column is populated when an entropy-prior run is available; otherwise, the codec CR is the release-facing operating point. CRs are reported alongside the noise-aware metrics so it is easy to inspect the difference between waveform faithfulness, clean-reference fidelity, and physiological metric preservation.
Data source: measurements preserved from the original documentation.
| CR | Codec CR | Effective CR | bits/tok | N | Faithful PRD% |
|---|
| 02x | 2.00 | — | — | 9788 | 2.59 |
|---|
| 04x | 4.00 | — | — | 9789 | 3.44 |
|---|
| 08x | 8.00 | — | — | 9789 | 5.10 |
|---|
| 16x | 16.00 | — | — | 9789 | 7.54 |
|---|
| 32x | 32.00 | — | — | 6336 | 12.08 |
|---|
| CR | Truth PRD% (clean) | PRDN-noise% | HR MAE (bpm) | Pulse-band PSD err | Coherence |
|---|
| 02x | 0.62 | 0.15 | 0.19 | 0.0155 | 0.9928 |
|---|
| 04x | 1.14 | 0.23 | 0.29 | 0.5885 | 0.9861 |
|---|
| 08x | 2.66 | 0.64 | 0.49 | 0.6623 | 0.9565 |
|---|
| 16x | 5.75 | 0.79 | 0.88 | 12.5676 | 0.8991 |
|---|
| 32x | 12.35 | 0.20 | 1.79 | 0.0356 | 0.8014 |
|---|
| CR | Seam ratio |
|---|
| 02x | — |
|---|
| 04x | — |
|---|
| 08x | — |
|---|
| 16x | — |
|---|
| 32x | — |
|---|
Tertiles are formed from a band-power noise estimate over each input recording; clean is the lowest-noise third, noisy is the highest.
| CR | Tertile | N | PRD% | PRDN-noise% | HR MAE |
|---|
| 02x | clean | 3263 | 4.98 | 0.46 | 0.22 |
|---|
| 02x | median | 3262 | 1.37 | 0.00 | 0.14 |
|---|
| 02x | noisy | 3263 | 1.42 | 0.00 | 0.22 |
|---|
| 04x | clean | 3263 | 6.34 | 0.69 | 0.33 |
|---|
| 04x | median | 3263 | 1.99 | 0.00 | 0.24 |
|---|
| 04x | noisy | 3263 | 1.98 | 0.00 | 0.30 |
|---|
| 08x | clean | 3263 | 8.60 | 1.92 | 0.52 |
|---|
| 08x | median | 3263 | 3.18 | 0.00 | 0.47 |
|---|
| 08x | noisy | 3263 | 3.51 | 0.00 | 0.49 |
|---|
| 16x | clean | 3263 | 10.36 | 2.32 | 0.75 |
|---|
| 16x | median | 3263 | 5.77 | 0.01 | 0.86 |
|---|
| 16x | noisy | 3263 | 6.49 | 0.04 | 1.04 |
|---|
| 32x | clean | 2112 | 11.17 | 0.59 | 1.54 |
|---|
| 32x | median | 2112 | 12.16 | 0.01 | 1.72 |
|---|
| 32x | noisy | 2112 | 12.90 | 0.00 | 2.11 |
|---|
| CR | Tertile | Pulse-band PSD err | Coherence |
|---|
| 02x | clean | — | 0.9862 |
|---|
| 02x | median | — | 0.9963 |
|---|
| 02x | noisy | — | 0.9959 |
|---|
| 04x | clean | — | 0.9784 |
|---|
| 04x | median | — | 0.9906 |
|---|
| 04x | noisy | — | 0.9894 |
|---|
| 08x | clean | — | 0.9449 |
|---|
| 08x | median | — | 0.9655 |
|---|
| 08x | noisy | — | 0.9592 |
|---|
| 16x | clean | — | 0.8928 |
|---|
| 16x | median | — | 0.9123 |
|---|
| 16x | noisy | — | 0.8922 |
|---|
| 32x | clean | — | 0.8367 |
|---|
| 32x | median | — | 0.8015 |
|---|
| 32x | noisy | — | 0.7660 |
|---|
- Faithful PRD compares reconstruction with the recorded input; Truth PRD uses the fixture’s clean reference. Their difference depends on the input noise and reference construction.
- PRDN-noise is a noise-normalized error measure. Interpret it with waveform and physiological errors, not as proof of denoising by itself.
- HR MAE and band error describe the measurements in this evaluation. Acceptance thresholds depend on your application.
- Seam ratio reports boundary energy relative to the window center. A missing value means no result was reported.
- Sample counts and noise-bucket composition can differ across ratios. Preserve that context when comparing candidates.
See the validation scorecard for definitions and customer evidence for a compact comparison.