# NanoPitch Student Leaderboard

*Last updated: 2026-04-21*

All metrics use the **realtime Viterbi decoder** (no lookahead), matching the browser deployment.

---

## 1. RPA Leaderboards

### RPA — Clean Audio ↑

Raw Pitch Accuracy on clean (no-noise) test clips. Higher is better.

| Rank | Student | RPA Clean ↑ | RPA +0 dB | RPA -5 dB | VAD Acc | Median Err | Note |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Festus Ewakaa Kahunla | 97.1% | 92.0% | 87.9% | 79.9% | 3.4¢ | Fixed train/eval label mismatch: VAD target now derived from f0>0 (RMVPE) instead of the RMS-energy vad label, which disagrees with RMVPE on 12.1% of frames. Added cosine annealing LR (CosineAnnealingWarmRestarts, T_0=10, T_mult=2, eta_min=1e-5). Reduced Viterbi onset penalty from 2.0 to 1.0. Trained 100 epochs from scratch (50 + 50 resume). |
| 2 | Dillon Best | 97.0% | 91.5% | 93.2% | 98.5% | 5.4¢ | 150 epoch, 112 GRU, tweaked VAD RPA loss weights, started with 5 + 20 individual feature training epochs, cosine annealing, 0.6 voiced/unvoiced weighting, augmentation + clean probability |
| 3 | Brady Chase | 96.8% | 91.4% | 91.5% | 80.0% | 5.5¢ | Codex updates when told to focus on changes in read me and with extra emphasiss on how VDR is being impacted so try for constant improvement(gru_size=96, cond_size=64, lr=0.0001). |
| 4 | Ilysia Krzywonos | 96.8% | 90.9% | 92.1% | 97.9% | 5.3¢ | Adjusted the weight for pitch/VAD loss, implemented noise augmentation and learning rate scheduler |
| 5 | Uddhav Jain | 96.7% | 91.9% | 91.6% | 98.6% | 5.5¢ | VDR-focused GRU-128 run (submitted from expD_best.pth). |
| 6 | Charis NoiseAugBaseline | 96.6% | 88.8% | 87.8% | 97.3% | 12.1¢ | Baseline run with default hyperparameters, and the baseline noise augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 7 | Dillon Positive Weight | 95.4% | 89.1% | 91.3% | 97.7% | 10.6¢ | Applied recommended augmentation w/ 25% chance of clean output and applied 0.6x weight to stop the model from over-favoring voiced predictions. |
| 8 | Charis Test | 94.8% | 87.3% | 88.0% | 98.3% | 6.5¢ | Baseline run with default hyperparameters (gru_size=96, cond_size=64, lr=1e-3). |
| 9 | Rajat Sharma | 93.8% | 88.6% | 88.6% | 97.0% | 17.2¢ | Baseline run with default hyperparameters and default augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 10 | Dabin Seomun | 92.8% | 87.3% | 88.6% | 97.3% | 12.4¢ | Noise augmentation using a normal distribution for SNR sampling (mean=10, std=5) & Ran 100 epochs |
| 11 | Stefan Snyder - Baseline | 91.1% | 85.4% | 81.0% | 98.4% | 13.0¢ | Baseline run with default hyperparameters, and the baseline noise augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 12 | Kim Huy Heng | 90.8% | 87.0% | 89.2% | 97.9% | 15.8¢ | Implemented a staged curriculum training pipeline |
| 13 | Charis - Noise Augmentation v2 | 90.1% | 86.3% | 88.4% | 95.8% | 18.0¢ | Baseline run with default hyperparameters, with the basic noise augmentation and clean signal 10% of the time (gru_size=96, cond_size=64, lr=1e-3). |
| 14 | Chris Rui Zhao | 90.0% | 82.9% | 83.4% | 93.0% | 12.6¢ | Added augmentations, specaugment, cosine annealing, and BCEWithLogitsLoss with Claude Opus 4.7's suggestions. |

### RPA — Macro Average (all SNR conditions) ↑

Mean RPA across all 6 SNR conditions (clean, −5 dB, 0 dB, +5 dB, +10 dB, +20 dB). Higher is better.

| Rank | Student | RPA Macro Avg ↑ | RPA Clean | RPA +0 dB | RPA -5 dB | Note |
| --- | --- | --- | --- | --- | --- | --- |
| 1 | Dillon Best | 94.1% | 97.0% | 91.5% | 93.2% | 150 epoch, 112 GRU, tweaked VAD RPA loss weights, started with 5 + 20 individual feature training epochs, cosine annealing, 0.6 voiced/unvoiced weighting, augmentation + clean probability |
| 2 | Ilysia Krzywonos | 93.9% | 96.8% | 90.9% | 92.1% | Adjusted the weight for pitch/VAD loss, implemented noise augmentation and learning rate scheduler |
| 3 | Uddhav Jain | 93.7% | 96.7% | 91.9% | 91.6% | VDR-focused GRU-128 run (submitted from expD_best.pth). |
| 4 | Brady Chase | 93.4% | 96.8% | 91.4% | 91.5% | Codex updates when told to focus on changes in read me and with extra emphasiss on how VDR is being impacted so try for constant improvement(gru_size=96, cond_size=64, lr=0.0001). |
| 5 | Dillon Positive Weight | 92.2% | 95.4% | 89.1% | 91.3% | Applied recommended augmentation w/ 25% chance of clean output and applied 0.6x weight to stop the model from over-favoring voiced predictions. |
| 6 | Charis NoiseAugBaseline | 91.9% | 96.6% | 88.8% | 87.8% | Baseline run with default hyperparameters, and the baseline noise augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 7 | Rajat Sharma | 91.9% | 93.8% | 88.6% | 88.6% | Baseline run with default hyperparameters and default augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 8 | Festus Ewakaa Kahunla | 91.7% | 97.1% | 92.0% | 87.9% | Fixed train/eval label mismatch: VAD target now derived from f0>0 (RMVPE) instead of the RMS-energy vad label, which disagrees with RMVPE on 12.1% of frames. Added cosine annealing LR (CosineAnnealingWarmRestarts, T_0=10, T_mult=2, eta_min=1e-5). Reduced Viterbi onset penalty from 2.0 to 1.0. Trained 100 epochs from scratch (50 + 50 resume). |
| 9 | Dabin Seomun | 90.2% | 92.8% | 87.3% | 88.6% | Noise augmentation using a normal distribution for SNR sampling (mean=10, std=5) & Ran 100 epochs |
| 10 | Charis Test | 90.2% | 94.8% | 87.3% | 88.0% | Baseline run with default hyperparameters (gru_size=96, cond_size=64, lr=1e-3). |
| 11 | Kim Huy Heng | 89.5% | 90.8% | 87.0% | 89.2% | Implemented a staged curriculum training pipeline |
| 12 | Charis - Noise Augmentation v2 | 88.0% | 90.1% | 86.3% | 88.4% | Baseline run with default hyperparameters, with the basic noise augmentation and clean signal 10% of the time (gru_size=96, cond_size=64, lr=1e-3). |
| 13 | Stefan Snyder - Baseline | 86.3% | 91.1% | 85.4% | 81.0% | Baseline run with default hyperparameters, and the baseline noise augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 14 | Chris Rui Zhao | 86.1% | 90.0% | 82.9% | 83.4% | Added augmentations, specaugment, cosine annealing, and BCEWithLogitsLoss with Claude Opus 4.7's suggestions. |

---

## 2. Gross Error Rate Leaderboards

### Gross Error Rate — Clean Audio ↓

Fraction of voiced frames with pitch error > 50 cents on clean audio. Lower is better.

| Rank | Student | Gross Err Clean ↓ | GER +0 dB | GER -5 dB | Note |
| --- | --- | --- | --- | --- | --- |
| 1 | Festus Ewakaa Kahunla | 2.9% | 8.0% | 12.1% | Fixed train/eval label mismatch: VAD target now derived from f0>0 (RMVPE) instead of the RMS-energy vad label, which disagrees with RMVPE on 12.1% of frames. Added cosine annealing LR (CosineAnnealingWarmRestarts, T_0=10, T_mult=2, eta_min=1e-5). Reduced Viterbi onset penalty from 2.0 to 1.0. Trained 100 epochs from scratch (50 + 50 resume). |
| 2 | Dillon Best | 3.0% | 8.5% | 6.9% | 150 epoch, 112 GRU, tweaked VAD RPA loss weights, started with 5 + 20 individual feature training epochs, cosine annealing, 0.6 voiced/unvoiced weighting, augmentation + clean probability |
| 3 | Brady Chase | 3.2% | 8.6% | 8.5% | Codex updates when told to focus on changes in read me and with extra emphasiss on how VDR is being impacted so try for constant improvement(gru_size=96, cond_size=64, lr=0.0001). |
| 4 | Ilysia Krzywonos | 3.2% | 9.1% | 7.9% | Adjusted the weight for pitch/VAD loss, implemented noise augmentation and learning rate scheduler |
| 5 | Uddhav Jain | 3.3% | 8.1% | 8.4% | VDR-focused GRU-128 run (submitted from expD_best.pth). |
| 6 | Charis NoiseAugBaseline | 3.4% | 11.2% | 12.2% | Baseline run with default hyperparameters, and the baseline noise augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 7 | Dillon Positive Weight | 4.6% | 10.9% | 8.7% | Applied recommended augmentation w/ 25% chance of clean output and applied 0.6x weight to stop the model from over-favoring voiced predictions. |
| 8 | Charis Test | 5.2% | 12.7% | 12.0% | Baseline run with default hyperparameters (gru_size=96, cond_size=64, lr=1e-3). |
| 9 | Rajat Sharma | 6.2% | 11.4% | 11.3% | Baseline run with default hyperparameters and default augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 10 | Dabin Seomun | 7.2% | 12.7% | 11.4% | Noise augmentation using a normal distribution for SNR sampling (mean=10, std=5) & Ran 100 epochs |
| 11 | Stefan Snyder - Baseline | 8.9% | 14.6% | 19.0% | Baseline run with default hyperparameters, and the baseline noise augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 12 | Kim Huy Heng | 9.2% | 13.0% | 10.8% | Implemented a staged curriculum training pipeline |
| 13 | Charis - Noise Augmentation v2 | 9.9% | 13.7% | 11.6% | Baseline run with default hyperparameters, with the basic noise augmentation and clean signal 10% of the time (gru_size=96, cond_size=64, lr=1e-3). |
| 14 | Chris Rui Zhao | 10.1% | 17.1% | 16.7% | Added augmentations, specaugment, cosine annealing, and BCEWithLogitsLoss with Claude Opus 4.7's suggestions. |

### Gross Error Rate — Macro Average (all SNR conditions) ↓

Mean gross error rate across all 6 SNR conditions. Lower is better.

| Rank | Student | Gross Err Macro Avg ↓ | GER Clean | GER +0 dB | GER -5 dB | Note |
| --- | --- | --- | --- | --- | --- | --- |
| 1 | Dillon Best | 5.9% | 3.0% | 8.5% | 6.9% | 150 epoch, 112 GRU, tweaked VAD RPA loss weights, started with 5 + 20 individual feature training epochs, cosine annealing, 0.6 voiced/unvoiced weighting, augmentation + clean probability |
| 2 | Ilysia Krzywonos | 6.1% | 3.2% | 9.1% | 7.9% | Adjusted the weight for pitch/VAD loss, implemented noise augmentation and learning rate scheduler |
| 3 | Uddhav Jain | 6.3% | 3.3% | 8.1% | 8.4% | VDR-focused GRU-128 run (submitted from expD_best.pth). |
| 4 | Brady Chase | 6.6% | 3.2% | 8.6% | 8.5% | Codex updates when told to focus on changes in read me and with extra emphasiss on how VDR is being impacted so try for constant improvement(gru_size=96, cond_size=64, lr=0.0001). |
| 5 | Dillon Positive Weight | 7.8% | 4.6% | 10.9% | 8.7% | Applied recommended augmentation w/ 25% chance of clean output and applied 0.6x weight to stop the model from over-favoring voiced predictions. |
| 6 | Charis NoiseAugBaseline | 8.1% | 3.4% | 11.2% | 12.2% | Baseline run with default hyperparameters, and the baseline noise augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 7 | Rajat Sharma | 8.1% | 6.2% | 11.4% | 11.3% | Baseline run with default hyperparameters and default augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 8 | Festus Ewakaa Kahunla | 8.3% | 2.9% | 8.0% | 12.1% | Fixed train/eval label mismatch: VAD target now derived from f0>0 (RMVPE) instead of the RMS-energy vad label, which disagrees with RMVPE on 12.1% of frames. Added cosine annealing LR (CosineAnnealingWarmRestarts, T_0=10, T_mult=2, eta_min=1e-5). Reduced Viterbi onset penalty from 2.0 to 1.0. Trained 100 epochs from scratch (50 + 50 resume). |
| 9 | Dabin Seomun | 9.8% | 7.2% | 12.7% | 11.4% | Noise augmentation using a normal distribution for SNR sampling (mean=10, std=5) & Ran 100 epochs |
| 10 | Charis Test | 9.8% | 5.2% | 12.7% | 12.0% | Baseline run with default hyperparameters (gru_size=96, cond_size=64, lr=1e-3). |
| 11 | Kim Huy Heng | 10.5% | 9.2% | 13.0% | 10.8% | Implemented a staged curriculum training pipeline |
| 12 | Charis - Noise Augmentation v2 | 12.0% | 9.9% | 13.7% | 11.6% | Baseline run with default hyperparameters, with the basic noise augmentation and clean signal 10% of the time (gru_size=96, cond_size=64, lr=1e-3). |
| 13 | Stefan Snyder - Baseline | 13.7% | 8.9% | 14.6% | 19.0% | Baseline run with default hyperparameters, and the baseline noise augmentation (gru_size=96, cond_size=64, lr=1e-3). |
| 14 | Chris Rui Zhao | 13.9% | 10.1% | 17.1% | 16.7% | Added augmentations, specaugment, cosine annealing, and BCEWithLogitsLoss with Claude Opus 4.7's suggestions. |

---

## Metrics glossary

| Metric | Description |
|--------|-------------|
| RPA | Raw Pitch Accuracy — % of voiced frames within 50 cents of ground truth (higher = better) |
| Gross Error Rate (GER) | % of voiced frames with pitch error > 50 cents (lower = better) |
| VAD Acc | Voice Activity Detection accuracy — % of frames correctly classified as voiced/unvoiced |
| Median Err | Median pitch error in cents across voiced frames (100 cents = 1 semitone) |
| Macro Avg | Mean of the metric across all 6 SNR conditions: clean, −5 dB, 0 dB, +5 dB, +10 dB, +20 dB |