benchmark: final gs-2.0M WideCodec numbers (UTMOSv2 2.788, beats DAC)
Browse files
README.md
CHANGED
|
@@ -53,8 +53,8 @@ the only one here whose token stream an autoregressive LLM/TTS model can practic
|
|
| 53 |
| ground truth | — | — | — | — | 2.822 | — | 0.00244 | 8263 |
|
| 54 |
| nvidia nemo44k | 44.1k | ~86 | RVQ (many) | ~6–9 kbps | **2.903** | 0.379 | 0.00196 | 8372 |
|
| 55 |
| neucodec (24k anchor) | 24k | 50 | 1 | ~0.8 kbps | 2.848 | 0.706 | 0.00005 | 6214 |
|
|
|
|
| 56 |
| dac | 44.1k | 86 | 9 | ~8 kbps | 2.672 | **0.341** | 0.00156 | 8185 |
|
| 57 |
-
| **WideCodec (ours)** | 44.1k | **50** | **1** | **~0.8 kbps** | 2.590 ∗ | 0.655 | 0.00229 | 8546 |
|
| 58 |
| snac44k | 44.1k | multi-scale | 3–4 | ~2.6 kbps | 2.340 | 0.493 | 0.00158 | 8402 |
|
| 59 |
| encodec48k | 48k | 150 | RVQ | 24 kbps | 2.042 | 0.458 | 0.00169 | 8800 |
|
| 60 |
|
|
@@ -62,11 +62,7 @@ the only one here whose token stream an autoregressive LLM/TTS model can practic
|
|
| 62 |
num_quantizers=1` → 8·log₂4 = **16 bits/frame** × 50 frames/s = **800 bps = 0.8 kbps**
|
| 63 |
(vs DAC's 9 codebooks × 10 bits × 86 fps ≈ 8 kbps).
|
| 64 |
|
| 65 |
-
|
| 66 |
-
checkpoint (gs-1.875M, clean-finetuned) is **not yet re-scored** on this benchmark. At
|
| 67 |
-
~0.8 kbps / 1 codebook it matches DAC's naturalness class using ~10× fewer bits, and unlike
|
| 68 |
-
the 24 kHz anchor it carries real high-band content (rolloff 8.5 kHz; HF energy ≈ 94% of GT
|
| 69 |
-
vs the anchor's ~0). Full report:
|
| 70 |
[evaluation/REPORT_44k_class_codecs.md](https://github.com/Scicom-AI-Enterprise-Organization/WideCodec/blob/main/evaluation/REPORT_44k_class_codecs.md).
|
| 71 |
|
| 72 |
## Training stages / which revision to pull
|
|
|
|
| 53 |
| ground truth | — | — | — | — | 2.822 | — | 0.00244 | 8263 |
|
| 54 |
| nvidia nemo44k | 44.1k | ~86 | RVQ (many) | ~6–9 kbps | **2.903** | 0.379 | 0.00196 | 8372 |
|
| 55 |
| neucodec (24k anchor) | 24k | 50 | 1 | ~0.8 kbps | 2.848 | 0.706 | 0.00005 | 6214 |
|
| 56 |
+
| **WideCodec (ours, gs-2.0M)** | 44.1k | **50** | **1** | **~0.8 kbps** | **2.788** | 0.571 | 0.00223 | 9011 |
|
| 57 |
| dac | 44.1k | 86 | 9 | ~8 kbps | 2.672 | **0.341** | 0.00156 | 8185 |
|
|
|
|
| 58 |
| snac44k | 44.1k | multi-scale | 3–4 | ~2.6 kbps | 2.340 | 0.493 | 0.00158 | 8402 |
|
| 59 |
| encodec48k | 48k | 150 | RVQ | 24 kbps | 2.042 | 0.458 | 0.00169 | 8800 |
|
| 60 |
|
|
|
|
| 62 |
num_quantizers=1` → 8·log₂4 = **16 bits/frame** × 50 frames/s = **800 bps = 0.8 kbps**
|
| 63 |
(vs DAC's 9 codebooks × 10 bits × 86 fps ≈ 8 kbps).
|
| 64 |
|
| 65 |
+
The **WideCodec row is the final gs-2.0M checkpoint** (9,291-clip set). At ~0.8 kbps / 1 codebook it **beats DAC on naturalness** (UTMOSv2 2.79 vs 2.67) using ~10x fewer bits, and unlike the 24 kHz anchor it carries real high-band content (rolloff 9.0 kHz; HF energy ~91% of GT vs the anchor's ~0). On reference-based metrics it beats the 24 kHz anchor on all six (PESQ 2.08, ESTOI 0.79, LSD-HF 0.96 vs 2.26, ECAPA-sim 0.87). Full report:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
[evaluation/REPORT_44k_class_codecs.md](https://github.com/Scicom-AI-Enterprise-Organization/WideCodec/blob/main/evaluation/REPORT_44k_class_codecs.md).
|
| 67 |
|
| 68 |
## Training stages / which revision to pull
|