For final results that made it into the paper, search for final.
VQ-VAE DP penalized val:
# ./eval_segmentation.py vqvae buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 66.42%
Recall: 75.77%
F-score: 70.79%
OS: 14.08%
R-value: 72.44%
---------------------------------------------------------------------------
Word boundaries:
Precision: 15.81%
Recall: 68.13%
F-score: 25.66%
OS: 330.90%
R-value: -194.47%
---------------------------------------------------------------------------
Merged GMM val, MFCCs, 50 components:
# ./eval_segmentation.py gmm buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 29.21%
Recall: 95.00%
F-score: 44.69%
OS: 225.22%
R-value: -94.03%
---------------------------------------------------------------------------
Phone clusters:
Purity: 21.61%
Homogeneity: 19.66%
Completeness: 17.87%
V-measure: 18.72%
---------------------------------------------------------------------------
DPDP (dur_weight=6), GMM val, MFCCs, 50 components:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 52.96%
Recall: 83.28%
F-score: 64.74%
OS: 57.26%
R-value: 44.02%
---------------------------------------------------------------------------
Phone clusters:
Purity: 23.04%
Homogeneity: 21.63%
Completeness: 24.04%
V-measure: 22.77%
---------------------------------------------------------------------------
VQ-VAE DP penalized val, dur_weight=4:
# ./eval_segmentation.py vqvae buckeye val phoneseg_dp_penalized_tune
---------------------------------------------------------------------------
Phone boundaries:
Precision: 70.54%
Recall: 71.14%
F-score: 70.84%
OS: 0.86%
R-value: 75.06%
---------------------------------------------------------------------------
Word boundaries:
Precision: 16.37%
Recall: 62.36%
F-score: 25.94%
OS: 280.91%
R-value: -154.34%
---------------------------------------------------------------------------
VQ-CPC DP penalized val:
# ./eval_segmentation.py vqcpc buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 55.72%
Recall: 74.68%
F-score: 63.82%
OS: 34.04%
R-value: 57.80%
---------------------------------------------------------------------------
Word boundaries:
Precision: 15.54%
Recall: 81.03%
F-score: 26.08%
OS: 421.37%
R-value: -266.58%
---------------------------------------------------------------------------
CPC-big normalized DP penalized val, dur_weight=200:
./eval_segmentation.py cpc_big_normalized buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 55.42%
Recall: 85.47%
F-score: 67.24%
OS: 54.23%
R-value: 47.62%
---------------------------------------------------------------------------
Clusters:
Phone purity: 36.86%
---------------------------------------------------------------------------
Word boundaries:
Precision: 14.91%
Recall: 87.52%
F-score: 25.48%
OS: 486.97%
R-value: -320.15%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 1.90%
Recall: 8.47%
F-score: 3.11%
OS: 345.42%
---------------------------------------------------------------------------
CPC-big DP penalized val, Gamma:
# ./utils/eval_segmentation.py cpc_big buckeye val phoneseg_dp_penalized_gamma
---------------------------------------------------------------------------
Phone boundaries:
Precision: 62.81%
Recall: 76.87%
F-score: 69.13%
OS: 22.39%
R-value: 67.81%
---------------------------------------------------------------------------
Clusters:
Phone purity: 40.07%
Phone homogeneity: 40.95%
Phone completeness: 38.84%
Phone V-measure: 39.87%
---------------------------------------------------------------------------
Word boundaries:
Precision: 17.09%
Recall: 79.75%
F-score: 28.14%
OS: 366.72%
R-value: -220.46%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 2.73%
Recall: 9.83%
F-score: 4.27%
OS: 260.07%
---------------------------------------------------------------------------
CPC-big DP penalized val, dur_weight=2:
./eval_segmentation.py cpc_big buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 51.04%
Recall: 88.75%
F-score: 64.81%
OS: 73.87%
R-value: 32.54%
---------------------------------------------------------------------------
Phone clusters:
Purity: 36.85%
Homogeneity: 37.25%
Completeness: 35.53%
V-measure: 36.37%
---------------------------------------------------------------------------
Word boundaries:
Precision: 13.60%
Recall: 90.20%
F-score: 23.64%
OS: 563.17%
R-value: -384.20%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 1.70%
Recall: 8.47%
F-score: 2.83%
OS: 399.25%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 50
uWER many: 473.16%
Purity: 7.44%
Homogeneity: 15.91%
Completeness: 28.84%
V-measure: 20.51%
---------------------------------------------------------------------------
CPC-big DP penalized val, dur_weight=3:
./eval_segmentation.py cpc_big buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 54.59%
Recall: 85.91%
F-score: 66.76%
OS: 57.36%
R-value: 45.20%
---------------------------------------------------------------------------
Clusters:
Phone purity: 38.10%
Phone homogeneity: 38.65%
Phone completeness: 36.77%
Phone V-measure: 37.69%
---------------------------------------------------------------------------
Word boundaries:
Precision: 14.62%
Recall: 87.73%
F-score: 25.06%
OS: 500.14%
R-value: -331.31%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 1.94%
Recall: 8.82%
F-score: 3.18%
OS: 354.59%
---------------------------------------------------------------------------
CPC-big DP penalized val, dur_weight=5:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 60.18%
Recall: 79.95%
F-score: 68.67%
OS: 32.86%
R-value: 62.05%
---------------------------------------------------------------------------
Word boundaries:
Precision: 16.04%
Recall: 81.27%
F-score: 26.79%
OS: 406.79%
R-value: -254.06%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 2.34%
Recall: 9.08%
F-score: 3.72%
OS: 288.34%
---------------------------------------------------------------------------
CPC-big DP penalized val, dur_weight=10:
./eval_segmentation.py cpc_big buckeye val phoneseg_dp_penalized_tune
---------------------------------------------------------------------------
Phone boundaries:
Precision: 67.80%
Recall: 66.58%
F-score: 67.19%
OS: -1.80%
R-value: 72.09%
---------------------------------------------------------------------------
Clusters:
Phone purity: 41.74%
Phone homogeneity: 42.57%
Phone completeness: 40.42%
Phone V-measure: 41.47%
---------------------------------------------------------------------------
Word boundaries:
Precision: 18.04%
Recall: 67.56%
F-score: 28.48%
OS: 274.48%
R-value: -146.70%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 3.57%
Recall: 10.51%
F-score: 5.33%
OS: 194.71%
---------------------------------------------------------------------------
CPC-big DP penalized val, dur_weight=50:
./utils/eval_segmentation.py cpc_big buckeye val dp_penalized_tune
---------------------------------------------------------------------------
Phone boundaries:
Precision: 69.48%
Recall: 36.79%
F-score: 48.11%
OS: -47.04%
R-value: 54.89%
---------------------------------------------------------------------------
Clusters:
Phone purity: 40.20%
Phone homogeneity: 38.64%
Phone completeness: 37.47%
Phone V-measure: 38.05%
---------------------------------------------------------------------------
Word boundaries:
Precision: 18.75%
Recall: 37.88%
F-score: 25.08%
OS: 102.04%
R-value: -17.77%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 6.88%
Recall: 11.86%
F-score: 8.71%
OS: 72.54%
---------------------------------------------------------------------------
HuBERT DP penalizes val, dur_weight=2:
./eval_segmentation.py hubert buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 45.52%
Recall: 93.01%
F-score: 61.13%
OS: 104.31%
R-value: 8.37%
---------------------------------------------------------------------------
Phone clusters:
Purity: 43.16%
Homogeneity: 41.14%
Completeness: 33.75%
V-measure: 37.08%
---------------------------------------------------------------------------
HuBERT DP penalizes val, dur_weight=3:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 47.15%
Recall: 92.15%
F-score: 62.39%
OS: 95.42%
R-value: 15.62%
---------------------------------------------------------------------------
Phone clusters:
Purity: 43.77%
Homogeneity: 41.80%
Completeness: 34.31%
V-measure: 37.69%
---------------------------------------------------------------------------
HuBERT DP penalizes val, dur_weight=5, select:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 49.95%
Recall: 90.27%
F-score: 64.32%
OS: 80.72%
R-value: 27.37%
---------------------------------------------------------------------------
Phone clusters:
Purity: 44.81%
Homogeneity: 42.92%
Completeness: 35.22%
V-measure: 38.69%
---------------------------------------------------------------------------
Word boundaries:
Precision: 13.48%
Recall: 92.56%
F-score: 23.53%
OS: 586.82%
R-value: -403.54%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 1.87%
Recall: 9.65%
F-score: 3.14%
OS: 414.70%
---------------------------------------------------------------------------
HuBERT DP penalizes val, dur_weight=12:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 56.97%
Recall: 82.94%
F-score: 67.55%
OS: 45.58%
R-value: 53.52%
---------------------------------------------------------------------------
Phone clusters:
Purity: 47.00%
Homogeneity: 45.46%
Completeness: 37.24%
V-measure: 40.94%
---------------------------------------------------------------------------
Word boundaries:
Precision: 15.78%
Recall: 87.29%
F-score: 26.73%
OS: 453.01%
R-value: -291.25%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 2.64%
Recall: 11.11%
F-score: 4.27%
OS: 320.15%
---------------------------------------------------------------------------
HuBERT DP penalizes val, dur_weight=20:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 61.79%
Recall: 73.81%
F-score: 67.26%
OS: 19.45%
R-value: 67.55%
---------------------------------------------------------------------------
Phone clusters:
Purity: 48.65%
Homogeneity: 47.02%
Completeness: 38.42%
V-measure: 42.29%
---------------------------------------------------------------------------
Word boundaries:
Precision: 17.69%
Recall: 80.25%
F-score: 28.99%
OS: 353.65%
R-value: -209.12%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 3.56%
Recall: 12.46%
F-score: 5.54%
OS: 249.94%
---------------------------------------------------------------------------
HuBERT DP penalizes val, dur_weight=25:
Phone boundaries:
Precision: 63.96%
Recall: 68.02%
F-score: 65.93%
OS: 6.34%
R-value: 70.15%
---------------------------------------------------------------------------
Phone clusters:
Purity: 49.38%
Homogeneity: 47.73%
Completeness: 38.92%
V-measure: 42.87%
---------------------------------------------------------------------------
Word boundaries:
Precision: 18.64%
Recall: 75.27%
F-score: 29.88%
OS: 303.84%
R-value: -168.59%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 4.14%
Recall: 13.04%
F-score: 6.29%
OS: 214.75%
---------------------------------------------------------------------------
HuBERT DP penalizes val, dur_weight=30:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 65.78%
Recall: 62.40%
F-score: 64.04%
OS: -5.13%
R-value: 69.55%
---------------------------------------------------------------------------
Phone clusters:
Purity: 49.87%
Homogeneity: 48.12%
Completeness: 39.19%
V-measure: 43.20%
---------------------------------------------------------------------------
Word boundaries:
Precision: 19.35%
Recall: 69.71%
F-score: 30.29%
OS: 260.25%
R-value: -133.73%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 4.67%
Recall: 13.25%
F-score: 6.90%
OS: 183.95%
---------------------------------------------------------------------------
XLSR DP penalized val, dur_weight=2500, select:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 49.32%
Recall: 84.54%
F-score: 62.30%
OS: 71.40%
R-value: 32.76%
---------------------------------------------------------------------------
Phone clusters:
Purity: 30.37%
Homogeneity: 31.30%
Completeness: 28.32%
V-measure: 29.74%
---------------------------------------------------------------------------
Word boundaries:
Precision: 13.18%
Recall: 85.72%
F-score: 22.85%
OS: 550.30%
R-value: -374.85%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 2.33%
Recall: 11.38%
F-score: 3.87%
OS: 388.49%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 50
uWER many: 463.96%
Purity: 6.40%
Homogeneity: 14.15%
Completeness: 24.51%
V-measure: 17.94%
---------------------------------------------------------------------------
ResDAVENet-VQ merged val:
./eval_segmentation.py resdavenet_vq buckeye val phoneseg_merge
---------------------------------------------------------------------------
Phone boundaries:
Precision: 30.99%
Recall: 98.06%
F-score: 47.10%
OS: 216.40%
R-value: -85.40%
---------------------------------------------------------------------------
Clusters:
Phone purity: 34.65%
---------------------------------------------------------------------------
Word boundaries:
Precision: 9.79%
Recall: 95.72%
F-score: 17.77%
OS: 877.57%
R-value: -650.57%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 0.64%
Recall: 5.95%
F-score: 1.16%
OS: 829.73%
---------------------------------------------------------------------------
ResDAVENet-VQ DP penalized val, dur_weight=50:
./eval_segmentation.py resdavenet_vq buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 45.52%
Recall: 86.07%
F-score: 59.55%
OS: 89.07%
R-value: 18.51%
---------------------------------------------------------------------------
Clusters:
Phone purity: 35.50%
---------------------------------------------------------------------------
Word boundaries:
Precision: 13.67%
Recall: 79.84%
F-score: 23.35%
OS: 483.95%
R-value: -320.41%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 1.55%
Recall: 8.77%
F-score: 2.63%
OS: 467.21%
---------------------------------------------------------------------------
ResDAVENet-VQ DP penalized val, dur_weight=200:
./eval_segmentation.py resdavenet_vq buckeye val phoneseg_dp_penalized_tune
---------------------------------------------------------------------------
Phone boundaries:
Precision: 59.44%
Recall: 58.29%
F-score: 58.86%
OS: -1.94%
R-value: 65.06%
---------------------------------------------------------------------------
Clusters:
Phone purity: 39.61%
---------------------------------------------------------------------------
Word boundaries:
Precision: 18.53%
Recall: 56.11%
F-score: 27.86%
OS: 202.78%
R-value: -90.95%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 3.44%
Recall: 10.61%
F-score: 5.20%
OS: 208.26%
---------------------------------------------------------------------------
ResDAVENet-VQ clustered codebook DP penalized val, dur_weight=50:
./eval_segmentation.py resdavenet_vq_clust50 buckeye val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 51.07%
Recall: 76.11%
F-score: 61.13%
OS: 49.02%
R-value: 46.95%
---------------------------------------------------------------------------
Clusters:
Phone purity: 27.63%
---------------------------------------------------------------------------
Word boundaries:
Precision: 15.64%
Recall: 71.97%
F-score: 25.69%
OS: 360.22%
R-value: -217.92%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 2.09%
Recall: 9.48%
F-score: 3.43%
OS: 353.26%
---------------------------------------------------------------------------
TP word segmentation on VQ-VAE DP penalized val:
# ./eval_segmentation.py vqvae buckeye val wordseg_tp_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 72.27%
Recall: 28.67%
F-score: 41.06%
OS: -60.33%
R-value: 49.40%
---------------------------------------------------------------------------
Word boundaries:
Precision: 18.07%
Recall: 27.07%
F-score: 21.67%
OS: 49.83%
R-value: 12.43%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 7.49%
Recall: 10.12%
F-score: 8.61%
---------------------------------------------------------------------------
TP word segmentation on VQ-VAE DP penalized val, Gamma prior, dur_weight=15:
# ./eval_segmentation.py vqvae buckeye val wordseg_tp_dp_penalized_gamma
---------------------------------------------------------------------------
Phone boundaries:
Precision: 78.50%
Recall: 10.86%
F-score: 19.08%
OS: -86.17%
R-value: 36.96%
---------------------------------------------------------------------------
Word boundaries:
Precision: 19.15%
Recall: 10.00%
F-score: 13.14%
OS: -47.78%
R-value: 34.12%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 16.86%
Recall: 11.16%
F-score: 13.43%
OS: -33.76%
---------------------------------------------------------------------------
AG word segmentation on VQ-VAE DP penalized val:
# ./eval_segmentation.py vqvae buckeye val wordseg_ag_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 69.29%
Recall: 63.37%
F-score: 66.20%
OS: -8.54%
R-value: 71.27%
---------------------------------------------------------------------------
Word boundaries:
Precision: 16.49%
Recall: 56.96%
F-score: 25.57%
OS: 245.45%
R-value: -126.59%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 3.63%
Recall: 9.92%
F-score: 5.31%
---------------------------------------------------------------------------
AG word segmentation on MFCC val, final:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 56.14%
Recall: 58.49%
F-score: 57.29%
OS: 4.18%
R-value: 62.98%
---------------------------------------------------------------------------
Phone clusters:
Purity: 29.48%
Homogeneity: 35.36%
Completeness: 16.81%
V-measure: 22.79%
---------------------------------------------------------------------------
Word boundaries:
Precision: 16.69%
Recall: 57.85%
F-score: 25.90%
OS: 246.70%
R-value: -127.26%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 3.10%
Recall: 9.96%
F-score: 4.73%
OS: 221.49%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 7593
uWER many: 291.89%
Purity: 11.43%
Homogeneity: 40.40%
Completeness: 36.01%
V-measure: 38.08%
---------------------------------------------------------------------------
AG word segmentation on MFCC+GMM test, final:
./eval_segmentation.py gmm buckeye test wordseg_ag_merge
---------------------------------------------------------------------------
Phone boundaries:
Precision: 54.10%
Recall: 57.55%
F-score: 55.77%
OS: 6.38%
R-value: 61.27%
---------------------------------------------------------------------------
Phone clusters:
Purity: 29.49%
Homogeneity: 35.06%
Completeness: 16.69%
V-measure: 22.61%
---------------------------------------------------------------------------
Word boundaries:
Precision: 15.95%
Recall: 57.67%
F-score: 24.99%
OS: 261.53%
R-value: -139.89%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 2.83%
Recall: 9.40%
F-score: 4.35%
OS: 231.96%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 8261
uWER many: 302.12%
Purity: 11.29%
Homogeneity: 39.04%
Completeness: 35.01%
V-measure: 36.91%
---------------------------------------------------------------------------
AG word segmentation on DPDP CPC-big val, final:
./eval_segmentation.py cpc_big buckeye val wordseg_ag_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.41%
Recall: 52.91%
F-score: 62.86%
OS: -31.65%
R-value: 66.17%
---------------------------------------------------------------------------
Phone clusters:
Purity: 49.27%
Homogeneity: 51.99%
Completeness: 29.83%
V-measure: 37.91%
---------------------------------------------------------------------------
Word boundaries:
Precision: 21.72%
Recall: 56.61%
F-score: 31.39%
OS: 160.66%
R-value: -55.35%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 6.36%
Recall: 13.61%
F-score: 8.67%
OS: 114.08%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 2558
uWER many: 184.73%
Purity: 15.43%
Homogeneity: 39.31%
Completeness: 42.58%
V-measure: 40.88%
---------------------------------------------------------------------------
AG word segmentation on VQ-VAE DP penalized val, Gamma prior, dur_weight=15:
# ./eval_segmentation.py vqvae buckeye val wordseg_ag_dp_penalized_gamma
---------------------------------------------------------------------------
Phone boundaries:
Precision: 78.03%
Recall: 27.07%
F-score: 40.19%
OS: -65.31%
R-value: 48.35%
---------------------------------------------------------------------------
Word boundaries:
Precision: 17.29%
Recall: 22.63%
F-score: 19.60%
OS: 30.93%
R-value: 20.05%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 9.34%
Recall: 11.39%
F-score: 10.27%
OS: 21.86%
---------------------------------------------------------------------------
AG word segmentation on CPC-big DP penalized val, dur_weight=3:
# ./utils/eval_segmentation.py cpc_big buckeye val wordseg_ag_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 78.76%
Recall: 51.77%
F-score: 62.48%
OS: -34.27%
R-value: 65.48%
---------------------------------------------------------------------------
Word boundaries:
Precision: 22.07%
Recall: 55.32%
F-score: 31.55%
OS: 150.63%
R-value: -47.61%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 6.85%
Recall: 14.17%
F-score: 9.23%
OS: 106.97%
---------------------------------------------------------------------------
SegAE-RNN Gamma word segmentation on CPC-big DP penalized val, dur_weight=3:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 76.82%
Recall: 23.03%
F-score: 35.44%
OS: -70.02%
R-value: 45.52%
---------------------------------------------------------------------------
Phone clusters:
Purity: 69.88%
Homogeneity: 78.74%
Completeness: 28.07%
V-measure: 41.39%
---------------------------------------------------------------------------
Word boundaries:
Precision: 33.75%
Recall: 38.55%
F-score: 35.99%
OS: 14.22%
R-value: 41.71%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 22.14%
Recall: 24.43%
F-score: 23.23%
OS: 10.33%
---------------------------------------------------------------------------
Word clusters:
Purity: 59.36%
Homogeneity: 83.10%
Completeness: 57.88%
V-measure: 68.24%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=4) word segmentation on CPC-big DP penalized val (dur_weight=5):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 78.63%
Recall: 16.80%
F-score: 27.69%
OS: -78.63%
R-value: 41.15%
---------------------------------------------------------------------------
Phone clusters:
Purity: 79.61%
Homogeneity: 86.95%
Completeness: 29.56%
V-measure: 44.13%
---------------------------------------------------------------------------
Word boundaries:
Precision: 35.36%
Recall: 28.79%
F-score: 31.74%
OS: -18.59%
R-value: 44.60%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 22.99%
Recall: 20.02%
F-score: 21.40%
OS: -12.92%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 30835
uWER many: 49.58%
Purity: 74.44%
Homogeneity: 90.97%
Completeness: 60.98%
V-measure: 73.02%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=5) word segmentation on CPC-big DP penalized val (dur_weight=3):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 76.40%
Recall: 15.87%
F-score: 26.28%
OS: -79.23%
R-value: 40.48%
---------------------------------------------------------------------------
Phone clusters:
Purity: 85.20%
Homogeneity: 90.90%
Completeness: 29.98%
V-measure: 45.09%
---------------------------------------------------------------------------
Word boundaries:
Precision: 36.57%
Recall: 28.94%
F-score: 32.31%
OS: -20.87%
R-value: 45.22%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 24.15%
Recall: 20.64%
F-score: 22.26%
OS: -14.54%
---------------------------------------------------------------------------
Word clusters:
uWER many: 44.77%
Purity: 81.74%
Homogeneity: 93.95%
Completeness: 61.28%
V-measure: 74.18%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=4) word segmentation on CPC-big DP penalized val (dur_weight=3):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.17%
Recall: 16.63%
F-score: 27.36%
OS: -78.45%
R-value: 41.02%
---------------------------------------------------------------------------
Phone clusters:
Purity: 83.65%
Homogeneity: 89.62%
Completeness: 29.87%
V-measure: 44.81%
---------------------------------------------------------------------------
Word boundaries:
Precision: 36.65%
Recall: 30.09%
F-score: 33.05%
OS: -17.90%
R-value: 45.53%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 24.06%
Recall: 21.07%
F-score: 22.47%
OS: -12.43%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 34671
uWER many: 45.71%
Purity: 79.65%
Homogeneity: 92.85%
Completeness: 60.79%
V-measure: 73.48%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=4) word segmentation on CPC-big DP penalized val (dur_weight=2), training on first 10k for 5 epochs, best:
./eval_segmentation.py cpc_big buckeye val wordseg_segaernn_dp_penalized_tune
---------------------------------------------------------------------------
Phone boundaries:
Precision: 78.80%
Recall: 18.62%
F-score: 30.12%
OS: -76.37%
R-value: 42.43%
---------------------------------------------------------------------------
Phone clusters:
Purity: 83.07%
Homogeneity: 89.31%
Completeness: 29.67%
V-measure: 44.54%
---------------------------------------------------------------------------
Word boundaries:
Precision: 36.78%
Recall: 33.12%
F-score: 34.85%
OS: -9.96%
R-value: 46.07%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 24.64%
Recall: 22.96%
F-score: 23.77%
OS: -6.80%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 36308
uWER many: 45.40%
Purity: 78.45%
Homogeneity: 92.44%
Completeness: 60.16%
V-measure: 72.89%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=2) word segmentation on CPC-big DP penalized val (dur_weight=2), training on all for 5 epochs:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 75.99%
Recall: 19.96%
F-score: 31.61%
OS: -73.74%
R-value: 43.36%
---------------------------------------------------------------------------
Phone clusters:
Purity: 78.17%
Homogeneity: 85.46%
Completeness: 29.05%
V-measure: 43.36%
---------------------------------------------------------------------------
Word boundaries:
Precision: 33.80%
Recall: 33.83%
F-score: 33.81%
OS: 0.09%
R-value: 43.49%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 22.05%
Recall: 22.12%
F-score: 22.08%
OS: 0.31%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 34608
uWER many: 51.43%
Purity: 71.65%
Homogeneity: 89.23%
Completeness: 59.25%
V-measure: 71.22%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=3) word segmentation on CPC-big DP penalized val (dur_weight=2), final:
./eval_segmentation.py cpc_big buckeye val wordseg_dpdp_aernn_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 78.32%
Recall: 20.84%
F-score: 32.92%
OS: -73.39%
R-value: 43.99%
---------------------------------------------------------------------------
Phone clusters:
Purity: 77.42%
Homogeneity: 84.85%
Completeness: 29.00%
V-measure: 43.23%
---------------------------------------------------------------------------
Word boundaries:
Precision: 35.80%
Recall: 36.30%
F-score: 36.05%
OS: 1.40%
R-value: 45.13%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 23.93%
Recall: 24.23%
F-score: 24.08%
OS: 1.24%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 33336
uWER many: 51.15%
Purity: 70.36%
Homogeneity: 88.59%
Completeness: 59.09%
V-measure: 70.90%
---------------------------------------------------------------------------
Merged CPC-big test, final:
./eval_segmentation.py cpc_big buckeye test phoneseg_merge
---------------------------------------------------------------------------
Phone boundaries:
Precision: 35.55%
Recall: 93.94%
F-score: 51.59%
OS: 164.21%
R-value: -42.36%
---------------------------------------------------------------------------
Phone clusters:
Purity: 32.33%
Homogeneity: 31.21%
Completeness: 29.88%
V-measure: 30.53%
---------------------------------------------------------------------------
Word boundaries:
Precision: 9.12%
Recall: 94.49%
F-score: 16.63%
OS: 936.59%
R-value: -701.39%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 0.86%
Recall: 6.47%
F-score: 1.52%
OS: 650.28%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 50
uWER many: 721.99%
Purity: 6.76%
Homogeneity: 12.55%
Completeness: 22.73%
V-measure: 16.17%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=3) word segmentation on CPC-big DP penalized test (dur_weight=2), final:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.86%
Recall: 21.16%
F-score: 33.27%
OS: -72.82%
R-value: 44.21%
---------------------------------------------------------------------------
Phone clusters:
Purity: 75.15%
Homogeneity: 82.81%
Completeness: 28.25%
V-measure: 42.13%
---------------------------------------------------------------------------
Word boundaries:
Precision: 35.30%
Recall: 37.66%
F-score: 36.44%
OS: 6.68%
R-value: 44.25%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 24.40%
Recall: 25.59%
F-score: 24.98%
OS: 4.88%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 38981
uWER: 105.00%
uWER many: 54.13%
Purity: 67.10%
Homogeneity: 86.92%
Completeness: 58.14%
V-measure: 69.68%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=3) word segmentation on CPC-big DP penalized val (dur_weight=1):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 75.73%
Recall: 20.71%
F-score: 32.52%
OS: -72.65%
R-value: 43.88%
---------------------------------------------------------------------------
Phone clusters:
Purity: 80.85%
Homogeneity: 87.22%
Completeness: 29.12%
V-measure: 43.66%
---------------------------------------------------------------------------
Word boundaries:
Precision: 34.94%
Recall: 36.41%
F-score: 35.66%
OS: 4.22%
R-value: 44.16%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 23.21%
Recall: 23.96%
F-score: 23.58%
OS: 3.25%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 38268
uWER many: 48.26%
Purity: 75.26%
Homogeneity: 90.70%
Completeness: 58.93%
V-measure: 71.44%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=4) word segmentation on CPC-big DP penalized val (dur_weight=0.1):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 75.17%
Recall: 18.91%
F-score: 30.21%
OS: -74.85%
R-value: 42.61%
---------------------------------------------------------------------------
Phone clusters:
Purity: 91.13%
Homogeneity: 94.76%
Completeness: 30.32%
V-measure: 45.94%
---------------------------------------------------------------------------
Word boundaries:
Precision: 34.42%
Recall: 33.00%
F-score: 33.70%
OS: -4.14%
R-value: 44.21%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 22.46%
Recall: 21.86%
F-score: 22.15%
OS: -2.68%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 46416
uWER many: 40.52%
Purity: 88.69%
Homogeneity: 96.36%
Completeness: 59.84%
V-measure: 73.83%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=5) word segmentation on HuBERT DPDP val
(dur_weight=2) best:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.04%
Recall: 17.63%
F-score: 28.69%
OS: -77.12%
R-value: 41.72%
---------------------------------------------------------------------------
Phone clusters:
Purity: 89.89%
Homogeneity: 93.06%
Completeness: 30.28%
V-measure: 45.70%
---------------------------------------------------------------------------
Word boundaries:
Precision: 41.61%
Recall: 36.15%
F-score: 38.69%
OS: -13.13%
R-value: 49.47%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 27.32%
Recall: 24.81%
F-score: 26.00%
OS: -9.22%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 40927
uWER many: 40.17%
Purity: 86.88%
Homogeneity: 94.71%
Completeness: 60.41%
V-measure: 73.77%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=6) word segmentation on HuBERT DPDP val
(dur_weight=2):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.26%
Recall: 16.38%
F-score: 27.03%
OS: -78.80%
R-value: 40.84%
---------------------------------------------------------------------------
Phone clusters:
Purity: 92.04%
Homogeneity: 94.62%
Completeness: 30.61%
V-measure: 46.25%
---------------------------------------------------------------------------
Word boundaries:
Precision: 42.27%
Recall: 34.03%
F-score: 37.71%
OS: -19.51%
R-value: 49.17%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 27.46%
Recall: 23.69%
F-score: 25.44%
OS: -13.72%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 41320
uWER many: 38.65%
Purity: 90.15%
Homogeneity: 96.03%
Completeness: 60.81%
V-measure: 74.47%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=6) word segmentation on HuBERT DPDP val
(dur_weight=3) final:
./eval_segmentation.py hubert buckeye val wordseg_dpdp_aernn_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.62%
Recall: 15.85%
F-score: 26.33%
OS: -79.58%
R-value: 40.47%
---------------------------------------------------------------------------
Phone clusters:
Purity: 92.43%
Homogeneity: 94.99%
Completeness: 30.59%
V-measure: 46.28%
---------------------------------------------------------------------------
Word boundaries:
Precision: 43.00%
Recall: 33.34%
F-score: 37.56%
OS: -22.47%
R-value: 49.20%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 27.90%
Recall: 23.49%
F-score: 25.51%
OS: -15.82%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 40213
uWER many: 38.61%
Purity: 90.58%
Homogeneity: 96.44%
Completeness: 61.11%
V-measure: 74.81%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=6) word segmentation on HuBERT DPDP val
(dur_weight=5):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 78.23%
Recall: 15.79%
F-score: 26.27%
OS: -79.82%
R-value: 40.43%
---------------------------------------------------------------------------
Phone clusters:
Purity: 90.98%
Homogeneity: 93.88%
Completeness: 30.57%
V-measure: 46.12%
---------------------------------------------------------------------------
Word boundaries:
Precision: 42.50%
Recall: 32.57%
F-score: 36.88%
OS: -23.38%
R-value: 48.74%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 27.61%
Recall: 23.07%
F-score: 25.14%
OS: -16.46%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 38695
uWER many: 39.76%
Purity: 88.82%
Homogeneity: 95.57%
Completeness: 61.13%
V-measure: 74.57%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=7) word segmentation on HuBERT DPDP val
(dur_weight=3):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.82%
Recall: 14.39%
F-score: 24.28%
OS: -81.51%
R-value: 39.44%
---------------------------------------------------------------------------
Phone clusters:
Purity: 94.23%
Homogeneity: 96.33%
Completeness: 30.85%
V-measure: 46.73%
---------------------------------------------------------------------------
Word boundaries:
Precision: 43.79%
Recall: 30.73%
F-score: 36.12%
OS: -29.82%
R-value: 48.35%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 27.99%
Recall: 22.11%
F-score: 24.70%
OS: -21.01%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 39301
uWER many: 38.48%
Purity: 93.01%
Homogeneity: 97.53%
Completeness: 61.68%
V-measure: 75.57%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=6) word segmentation on HuBERT DPDP test
(dur_weight=3) final:
# ./eval_segmentation.py hubert buckeye test wordseg_dpdp_aernn_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.93%
Recall: 15.52%
F-score: 25.88%
OS: -80.09%
R-value: 40.24%
---------------------------------------------------------------------------
Phone clusters:
Purity: 92.21%
Homogeneity: 94.75%
Completeness: 30.12%
V-measure: 45.71%
---------------------------------------------------------------------------
Word boundaries:
Precision: 43.81%
Recall: 34.11%
F-score: 38.36%
OS: -22.15%
R-value: 49.78%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 29.55%
Recall: 25.04%
F-score: 27.11%
OS: -15.25%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 49728
uWER many: 37.78%
Purity: 90.54%
Homogeneity: 96.40%
Completeness: 60.27%
V-measure: 74.17%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=3) word segmentation on XLSR DPDP val (dur_weight=2500):
---------------------------------------------------------------------------
Phone boundaries:
Precision: 72.49%
Recall: 16.27%
F-score: 26.57%
OS: -77.56%
R-value: 40.75%
---------------------------------------------------------------------------
Phone clusters:
Purity: 85.83%
Homogeneity: 90.63%
Completeness: 30.04%
V-measure: 45.12%
---------------------------------------------------------------------------
Word boundaries:
Precision: 26.81%
Recall: 22.80%
F-score: 24.64%
OS: -14.94%
R-value: 38.67%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 17.59%
Recall: 15.74%
F-score: 16.61%
OS: -10.55%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 38852
uWER many: 49.85%
Purity: 81.68%
Homogeneity: 93.28%
Completeness: 60.98%
V-measure: 73.75%
---------------------------------------------------------------------------
TP, DPDP GMM val, MFCCs, 50 components:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 59.17%
Recall: 60.21%
F-score: 59.69%
OS: 1.76%
R-value: 65.40%
---------------------------------------------------------------------------
Phone clusters:
Purity: 42.35%
Homogeneity: 50.11%
Completeness: 20.96%
V-measure: 29.56%
---------------------------------------------------------------------------
Word boundaries:
Precision: 17.94%
Recall: 60.73%
F-score: 27.69%
OS: 238.59%
R-value: -119.14%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 3.80%
Recall: 11.96%
F-score: 5.77%
OS: 214.65%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 35842
uWER many: 256.14%
Purity: 26.65%
Homogeneity: 54.32%
Completeness: 42.78%
V-measure: 47.86%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=3), DPDP GMM val, MFCCs, 50 components:
./eval_segmentation.py gmm buckeye val wordseg_dpdp_aernn_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 65.03%
Recall: 29.10%
F-score: 40.21%
OS: -55.25%
R-value: 49.52%
---------------------------------------------------------------------------
Phone clusters:
Purity: 86.63%
Homogeneity: 91.64%
Completeness: 28.96%
V-measure: 44.02%
---------------------------------------------------------------------------
Word boundaries:
Precision: 23.63%
Recall: 35.15%
F-score: 28.26%
OS: 48.78%
R-value: 19.25%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 11.76%
Recall: 18.18%
F-score: 14.28%
OS: 54.59%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 70916
uWER many: 74.95%
Purity: 83.19%
Homogeneity: 94.21%
Completeness: 57.22%
V-measure: 71.20%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=6), DPDP GMM val, MFCCs, 50 components, final:
./eval_segmentation.py gmm buckeye val wordseg_dpdp_aernn_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 67.76%
Recall: 20.62%
F-score: 31.62%
OS: -69.58%
R-value: 43.75%
---------------------------------------------------------------------------
Phone clusters:
Purity: 97.15%
Homogeneity: 98.54%
Completeness: 30.56%
V-measure: 46.65%
---------------------------------------------------------------------------
Word boundaries:
Precision: 26.45%
Recall: 26.74%
F-score: 26.60%
OS: 1.11%
R-value: 37.07%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 16.15%
Recall: 18.48%
F-score: 17.24%
OS: 14.39%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 61639
uWER many: 43.75%
Purity: 96.45%
Homogeneity: 99.05%
Completeness: 59.89%
V-measure: 74.64%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=3), DPDP GMM val, MFCCs, 25 components:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 65.83%
Recall: 25.94%
F-score: 37.22%
OS: -60.60%
R-value: 47.39%
---------------------------------------------------------------------------
Phone clusters:
Purity: 77.07%
Homogeneity: 84.62%
Completeness: 27.75%
V-measure: 41.79%
---------------------------------------------------------------------------
Word boundaries:
Precision: 23.89%
Recall: 31.29%
F-score: 27.09%
OS: 30.99%
R-value: 27.06%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 12.45%
Recall: 17.38%
F-score: 14.51%
OS: 39.58%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 54535
uWER many: 71.47%
Purity: 72.20%
Homogeneity: 89.50%
Completeness: 56.52%
V-measure: 69.28%
---------------------------------------------------------------------------
DPDP AE-RNN (dur_weight=6), DPDP GMM val, MFCCs, 25 components:
./eval_segmentation.py gmm buckeye val wordseg_segaernn_dp_penalized_25
---------------------------------------------------------------------------
Phone boundaries:
Precision: 69.00%
Recall: 17.68%
F-score: 28.15%
OS: -74.38%
R-value: 41.72%
---------------------------------------------------------------------------
Phone clusters:
Purity: 94.27%
Homogeneity: 96.84%
Completeness: 30.45%
V-measure: 46.33%
---------------------------------------------------------------------------
Word boundaries:
Precision: 26.85%
Recall: 22.86%
F-score: 24.69%
OS: -14.86%
R-value: 38.70%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 16.70%
Recall: 16.85%
F-score: 16.77%
OS: 0.92%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 52243
uWER many: 42.06%
Purity: 93.20%
Homogeneity: 98.05%
Completeness: 60.43%
V-measure: 74.77%
---------------------------------------------------------------------------
K-means clustering on the CPC-big AE-RNN embeddings:
./eval_segmentation.py cpc_big buckeye val wordseg_dpdp_aernn_dp_penalized_kmeans14000
---------------------------------------------------------------------------
Phone boundaries:
Precision: 78.53%
Recall: 20.24%
F-score: 32.19%
OS: -74.22%
R-value: 43.57%
---------------------------------------------------------------------------
Phone clusters:
Purity: 54.81%
Homogeneity: 70.66%
Completeness: 26.04%
V-measure: 38.05%
---------------------------------------------------------------------------
Word boundaries:
Precision: 36.31%
Recall: 35.67%
F-score: 35.99%
OS: -1.78%
R-value: 45.70%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 24.09%
Recall: 23.84%
F-score: 23.97%
OS: -1.01%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 14000
uWER many: 72.18%
Purity: 42.74%
Homogeneity: 79.72%
Completeness: 57.32%
V-measure: 66.69%
---------------------------------------------------------------------------
K-means clustering on the HuBERT AE-RNN embeddings (dur_weight=3,
dur_weight=6):
./eval_segmentation.py hubert buckeye val wordseg_dpdp_aernn_dp_penalized_kmeans14000
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.62%
Recall: 15.85%
F-score: 26.33%
OS: -79.58%
R-value: 40.47%
---------------------------------------------------------------------------
Phone clusters:
Purity: 55.06%
Homogeneity: 71.91%
Completeness: 26.08%
V-measure: 38.28%
---------------------------------------------------------------------------
Word boundaries:
Precision: 43.00%
Recall: 33.34%
F-score: 37.56%
OS: -22.47%
R-value: 49.20%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 27.90%
Recall: 23.49%
F-score: 25.51%
OS: -15.82%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 13381
uWER many: 69.05%
Purity: 45.19%
Homogeneity: 82.10%
Completeness: 58.59%
V-measure: 68.38%
---------------------------------------------------------------------------
K-means clustering on the HuBERT averaged embeddings (dur_weight=2,
dur_weight=5):
./eval_segmentation.py hubert buckeye val wordseg_dpdp_avgembed_kmeans14000
---------------------------------------------------------------------------
Phone boundaries:
Precision: 76.85%
Recall: 18.16%
F-score: 29.38%
OS: -76.36%
R-value: 42.10%
---------------------------------------------------------------------------
Phone clusters:
Purity: 62.84%
Homogeneity: 76.23%
Completeness: 27.69%
V-measure: 40.63%
---------------------------------------------------------------------------
Word boundaries:
Precision: 41.21%
Recall: 36.99%
F-score: 38.99%
OS: -10.24%
R-value: 49.43%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 27.11%
Recall: 25.17%
F-score: 26.10%
OS: -7.17%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 14000
uWER many: 56.58%
Purity: 61.89%
Homogeneity: 87.05%
Completeness: 61.62%
V-measure: 72.16%
---------------------------------------------------------------------------
K-means clustering on the HuBERT averaged embeddings (dur_weight=6):
./eval_segmentation.py hubert buckeye val wordseg_dpdp_avgembed_kmeans14000
---------------------------------------------------------------------------
Phone boundaries:
Precision: 77.62%
Recall: 15.85%
F-score: 26.33%
OS: -79.58%
R-value: 40.47%
---------------------------------------------------------------------------
Phone clusters:
Purity: 63.72%
Homogeneity: 77.16%
Completeness: 27.82%
V-measure: 40.90%
---------------------------------------------------------------------------
Word boundaries:
Precision: 43.00%
Recall: 33.34%
F-score: 37.56%
OS: -22.47%
R-value: 49.20%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 27.90%
Recall: 23.49%
F-score: 25.51%
OS: -15.82%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 14000
uWER many: 55.21%
Purity: 63.52%
Homogeneity: 88.01%
Completeness: 62.43%
V-measure: 73.04%
---------------------------------------------------------------------------
SegAE-RNN Chorowski (dur_weight=3) word segmentation on CPC-big DP penalized val (dur_weight=2):
---------------------------------------------------------------------------
Word boundaries:
Precision: 20.59%
Recall: 19.75%
F-score: 20.16%
OS: -4.06%
R-value: 32.89%
---------------------------------------------------------------------------
Word token boundaries:
Precision: 6.34%
Recall: 6.14%
F-score: 6.24%
OS: -3.20%
---------------------------------------------------------------------------
Word clusters:
No. clusters: 15718
uWER many: 51.12%
Purity: 84.67%
Homogeneity: 93.22%
Completeness: 65.88%
V-measure: 77.20%
---------------------------------------------------------------------------
VQ-VAE DP penalized val, dur_weight=3:
# ./eval_segmentation.py vqvae buckeye_felix val phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 65.64%
Recall: 75.97%
F-score: 70.43%
OS: 15.73%
R-value: 71.58%
---------------------------------------------------------------------------
VQ-VAE DP penalized val, dur_weight=4:
# ./eval_segmentation.py vqvae buckeye_felix val phoneseg_dp_penalized_tune
---------------------------------------------------------------------------
Phone boundaries:
Precision: 69.40%
Recall: 71.23%
F-score: 70.31%
OS: 2.64%
R-value: 74.45%
---------------------------------------------------------------------------
VQ-VAE DP penalized val, Poisson prior, dur_weight=2:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 67.46%
Recall: 70.33%
F-score: 68.87%
OS: 4.27%
R-value: 73.02%
---------------------------------------------------------------------------
VQ-VAE DP penalized val, histogram prior, dur_weight=2:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 70.95%
Recall: 66.54%
F-score: 68.68%
OS: -6.22%
R-value: 73.35%
---------------------------------------------------------------------------
VQ-VAE DP penalized val, Gamma prior, dur_weight=2:
# ./eval_segmentation.py vqvae buckeye_felix val phoneseg_dp_penalized_gamma
---------------------------------------------------------------------------
Phone boundaries:
Precision: 71.70%
Recall: 65.85%
F-score: 68.65%
OS: -8.17%
R-value: 73.25%
---------------------------------------------------------------------------
VQ-VAE DP penalized test:
# ./eval_segmentation.py vqvae buckeye_felix test phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 63.80%
Recall: 77.14%
F-score: 69.83%
OS: 20.91%
R-value: 69.03%
---------------------------------------------------------------------------
VQ-VAE DP penalized N-seg. test:
# ./eval_segmentation.py --phone_tolerance 3 vqvae buckeye_felix test phoneseg_dp_penalized_n_seg
---------------------------------------------------------------------------
Phone boundaries:
Precision: 61.74%
Recall: 90.90%
F-score: 73.53%
OS: 47.24%
R-value: 56.03%
---------------------------------------------------------------------------
DPDP VQ-CPC val, dur_weight=700:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 70.55%
Recall: 69.92%
F-score: 70.23%
OS: -0.89%
R-value: 74.64%
---------------------------------------------------------------------------
Phone clusters:
Purity: 38.48%
Homogeneity: 46.17%
Completeness: 27.05%
V-measure: 34.12%
---------------------------------------------------------------------------
DPDP VQ-CPC val, dur_weight=800:
---------------------------------------------------------------------------
Phone boundaries:
Precision: 71.93%
Recall: 65.58%
F-score: 68.61%
OS: -8.83%
R-value: 73.18%
---------------------------------------------------------------------------
Phone clusters:
Purity: 38.70%
Homogeneity: 46.61%
Completeness: 27.34%
V-measure: 34.46%
---------------------------------------------------------------------------
DPDP VQ-CPC test, dur_weight=700, final:
./eval_segmentation.py --phone_tolerance 3 vqcpc buckeye_felix test phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 69.64%
Recall: 73.41%
F-score: 71.48%
OS: 5.42%
R-value: 75.12%
---------------------------------------------------------------------------
Phone clusters:
Purity: 38.83%
Homogeneity: 45.94%
Completeness: 27.01%
V-measure: 34.02%
---------------------------------------------------------------------------
DPDP CPC-big val, dur_weight=10:
./eval_segmentation.py cpc_big buckeye_felix val phoneseg_dp_penalized_tune
---------------------------------------------------------------------------
Phone boundaries:
Precision: 65.46%
Recall: 65.78%
F-score: 65.62%
OS: 0.48%
R-value: 70.62%
---------------------------------------------------------------------------
Phone clusters:
Purity: 44.36%
Homogeneity: 47.30%
Completeness: 44.13%
V-measure: 45.66%
---------------------------------------------------------------------------
DPDP CPC-big test, dur_weight=10, final:
./eval_segmentation.py --phone_tolerance 3 cpc_big buckeye_felix test phoneseg_dp_penalized
---------------------------------------------------------------------------
Phone boundaries:
Precision: 73.15%
Recall: 77.68%
F-score: 75.35%
OS: 6.19%
R-value: 78.34%
---------------------------------------------------------------------------
Phone clusters:
Purity: 45.55%
Homogeneity: 48.81%
Completeness: 45.62%
V-measure: 47.16%
---------------------------------------------------------------------------
Merged CPC-big test, final:
./eval_segmentation.py --phone_tolerance 3 cpc_big buckeye_felix test phoneseg_merge
---------------------------------------------------------------------------
Phone boundaries:
Precision: 36.89%
Recall: 97.20%
F-score: 53.48%
OS: 163.47%
R-value: -40.54%
---------------------------------------------------------------------------
Phone clusters:
Purity: 35.73%
Homogeneity: 38.38%
Completeness: 36.51%
V-measure: 37.42%
---------------------------------------------------------------------------
DP penalized CPC-big (dur_weight=2), DP penalized AE-RNN (dur_weight=3), final:
boundary_precision: 0.6616161616161617
boundary_recall: 0.7065405514187966
boundary_fscore: 0.6833407945189482
token_precision: 0.2688495201940314
type_precision: 0.25486822554156563
token_recall: 0.257577288341079
type_recall: 0.3673768459023785
token_fscore: 0.2630927196739074
type_fscore: 0.30095114969064546
words: 12787
coverage: 0.9999693444306494
ned: 0.8897823993837859
pairs: 17310
DP penalized CPC-big (dur_weight=2), DP penalized AE-RNN (dur_weight=1):
boundary_precision: 0.5922140060896042
boundary_recall: 0.7671966190380358
boundary_fscore: 0.6684434781083972
token_precision: 0.23382054243498346
type_precision: 0.26392887383573244
token_recall: 0.29657506566983227
type_recall: 0.35136963138315863
token_fscore: 0.2614853579779534
type_fscore: 0.3014361007688216
words: 11810
coverage: 0.9998313943685719
ned: 0.958675171906796
pairs: 564190
DP penalized CPC-big (dur_weight=1), DP penalized AE-RNN (dur_weight=3):
boundary_precision: 0.6525670837485698
boundary_recall: 0.7116119943650634
boundary_fscore: 0.6808117370711232
token_precision: 0.2663520576444918
type_precision: 0.25612908279157204
token_recall: 0.265154576682158
type_recall: 0.36861684139330403
token_fscore: 0.26575196820494645
type_fscore: 0.3022460486181717
words: 12767
coverage: 0.9999693444306494
ned: 0.9174382816394651
pairs: 15379
DP penalized CPC-big (dur_weight=1), DP penalized AE-RNN (dur_weight=1):
boundary_precision: 0.573416549705856
boundary_recall: 0.7728718051921916
boundary_fscore: 0.6583693341562092
token_precision: 0.21634615384615385
type_precision: 0.25046713096653644
token_recall: 0.28869468579511015
type_recall: 0.33243151843084207
token_fscore: 0.24733835367437032
type_fscore: 0.2856866069266166
words: 11774
coverage: 0.9997700832298708
ned: 0.9616700675645037
pairs: 477815
DP penalized XLSR (dur_weight=2500), DP penalized AE-RNN (dur_weight=3):
boundary_precision: 0.5780879175374588
boundary_recall: 0.6568726101831355
boundary_fscore: 0.6149672168211621
token_precision: 0.1839140413102441
type_precision: 0.17905038057267125
token_recall: 0.17811679127096383
type_recall: 0.27843535114417767
token_fscore: 0.18096900020529666
type_fscore: 0.21794758669372627
words: 13795
coverage: 0.9997394276605203
ned: 0.8476354658834546
pairs: 7831
DP penalized XLSR (dur_weight=1500), DP penalized AE-RNN (dur_weight=3):
boundary_precision: 0.5632707950669116
boundary_recall: 0.6912054739384182
boundary_fscore: 0.6207145826179676
token_precision: 0.19804022692109335
type_precision: 0.1994954954954955
token_recall: 0.2133764396847848
type_recall: 0.3120279562619772
token_fscore: 0.2054224924012158
type_fscore: 0.24338345203552275
words: 13875
coverage: 0.9997394276605203
ned: 0.8672395157803717
pairs: 15420
DP penalized XLSR (dur_weight=500), DP penalized AE-RNN (dur_weight=3):
boundary_precision: 0.5272764681676563
boundary_recall: 0.7220366270879452
boundary_fscore: 0.6094755975334228
token_precision: 0.18828590337524817
type_precision: 0.1983859649122807
token_recall: 0.22994544352394422
type_recall: 0.3186788411678503
token_fscore: 0.20704084417356497
type_fscore: 0.24453959603823366
words: 14250
coverage: 0.999724099875845
ned: 0.8283892604551045
pairs: 17344
HuBERT (dur_weight=2, dur_weight=5):
"boundary": {
"precision": 0.6191301982386345,
"recall": 0.732863755282753,
"fscore": 0.6712131824381613
},
"token": {
"precision": 0.2345940959409594,
"recall": 0.2569205900181855,
"fscore": 0.24525026521361748
},
"type": {
"precision": 0.24242646519474803,
"recall": 0.3725622815917033,
"fscore": 0.2937255599004621
},
"nlp": {
"nwords": [
13633
],
"ned": 0.8597740461054841,
"coverage": 0.9998160665838967,
"npairs": 130513
}
ES-KMeans:
"boundary": {
"precision": 0.4258554271470655,
"recall": 0.7564097403904206,
"fscore": 0.5449220732149329
},
"token": {
"precision": 0.07011532368004127,
"recall": 0.09613053142048898,
"fscore": 0.08108741504569955
},
"type": {
"precision": 0.06787171191447461,
"recall": 0.10878142261300867,
"fscore": 0.08358958811555287
},
"nlp": {
"nwords": [
14218
],
"ned": 0.880712846099123,
"coverage": 1.0,
"npairs": 357082
},
"grouping": {
"precision": 0.20240808264965252,
"recall": 0.31482247297217775,
"fscore": 0.24639926020113523
}
DP penalized CPC-big (dur_weight=2), DP penalized AE-RNN (dur_weight=3), segment length<=50:
boundary_precision: 0.4638637792059966
boundary_recall: 0.5817915641948714
boundary_fscore: 0.5161777929616368
token_precision: 0.10097769204095705
type_precision: 0.04817722938867078
token_recall: 0.11336990763943582
type_recall: 0.15810058436479088
token_fscore: 0.10681558027611417
type_fscore: 0.07385038633896811
words: 71320
coverage: 0.9994668206384628
ned: 0.7985821903307203
pairs: 5642231
DP penalized CPC-big (dur_weight=2), DP penalized AE-RNN (dur_weight=3), final:
boundary_precision: 0.4981988292390054
boundary_recall: 0.5787582779652644
boundary_fscore: 0.5354655158796178
token_precision: 0.11984520550303907
type_precision: 0.0522429243423646
token_recall: 0.1245838404811341
type_recall: 0.18294759122072424
token_fscore: 0.12216859010926324
type_fscore: 0.08127638262860414
words: 76106
coverage: 0.9994785245756672
ned: 0.8678625502797006
pairs: 4482962
DP penalized CPC-big (dur_weight=2), DP penalized AE-RNN (dur_weight=1):
boundary_precision: 0.44045020257251954
boundary_recall: 0.6867723291642297
boundary_fscore: 0.5366979509104225
token_precision: 0.11128960259377166
type_precision: 0.05380097401918776
token_recall: 0.16957381685401304
type_recall: 0.11843739934661575
token_fscore: 0.1343841980212064
type_fscore: 0.07399103139013453
words: 47843
coverage: 0.9992093340199644
ned: 0.7407467270862337
pairs: 34973886
DP penalized CPC-big (dur_weight=1), DP penalized AE-RNN (dur_weight=1):
boundary_precision: 0.43373805566862395
boundary_recall: 0.7006309961893207
boundary_fscore: 0.5357874062776931
token_precision: 0.10650271854604094
type_precision: 0.05249721804019114
token_recall: 0.17495704159805256
type_recall: 0.11070721943588092
token_fscore: 0.13240543194825424
type_fscore: 0.07122136048783377
words: 45831
coverage: 0.998635841096945
ned: 0.7472744666179789
pairs: 46830800
DP penalized XLSR (dur_weight=1500), DP penalized AE-RNN (dur_weight=1):
boundary_precision: 0.434751257850677
boundary_recall: 0.5891731661332684
boundary_fscore: 0.5003177364823178
token_precision: 0.08654744268318969
type_precision: 0.04939499419857451
token_recall: 0.11319538913152431
type_recall: 0.1508305342106474
token_fscore: 0.09809384761251069
type_fscore: 0.07441881583726843
words: 66363
coverage: 0.9979388066145451
ned: 0.5986752611793141
pairs: 10054784
HuBERT (dur_weight=2, dur_weight=5):
ned: 0.6601367208337441
coverage: 0.9990766893983136
boundary_precision: 0.5387142623267998
boundary_recall: 0.5259645668576245
boundary_fscore: 0.5322630747937205
grouping_precision: 0.2817639695874773
grouping_recall: 0.4437490853685101
grouping_fscore: 0.34467333961848207
token_precision: 0.126271164125655
type_precision: 0.05858942927190832
token_recall: 0.10729308369728646
type_recall: 0.23641466893664012
token_fscore: 0.11601109936447188
type_fscore: 0.09390649559527726
words: 87695
coverage: 0.9990766893983136
ned: 0.6601367208337441
pairs: 99903
ES-KMeans:
"boundary": {
"precision": 0.3702359159468867,
"recall": 0.5216380472175502,
"fscore": 0.4330861717705002
},
"token": {
"precision": 0.05693321581812708,
"recall": 0.07024146201761294,
"fscore": 0.06289101548336175
},
"type": {
"precision": 0.03963933940605791,
"recall": 0.0985137808862099,
"fscore": 0.05653178427618984
},
"nlp": {
"nwords": [
54012
],
"ned": 0.6878864516115213,
"coverage": 0.9996553840712016,
"npairs": 4354797
},
"grouping": {
"precision": null,
"recall": null,
"fscore": null
}
DP penalized CPC-big (dur_weight=2), DP penalized AE-RNN (dur_weight=3), final:
boundary_precision: 0.5272349929534665
boundary_recall: 0.6328394074696406
boundary_fscore: 0.5752304859347639
token_precision: 0.18578293430444576
type_precision: 0.06063438714420754
token_recall: 0.20514162595681987
type_recall: 0.2744864967668315
token_fscore: 0.19498295626538212
type_fscore: 0.09932726553225167
words: 95210
coverage: 0.9988127448148101
ned: 0.5709890053444755
pairs: 10327401
HuBERT (dur_weight=2, dur_weight=5):
"boundary": {
"precision": 0.5627599139360644,
"recall": 0.5977993536278821,
"fscore": 0.5797506809019994
},
"token": {
"precision": 0.2015219949697274,
"recall": 0.1901620639615389,
"fscore": 0.19567729461165043
},
"type": {
"precision": 0.06525181851530416,
"recall": 0.34675732217573224,
"fscore": 0.10983516442141884
},
"nlp": {
"nwords": [
111767
],
"ned": 0.41745488130523395,
"coverage": 0.9978178340211217,
"npairs": 1983038
},
"grouping": {
"precision": null,
"recall": null,
"fscore": null
}
ES-KMeans:
"matching": null,
"boundary": {
"precision": 0.49339736696212083,
"recall": 0.6673125870919804,
"fscore": 0.5673256652307774
},
"token": {
"precision": 0.18109334650322245,
"recall": 0.20411141758886933,
"fscore": 0.19191465483458126
},
"type": {
"precision": 0.0629473805544349,
"recall": 0.25965195891974135,
"fscore": 0.10132947386976165
},
"nlp": {
"nwords": [
86755
],
"ned": 0.7320613625212131,
"coverage": 0.9997616438128843,
"npairs": 6932727
},
"grouping": {
"precision": null,
"recall": null,
"fscore": null
}
CPC-big results. These were generated by Robin and can be found in
vqseg/zerospeech2017_eval/kamper_submission.zip.
metric: boundary
precision: 0.5053194153293602
recall: 0.6174457334184231
fscore: 0.5557837582625119
metric: token
precision: 0.14448148897050178
recall: 0.16002854211263717
fscore: 0.15185813064988854
metric: type
precision: 0.05823323136611207
recall: 0.19390501428512277
fscore: 0.08956765389360137
CPC-big results. The above is probably incorrect and the correct ones are produced by me below:
"matching": null,
"boundary": {
"precision": 0.41052507315537556,
"recall": 0.5678158985878697,
"fscore": 0.4765264259376142
},
"token": {
"precision": 0.10913348725882524,
"recall": 0.12087688870138596,
"fscore": 0.11470540264028034
},
"type": {
"precision": 0.045847944490039545,
"recall": 0.15266448594261106,
"fscore": 0.07051803115646128
},
"nlp": {
"nwords": [
80418
],
"ned": 0.5638102970665835,
"coverage": 0.9990113192179535,
"npairs": 1191592
},
"grouping": {
"precision": null,
"recall": null,
"fscore": null
}
HuBERT DPDP:
"boundary": {
"precision": 0.4466614587141488,
"recall": 0.5742976457898331,
"fscore": 0.5025012717413657
},
"token": {
"precision": 0.12448098408523768,
"recall": 0.12523112340296985,
"fscore": 0.12485492702861994
},
"type": {
"precision": 0.052714934882984295,
"recall": 0.17916442383338163,
"fscore": 0.08146167893518082
},
"nlp": {
"nwords": [
82083
],
"ned": 0.568326971960231,
"coverage": 0.9987253079917187,
"npairs": 970429
},
"grouping": {
"precision": null,
"recall": null,
"fscore": null
}
ES-KMeans:
"boundary": {
"precision": 0.3542984991771128,
"recall": 0.6252560537539706,
"fscore": 0.4523020811522307
},
"token": {
"precision": 0.09935391241923905,
"recall": 0.13696047334072928,
"fscore": 0.11516487944169385
},
"type": {
"precision": 0.03425854584406066,
"recall": 0.07548341683574179,
"fscore": 0.04712786308877514
},
"nlp": {
"nwords": [
53213
],
"ned": 0.6621976350816241,
"coverage": 0.9991572673333985,
"npairs": 4222872
},
"grouping": {
"precision": null,
"recall": null,
"fscore": null
}
CPC-big results:
"matching": null,
"boundary": {
"precision": 0.631188814763694,
"recall": 0.5646536940068141,
"fscore": 0.5960702906246983
},
"token": {
"precision": 0.17489360154232295,
"recall": 0.1308940433409561,
"fscore": 0.14972829048783146
},
"type": {
"precision": 0.08042438928548039,
"recall": 0.4967637540453074,
"fscore": 0.1384363902823967
},
"nlp": {
"nwords": [
15269
],
"ned": 0.6211650318898436,
"coverage": 0.9993713741760512,
"npairs": 18322
},
"grouping": {
"precision": 0.3711656441717786,
"recall": 0.15307277448565174,
"fscore": 0.2167538773393675
}
HuBERT DPDP:
"boundary": {
"precision": 0.6909318395086302,
"recall": 0.5555987469416688,
"fscore": 0.6159188846641318
},
"token": {
"precision": 0.16766949152542374,
"recall": 0.10772623325710552,
"fscore": 0.13117416959490816
},
"type": {
"precision": 0.0752817525642649,
"recall": 0.4809870550161812,
"fscore": 0.13018723311069744
},
"nlp": {
"nwords": [
15794
],
"ned": 0.7218721640620105,
"coverage": 0.9991289042153851,
"npairs": 123858
},
"grouping": {
"precision": 0.14945111937073183,
"recall": 0.3396189353761568,
"fscore": 0.20756302521008427
}
ES-KMeans:
"boundary": {
"precision": 0.5079948461229749,
"recall": 0.5499279720119818,
"fscore": 0.5281303526724933
},
"token": {
"precision": 0.11520667194253363,
"recall": 0.10304366764673854,
"fscore": 0.10878625010778031
},
"type": {
"precision": 0.08214285714285714,
"recall": 0.4186893203883495,
"fscore": 0.13734076433121017
},
"nlp": {
"nwords": [
12600
],
"ned": 0.7241772624947278,
"coverage": 0.9994162760206189,
"npairs": 439503
},
"grouping": {
"precision": null,
"recall": null,
"fscore": null
}