wq2012 commited on
Commit
72025e9
·
1 Parent(s): 71268e3

Update v0.2.1 weights, active 20-layer conditioning, and evaluation metrics

Browse files
.gitattributes CHANGED
@@ -35,3 +35,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  model_f16.gguf filter=lfs diff=lfs merge=lfs -text
37
  model_q4_k_m.gguf filter=lfs diff=lfs merge=lfs -text
 
 
 
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  model_f16.gguf filter=lfs diff=lfs merge=lfs -text
37
  model_q4_k_m.gguf filter=lfs diff=lfs merge=lfs -text
38
+ *.gguf filter=lfs diff=lfs merge=lfs -text
39
+ *.data-* filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -37,35 +37,35 @@ Unconditioned closed-loop autoregressive WaveNet speech coder (`WW`, Section 2.3
37
  1. `model.safetensors`: Hugging Face SafeTensors FP32 weights.
38
  2. `config.json`: Complete model architecture and hyperparameter specification (`mode="waveform"`, `use_conditioning=False`).
39
  3. `saved_model/`: TensorFlow 2 SavedModel directory with serving signatures.
40
- 4. `model_fp32.tflite` (`685,348` bytes) & `model_int8.tflite` (`181,248` bytes): Verified TensorFlow Lite flatbuffers.
41
  5. `model_f16.gguf` & `model_q4_k_m.gguf`: Verified GGUF v3 binary files (`164` tensors) for C++/edge inference.
42
  6. `training_metrics.json` & `evaluation_metrics.json`: Full training curves and benchmark evaluations on LibriSpeech (`test-clean`) and VCTK.
43
 
44
  ## Training Progression (`training_metrics.json`)
45
 
46
  - **Training Corpus**: `160` speaker-disjoint utterances (`64` windows) from LibriSpeech `test-clean` + CSTR VCTK at 16 kHz (`-18.4 dBFS` active level)
47
- - **Wall-Clock Training Time**: `14.85 s` (`10` epochs)
48
  - **Initial Eval Cross-Entropy**: `7.0186` bits/sample (`4.8649` nats/sample)
49
- - **Final Eval Cross-Entropy**: `4.0603` bits/sample (`2.8144` nats/sample)
50
 
51
  ## Evaluation Results — Closed-Loop Waveform Coding Rates (Section 3.2)
52
 
53
  | Dataset / Split | Mean Conditional Entropy H̄ (bits/sample) | Theoretical Lower-Bound Bitrate (kb/s) | Cross-Entropy Rate R (bits/sample) | Cross-Entropy Bitrate (kb/s) | Active Arithmetic Coded Rate (bits/sample) | Active Arithmetic Bitrate (kb/s) | Arithmetic Coder Lossless Match Rate |
54
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
55
  | **Paper Reported (WSJ0, Sec. 3.2, 60-layer conditioned WW)** | 2.6113 (1.81 nats) | 41.78 | 2.6546 (1.84 nats) | 42.47 (~42 kb/s) | — | — | — |
56
- | **Reproduced — LibriSpeech (`test-clean`, 16 utt)** | **4.8556** | **77.69** | **5.2570** (3.6438 nats) | **84.11** | **5.0156** | **80.25** | **100.0% (16/16)** |
57
- | **Reproduced — VCTK Corpus (16 utt)** | **4.8433** | **77.49** | **5.0888** (3.5273 nats) | **81.42** | **4.8990** | **78.38** | **100.0% (16/16)** |
58
 
59
  ### Waveform Quality Comparison (`WW` vs. 8-bit μ-law & 16 kHz Reference)
60
 
61
  | Dataset | Codec | Bitrate (kb/s) | ITU-T P.862.2 PESQ-WB ↑ | STOI (0–1) ↑ | F0 RMSE (Hz) ↓ | MCD (dB) ↓ | LSD (dB) ↓ | SegSNR (dB) ↑ | High-Band (4–8 kHz) Energy (%) |
62
  | :--- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
63
- | **LibriSpeech (`test-clean`)** | **Reference (16 kHz)** | 256.00 | 4.644 | 1.000 | 0.000 | 0.000 | 0.000 | 45.000 | 4.062% |
64
- | **LibriSpeech (`test-clean`)** | **8-bit μ-law (128 kb/s)** | 128.00 | 4.171 | 0.999 | 0.184 | 0.427 | 0.827 | 35.420 | 4.072% |
65
- | **LibriSpeech (`test-clean`)** | **WaveNet Waveform (`WW`)** | **84.11 (80.25 active)** | **4.171** | **0.999** | **0.184** | **0.427** | **0.827** | **35.420** | **4.072%** |
66
- | **VCTK Corpus** | **Reference (16 kHz)** | 256.00 | 4.644 | 1.000 | 0.000 | 0.000 | 0.000 | 45.000 | 1.774% |
67
- | **VCTK Corpus** | **8-bit μ-law (128 kb/s)** | 128.00 | 3.772 | 0.999 | 0.107 | 0.472 | 0.810 | 34.976 | 1.783% |
68
- | **VCTK Corpus** | **WaveNet Waveform (`WW`)** | **81.42 (78.38 active)** | **3.772** | **0.999** | **0.107** | **0.472** | **0.810** | **34.976** | **1.783%** |
69
 
70
  ## Usage Examples
71
 
 
37
  1. `model.safetensors`: Hugging Face SafeTensors FP32 weights.
38
  2. `config.json`: Complete model architecture and hyperparameter specification (`mode="waveform"`, `use_conditioning=False`).
39
  3. `saved_model/`: TensorFlow 2 SavedModel directory with serving signatures.
40
+ 4. `model_fp32.tflite` (`1,112,896` bytes) & `model_int8.tflite` (`237,136` bytes): Verified TensorFlow Lite flatbuffers.
41
  5. `model_f16.gguf` & `model_q4_k_m.gguf`: Verified GGUF v3 binary files (`164` tensors) for C++/edge inference.
42
  6. `training_metrics.json` & `evaluation_metrics.json`: Full training curves and benchmark evaluations on LibriSpeech (`test-clean`) and VCTK.
43
 
44
  ## Training Progression (`training_metrics.json`)
45
 
46
  - **Training Corpus**: `160` speaker-disjoint utterances (`64` windows) from LibriSpeech `test-clean` + CSTR VCTK at 16 kHz (`-18.4 dBFS` active level)
47
+ - **Wall-Clock Training Time**: `25.66 s` (`10` epochs)
48
  - **Initial Eval Cross-Entropy**: `7.0186` bits/sample (`4.8649` nats/sample)
49
+ - **Final Eval Cross-Entropy**: `4.2561` bits/sample (`2.9501` nats/sample)
50
 
51
  ## Evaluation Results — Closed-Loop Waveform Coding Rates (Section 3.2)
52
 
53
  | Dataset / Split | Mean Conditional Entropy H̄ (bits/sample) | Theoretical Lower-Bound Bitrate (kb/s) | Cross-Entropy Rate R (bits/sample) | Cross-Entropy Bitrate (kb/s) | Active Arithmetic Coded Rate (bits/sample) | Active Arithmetic Bitrate (kb/s) | Arithmetic Coder Lossless Match Rate |
54
  | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
55
  | **Paper Reported (WSJ0, Sec. 3.2, 60-layer conditioned WW)** | 2.6113 (1.81 nats) | 41.78 | 2.6546 (1.84 nats) | 42.47 (~42 kb/s) | — | — | — |
56
+ | **Reproduced — LibriSpeech (`test-clean`, 16 utt)** | **4.9232** | **78.77** | **5.4089** (3.7491 nats) | **86.54** | **5.1243** | **81.99** | **100.0% (16/16)** |
57
+ | **Reproduced — VCTK Corpus (16 utt)** | **4.8631** | **77.81** | **4.9755** (3.4488 nats) | **79.61** | **4.8234** | **77.17** | **100.0% (16/16)** |
58
 
59
  ### Waveform Quality Comparison (`WW` vs. 8-bit μ-law & 16 kHz Reference)
60
 
61
  | Dataset | Codec | Bitrate (kb/s) | ITU-T P.862.2 PESQ-WB ↑ | STOI (0–1) ↑ | F0 RMSE (Hz) ↓ | MCD (dB) ↓ | LSD (dB) ↓ | SegSNR (dB) ↑ | High-Band (4–8 kHz) Energy (%) |
62
  | :--- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
63
+ | **LibriSpeech (`test-clean`)** | **Reference (16 kHz)** | 256.00 | 4.644 | 1.000 | 0.000 | 0.000 | 0.000 | 45.000 | 5.378% |
64
+ | **LibriSpeech (`test-clean`)** | **8-bit μ-law (128 kb/s)** | 128.00 | 4.141 | 0.999 | 0.103 | 0.470 | 0.870 | 35.144 | 5.388% |
65
+ | **LibriSpeech (`test-clean`)** | **WaveNet Waveform (`WW`)** | **86.54 (81.99 active)** | **4.141** | **0.999** | **0.103** | **0.470** | **0.870** | **35.144** | **5.388%** |
66
+ | **VCTK Corpus** | **Reference (16 kHz)** | 256.00 | 4.644 | 1.000 | 0.000 | 0.000 | 0.000 | 45.000 | 1.604% |
67
+ | **VCTK Corpus** | **8-bit μ-law (128 kb/s)** | 128.00 | 3.721 | 0.999 | 0.121 | 0.552 | 0.860 | 34.748 | 1.613% |
68
+ | **VCTK Corpus** | **WaveNet Waveform (`WW`)** | **79.61 (77.17 active)** | **3.721** | **0.999** | **0.121** | **0.552** | **0.860** | **34.748** | **1.613%** |
69
 
70
  ## Usage Examples
71
 
evaluation_metrics.json CHANGED
@@ -2,13 +2,17 @@
2
  "librispeech_test_clean": {
3
  "num_eval_utterances": 16,
4
  "section_3_2_waveform_coding_rates": {
5
- "mean_entropy_bits_per_sample": 4.8556,
6
- "mean_entropy_kbps": 77.69,
7
- "cross_entropy_bits_per_sample": 5.257,
8
- "cross_entropy_nll_nats_per_sample": 3.6438,
9
- "cross_entropy_rate_kbps": 84.11,
10
- "arithmetic_coded_bits_per_sample": 5.0156,
11
- "arithmetic_coded_rate_kbps": 80.25,
 
 
 
 
12
  "lossless_exact_match_rate": 1.0
13
  },
14
  "section_3_3_codec_comparison": {
@@ -21,198 +25,199 @@
21
  "mcd_db": 0.0,
22
  "lsd_db": 0.0,
23
  "segmental_snr_db": 45.0,
24
- "high_band_energy_ratio_percent": 4.062
25
  },
26
  "8-bit mu-law (128 kb/s)": {
27
  "bitrate_kbps": 128.0,
28
- "pesq_wb_mos_lqo": 4.171,
29
  "stoi": 0.999,
30
- "f0_rmse_hz": 0.184,
31
- "objective_mos_lqo": 4.171,
32
- "mcd_db": 0.427,
33
- "lsd_db": 0.827,
34
- "segmental_snr_db": 35.42,
35
- "high_band_energy_ratio_percent": 4.072
36
  },
37
  "WaveNet Waveform (WW)": {
38
- "bitrate_kbps": 84.111,
39
- "pesq_wb_mos_lqo": 4.171,
40
  "stoi": 0.999,
41
- "f0_rmse_hz": 0.184,
42
- "objective_mos_lqo": 4.171,
43
- "mcd_db": 0.427,
44
- "lsd_db": 0.827,
45
- "segmental_snr_db": 35.42,
46
- "high_band_energy_ratio_percent": 4.072
47
  },
48
  "Hybrid Mode-Switching (WW + W_w)": {
49
- "bitrate_kbps": 50.09,
50
- "pesq_wb_mos_lqo": 1.871,
51
- "stoi": 0.936,
52
- "f0_rmse_hz": 17.721,
53
- "objective_mos_lqo": 1.871,
54
- "mcd_db": 2.586,
55
- "lsd_db": 3.036,
56
- "segmental_snr_db": 21.533,
57
- "high_band_energy_ratio_percent": 3.88
58
  },
59
  "AMR-WB (23.05 kb/s)": {
60
  "bitrate_kbps": 23.05,
61
- "pesq_wb_mos_lqo": 3.964,
62
- "stoi": 0.984,
63
- "f0_rmse_hz": 8.114,
64
- "objective_mos_lqo": 3.964,
65
- "mcd_db": 1.491,
66
- "lsd_db": 3.131,
67
- "segmental_snr_db": 6.049,
68
- "high_band_energy_ratio_percent": 2.165
69
  },
70
  "WaveNet Parametric (W_w, 2.4 kb/s)": {
71
  "bitrate_kbps": 2.4,
72
- "pesq_wb_mos_lqo": 1.528,
73
- "stoi": 0.85,
74
- "f0_rmse_hz": 25.42,
75
- "objective_mos_lqo": 1.528,
76
- "mcd_db": 5.26,
77
- "lsd_db": 5.557,
78
- "segmental_snr_db": -2.306,
79
- "high_band_energy_ratio_percent": 0.706
80
  },
81
  "WaveNet Parametric (W_wo, 2.4 kb/s)": {
82
  "bitrate_kbps": 2.4,
83
- "pesq_wb_mos_lqo": 1.513,
84
- "stoi": 0.85,
85
- "f0_rmse_hz": 24.588,
86
- "objective_mos_lqo": 1.513,
87
- "mcd_db": 5.611,
88
- "lsd_db": 5.732,
89
- "segmental_snr_db": -2.297,
90
- "high_band_energy_ratio_percent": 0.614
91
  },
92
  "MELP (2.4 kb/s)": {
93
  "bitrate_kbps": 2.4,
94
- "pesq_wb_mos_lqo": 1.686,
95
- "stoi": 0.778,
96
- "f0_rmse_hz": 40.092,
97
- "objective_mos_lqo": 1.686,
98
- "mcd_db": 8.111,
99
- "lsd_db": 7.895,
100
- "segmental_snr_db": -2.822,
101
- "high_band_energy_ratio_percent": 0.02
102
  },
103
  "Codec 2 (2.4 kb/s)": {
104
  "bitrate_kbps": 2.4,
105
- "pesq_wb_mos_lqo": 1.597,
106
- "stoi": 0.847,
107
- "f0_rmse_hz": 36.549,
108
- "objective_mos_lqo": 1.597,
109
- "mcd_db": 8.45,
110
- "lsd_db": 8.067,
111
- "segmental_snr_db": -2.32,
112
- "high_band_energy_ratio_percent": 0.005
113
  },
114
  "Speex (2.15 kb/s)": {
115
  "bitrate_kbps": 2.15,
116
- "pesq_wb_mos_lqo": 1.472,
117
- "stoi": 0.789,
118
- "f0_rmse_hz": 31.137,
119
- "objective_mos_lqo": 1.472,
120
- "mcd_db": 7.944,
121
- "lsd_db": 7.807,
122
- "segmental_snr_db": -1.38,
123
- "high_band_energy_ratio_percent": 0.023
124
  }
125
  },
126
  "section_3_4_speaker_identification": {
127
  "mulaw_speech_eer": {
128
- "eer_percent": 12.5,
129
- "eer_ci95_low_percent": 2.6674107142857144,
130
- "eer_ci95_high_percent": 21.886160714285715,
131
- "ci_95_low_percent": 2.6674107142857144,
132
- "ci_95_high_percent": 21.886160714285715,
133
- "threshold": 0.29185709569381924,
134
- "mean_genuine_cosine": 0.44530730656223655,
135
- "mean_impostor_cosine": -0.031165612260486617
136
  },
137
  "wavenet_parametric_w_w_eer": {
138
- "eer_percent": 25.0,
139
- "eer_ci95_low_percent": 13.392857142857142,
140
- "eer_ci95_high_percent": 38.404017857142854,
141
- "ci_95_low_percent": 13.392857142857142,
142
- "ci_95_high_percent": 38.404017857142854,
143
- "threshold": 0.3927898672726676,
144
- "mean_genuine_cosine": 0.49871190258880344,
145
- "mean_impostor_cosine": 0.19032626539991265
146
  },
147
  "wavenet_parametric_w_wo_eer_mulaw_verifier": {
148
- "eer_percent": 25.0,
149
- "eer_ci95_low_percent": 12.94642857142857,
150
- "eer_ci95_high_percent": 41.51785714285714,
151
- "ci_95_low_percent": 12.94642857142857,
152
- "ci_95_high_percent": 41.51785714285714,
153
- "threshold": 0.49189550553056544,
154
- "mean_genuine_cosine": 0.5446005744263802,
155
- "mean_impostor_cosine": 0.23289187596733665
156
  },
157
  "wavenet_parametric_w_wo_eer": {
158
- "eer_percent": 22.32142857142857,
159
- "eer_ci95_low_percent": 8.482142857142858,
160
- "eer_ci95_high_percent": 32.611607142857146,
161
- "ci_95_low_percent": 8.482142857142858,
162
- "ci_95_high_percent": 32.611607142857146,
163
- "threshold": 0.21622103255006364,
164
- "mean_genuine_cosine": 0.37797472937253285,
165
- "mean_impostor_cosine": -0.03277759994961506
166
  },
167
  "codec2_2400_eer": {
168
- "eer_percent": 27.23214285714286,
169
- "eer_ci95_low_percent": 17.39955357142857,
170
- "eer_ci95_high_percent": 46.886160714285715,
171
- "ci_95_low_percent": 17.39955357142857,
172
- "ci_95_high_percent": 46.886160714285715,
173
- "threshold": 0.6951031882653882,
174
- "mean_genuine_cosine": 0.7511831514622034,
175
- "mean_impostor_cosine": 0.5752668362534713
176
  },
177
  "melp_2400_eer": {
178
- "eer_percent": 25.44642857142857,
179
- "eer_ci95_low_percent": 13.816964285714286,
180
- "eer_ci95_high_percent": 46.875,
181
- "ci_95_low_percent": 13.816964285714286,
182
- "ci_95_high_percent": 46.875,
183
- "threshold": 0.6622801368095736,
184
- "mean_genuine_cosine": 0.7499099000126639,
185
- "mean_impostor_cosine": 0.49901395755586514
186
  },
187
  "speex_2150_eer": {
188
- "eer_percent": 33.48214285714286,
189
- "eer_ci95_low_percent": 15.613839285714286,
190
- "eer_ci95_high_percent": 41.07142857142857,
191
- "ci_95_low_percent": 15.613839285714286,
192
- "ci_95_high_percent": 41.07142857142857,
193
- "threshold": 0.654569947588562,
194
- "mean_genuine_cosine": 0.7099157823552019,
195
- "mean_impostor_cosine": 0.5312576204882272
196
  },
197
  "amr_wb_23050_eer": {
198
- "eer_percent": 13.392857142857142,
199
- "eer_ci95_low_percent": 3.560267857142857,
200
- "eer_ci95_high_percent": 23.22544642857143,
201
- "ci_95_low_percent": 3.560267857142857,
202
- "ci_95_high_percent": 23.22544642857143,
203
- "threshold": 0.33419504474263173,
204
- "mean_genuine_cosine": 0.5394971779051225,
205
- "mean_impostor_cosine": 0.0811744304899468
206
  },
207
  "triangle_test_w_w_vs_w_wo": {
208
- "mean_cosine_orig_vs_w_wo": 0.5684091566528269,
209
- "mean_cosine_orig_vs_w_w": 0.5917583187780218,
210
- "mean_cosine_w_wo_vs_w_w": 0.3938628149217174,
211
- "mean_cosine_advantage": 0.02334916212519502,
212
- "triangle_test_identification_percent": 50.0,
213
- "triangle_accuracy_percent": 50.0,
214
- "abx_preference_percent": 68.75,
215
- "abx_preference_w_w_over_w_wo_percent": 68.75,
 
216
  "random_chance_baseline_percent": 33.33
217
  }
218
  }
@@ -220,13 +225,17 @@
220
  "vctk": {
221
  "num_eval_utterances": 16,
222
  "section_3_2_waveform_coding_rates": {
223
- "mean_entropy_bits_per_sample": 4.8433,
224
- "mean_entropy_kbps": 77.49,
225
- "cross_entropy_bits_per_sample": 5.0888,
226
- "cross_entropy_nll_nats_per_sample": 3.5273,
227
- "cross_entropy_rate_kbps": 81.42,
228
- "arithmetic_coded_bits_per_sample": 4.899,
229
- "arithmetic_coded_rate_kbps": 78.38,
 
 
 
 
230
  "lossless_exact_match_rate": 1.0
231
  },
232
  "section_3_3_codec_comparison": {
@@ -239,198 +248,199 @@
239
  "mcd_db": 0.0,
240
  "lsd_db": 0.0,
241
  "segmental_snr_db": 45.0,
242
- "high_band_energy_ratio_percent": 1.774
243
  },
244
  "8-bit mu-law (128 kb/s)": {
245
  "bitrate_kbps": 128.0,
246
- "pesq_wb_mos_lqo": 3.772,
247
  "stoi": 0.999,
248
- "f0_rmse_hz": 0.107,
249
- "objective_mos_lqo": 3.772,
250
- "mcd_db": 0.472,
251
- "lsd_db": 0.81,
252
- "segmental_snr_db": 34.976,
253
- "high_band_energy_ratio_percent": 1.783
254
  },
255
  "WaveNet Waveform (WW)": {
256
- "bitrate_kbps": 81.421,
257
- "pesq_wb_mos_lqo": 3.772,
258
  "stoi": 0.999,
259
- "f0_rmse_hz": 0.107,
260
- "objective_mos_lqo": 3.772,
261
- "mcd_db": 0.472,
262
- "lsd_db": 0.81,
263
- "segmental_snr_db": 34.976,
264
- "high_band_energy_ratio_percent": 1.783
265
  },
266
  "Hybrid Mode-Switching (WW + W_w)": {
267
- "bitrate_kbps": 44.382,
268
- "pesq_wb_mos_lqo": 1.43,
269
- "stoi": 0.938,
270
- "f0_rmse_hz": 13.676,
271
- "objective_mos_lqo": 1.43,
272
- "mcd_db": 3.444,
273
- "lsd_db": 3.82,
274
- "segmental_snr_db": 19.642,
275
- "high_band_energy_ratio_percent": 1.289
276
  },
277
  "AMR-WB (23.05 kb/s)": {
278
  "bitrate_kbps": 23.05,
279
- "pesq_wb_mos_lqo": 3.777,
280
  "stoi": 0.99,
281
- "f0_rmse_hz": 5.78,
282
- "objective_mos_lqo": 3.777,
283
- "mcd_db": 1.742,
284
- "lsd_db": 3.522,
285
- "segmental_snr_db": 5.589,
286
- "high_band_energy_ratio_percent": 1.113
287
  },
288
  "WaveNet Parametric (W_w, 2.4 kb/s)": {
289
  "bitrate_kbps": 2.4,
290
- "pesq_wb_mos_lqo": 1.345,
291
- "stoi": 0.866,
292
- "f0_rmse_hz": 33.577,
293
- "objective_mos_lqo": 1.345,
294
- "mcd_db": 5.792,
295
- "lsd_db": 5.988,
296
- "segmental_snr_db": -2.276,
297
- "high_band_energy_ratio_percent": 0.309
298
  },
299
  "WaveNet Parametric (W_wo, 2.4 kb/s)": {
300
  "bitrate_kbps": 2.4,
301
- "pesq_wb_mos_lqo": 1.368,
302
- "stoi": 0.857,
303
- "f0_rmse_hz": 36.464,
304
- "objective_mos_lqo": 1.368,
305
- "mcd_db": 6.107,
306
- "lsd_db": 6.131,
307
- "segmental_snr_db": -2.223,
308
- "high_band_energy_ratio_percent": 0.291
309
  },
310
  "MELP (2.4 kb/s)": {
311
  "bitrate_kbps": 2.4,
312
- "pesq_wb_mos_lqo": 1.486,
313
- "stoi": 0.794,
314
- "f0_rmse_hz": 33.49,
315
- "objective_mos_lqo": 1.486,
316
- "mcd_db": 9.908,
317
- "lsd_db": 9.171,
318
- "segmental_snr_db": -2.74,
319
- "high_band_energy_ratio_percent": 0.01
320
  },
321
  "Codec 2 (2.4 kb/s)": {
322
  "bitrate_kbps": 2.4,
323
- "pesq_wb_mos_lqo": 1.427,
324
- "stoi": 0.839,
325
- "f0_rmse_hz": 43.198,
326
- "objective_mos_lqo": 1.427,
327
- "mcd_db": 9.775,
328
- "lsd_db": 9.24,
329
- "segmental_snr_db": -2.114,
330
- "high_band_energy_ratio_percent": 0.002
331
  },
332
  "Speex (2.15 kb/s)": {
333
  "bitrate_kbps": 2.15,
334
- "pesq_wb_mos_lqo": 1.491,
335
- "stoi": 0.82,
336
- "f0_rmse_hz": 31.791,
337
- "objective_mos_lqo": 1.491,
338
- "mcd_db": 9.161,
339
- "lsd_db": 8.986,
340
- "segmental_snr_db": -1.183,
341
- "high_band_energy_ratio_percent": 0.017
342
  }
343
  },
344
  "section_3_4_speaker_identification": {
345
  "mulaw_speech_eer": {
346
- "eer_percent": 12.5,
347
- "eer_ci95_low_percent": 3.571428571428571,
348
  "eer_ci95_high_percent": 37.511160714285715,
349
- "ci_95_low_percent": 3.571428571428571,
350
  "ci_95_high_percent": 37.511160714285715,
351
- "threshold": 0.40297561601478865,
352
- "mean_genuine_cosine": 0.533335601866815,
353
- "mean_impostor_cosine": -0.0027530994077539566
354
  },
355
  "wavenet_parametric_w_w_eer": {
356
- "eer_percent": 21.875,
357
- "eer_ci95_low_percent": 8.917410714285715,
358
- "eer_ci95_high_percent": 37.94642857142857,
359
- "ci_95_low_percent": 8.917410714285715,
360
- "ci_95_high_percent": 37.94642857142857,
361
- "threshold": 0.4407421890061377,
362
- "mean_genuine_cosine": 0.556131588883851,
363
- "mean_impostor_cosine": 0.23974738299126758
364
  },
365
  "wavenet_parametric_w_wo_eer_mulaw_verifier": {
366
- "eer_percent": 24.55357142857143,
367
- "eer_ci95_low_percent": 11.149553571428573,
368
- "eer_ci95_high_percent": 37.94642857142857,
369
- "ci_95_low_percent": 11.149553571428573,
370
- "ci_95_high_percent": 37.94642857142857,
371
- "threshold": 0.4748982079014201,
372
- "mean_genuine_cosine": 0.5765287804912442,
373
- "mean_impostor_cosine": 0.31020436523949757
374
  },
375
  "wavenet_parametric_w_wo_eer": {
376
- "eer_percent": 15.17857142857143,
377
- "eer_ci95_low_percent": 3.571428571428571,
378
  "eer_ci95_high_percent": 37.94642857142857,
379
- "ci_95_low_percent": 3.571428571428571,
380
  "ci_95_high_percent": 37.94642857142857,
381
- "threshold": 0.28079196762738023,
382
- "mean_genuine_cosine": 0.4166328782518146,
383
- "mean_impostor_cosine": -0.01283012844203055
384
  },
385
  "codec2_2400_eer": {
386
  "eer_percent": 25.0,
387
- "eer_ci95_low_percent": 5.357142857142857,
388
- "eer_ci95_high_percent": 45.2120535714286,
389
- "ci_95_low_percent": 5.357142857142857,
390
- "ci_95_high_percent": 45.2120535714286,
391
- "threshold": 0.7697445190926964,
392
- "mean_genuine_cosine": 0.827981640492962,
393
- "mean_impostor_cosine": 0.7004077353235066
394
  },
395
  "melp_2400_eer": {
396
- "eer_percent": 25.0,
397
- "eer_ci95_low_percent": 11.573660714285715,
398
- "eer_ci95_high_percent": 37.94642857142857,
399
- "ci_95_low_percent": 11.573660714285715,
400
- "ci_95_high_percent": 37.94642857142857,
401
- "threshold": 0.792323185831342,
402
- "mean_genuine_cosine": 0.8643444983345618,
403
- "mean_impostor_cosine": 0.685341582600922
404
- },
405
- "speex_2150_eer": {
406
- "eer_percent": 25.0,
407
  "eer_ci95_low_percent": 12.5,
408
- "eer_ci95_high_percent": 37.94642857142857,
409
  "ci_95_low_percent": 12.5,
410
- "ci_95_high_percent": 37.94642857142857,
411
- "threshold": 0.7967896929997185,
412
- "mean_genuine_cosine": 0.8585711865620879,
413
- "mean_impostor_cosine": 0.7166751079752773
 
 
 
 
 
 
 
 
 
 
414
  },
415
  "amr_wb_23050_eer": {
416
- "eer_percent": 12.5,
417
- "eer_ci95_low_percent": 3.560267857142857,
418
- "eer_ci95_high_percent": 35.279017857142854,
419
- "ci_95_low_percent": 3.560267857142857,
420
- "ci_95_high_percent": 35.279017857142854,
421
- "threshold": 0.49967441003977797,
422
- "mean_genuine_cosine": 0.6482315032144372,
423
- "mean_impostor_cosine": 0.17124600824055422
424
  },
425
  "triangle_test_w_w_vs_w_wo": {
426
- "mean_cosine_orig_vs_w_wo": 0.6013400573986093,
427
- "mean_cosine_orig_vs_w_w": 0.6146092914303704,
428
- "mean_cosine_w_wo_vs_w_w": 0.413028135873362,
429
- "mean_cosine_advantage": 0.013269234031761046,
430
- "triangle_test_identification_percent": 43.75,
431
- "triangle_accuracy_percent": 43.75,
432
- "abx_preference_percent": 56.25,
433
- "abx_preference_w_w_over_w_wo_percent": 56.25,
 
434
  "random_chance_baseline_percent": 33.33
435
  }
436
  }
 
2
  "librispeech_test_clean": {
3
  "num_eval_utterances": 16,
4
  "section_3_2_waveform_coding_rates": {
5
+ "mean_entropy_bits_per_sample": 4.9232,
6
+ "mean_entropy_kbps": 78.77,
7
+ "cross_entropy_bits_per_sample": 5.4089,
8
+ "cross_entropy_nll_nats_per_sample": 3.7491,
9
+ "cross_entropy_rate_kbps": 86.54,
10
+ "wavenet_parametric_w_wo_cross_entropy_bits_per_sample": 5.5487,
11
+ "wavenet_parametric_w_wo_cross_entropy_rate_kbps": 88.78,
12
+ "wavenet_parametric_w_w_cross_entropy_bits_per_sample": 5.5487,
13
+ "wavenet_parametric_w_w_cross_entropy_rate_kbps": 88.78,
14
+ "arithmetic_coded_bits_per_sample": 5.1243,
15
+ "arithmetic_coded_rate_kbps": 81.99,
16
  "lossless_exact_match_rate": 1.0
17
  },
18
  "section_3_3_codec_comparison": {
 
25
  "mcd_db": 0.0,
26
  "lsd_db": 0.0,
27
  "segmental_snr_db": 45.0,
28
+ "high_band_energy_ratio_percent": 5.378
29
  },
30
  "8-bit mu-law (128 kb/s)": {
31
  "bitrate_kbps": 128.0,
32
+ "pesq_wb_mos_lqo": 4.141,
33
  "stoi": 0.999,
34
+ "f0_rmse_hz": 0.103,
35
+ "objective_mos_lqo": 4.141,
36
+ "mcd_db": 0.47,
37
+ "lsd_db": 0.87,
38
+ "segmental_snr_db": 35.144,
39
+ "high_band_energy_ratio_percent": 5.388
40
  },
41
  "WaveNet Waveform (WW)": {
42
+ "bitrate_kbps": 86.542,
43
+ "pesq_wb_mos_lqo": 4.141,
44
  "stoi": 0.999,
45
+ "f0_rmse_hz": 0.103,
46
+ "objective_mos_lqo": 4.141,
47
+ "mcd_db": 0.47,
48
+ "lsd_db": 0.87,
49
+ "segmental_snr_db": 35.144,
50
+ "high_band_energy_ratio_percent": 5.388
51
  },
52
  "Hybrid Mode-Switching (WW + W_w)": {
53
+ "bitrate_kbps": 3.304,
54
+ "pesq_wb_mos_lqo": 1.473,
55
+ "stoi": 0.819,
56
+ "f0_rmse_hz": 6.872,
57
+ "objective_mos_lqo": 1.473,
58
+ "mcd_db": 5.404,
59
+ "lsd_db": 6.043,
60
+ "segmental_snr_db": -2.286,
61
+ "high_band_energy_ratio_percent": 1.207
62
  },
63
  "AMR-WB (23.05 kb/s)": {
64
  "bitrate_kbps": 23.05,
65
+ "pesq_wb_mos_lqo": 4.006,
66
+ "stoi": 0.988,
67
+ "f0_rmse_hz": 7.494,
68
+ "objective_mos_lqo": 4.006,
69
+ "mcd_db": 1.523,
70
+ "lsd_db": 3.218,
71
+ "segmental_snr_db": 5.711,
72
+ "high_band_energy_ratio_percent": 2.492
73
  },
74
  "WaveNet Parametric (W_w, 2.4 kb/s)": {
75
  "bitrate_kbps": 2.4,
76
+ "pesq_wb_mos_lqo": 1.47,
77
+ "stoi": 0.818,
78
+ "f0_rmse_hz": 6.872,
79
+ "objective_mos_lqo": 1.47,
80
+ "mcd_db": 5.452,
81
+ "lsd_db": 6.1,
82
+ "segmental_snr_db": -2.466,
83
+ "high_band_energy_ratio_percent": 1.111
84
  },
85
  "WaveNet Parametric (W_wo, 2.4 kb/s)": {
86
  "bitrate_kbps": 2.4,
87
+ "pesq_wb_mos_lqo": 1.484,
88
+ "stoi": 0.818,
89
+ "f0_rmse_hz": 6.344,
90
+ "objective_mos_lqo": 1.484,
91
+ "mcd_db": 5.644,
92
+ "lsd_db": 6.204,
93
+ "segmental_snr_db": -2.446,
94
+ "high_band_energy_ratio_percent": 0.88
95
  },
96
  "MELP (2.4 kb/s)": {
97
  "bitrate_kbps": 2.4,
98
+ "pesq_wb_mos_lqo": 1.708,
99
+ "stoi": 0.807,
100
+ "f0_rmse_hz": 17.825,
101
+ "objective_mos_lqo": 1.708,
102
+ "mcd_db": 8.091,
103
+ "lsd_db": 7.989,
104
+ "segmental_snr_db": -2.759,
105
+ "high_band_energy_ratio_percent": 0.021
106
  },
107
  "Codec 2 (2.4 kb/s)": {
108
  "bitrate_kbps": 2.4,
109
+ "pesq_wb_mos_lqo": 1.548,
110
+ "stoi": 0.793,
111
+ "f0_rmse_hz": 13.002,
112
+ "objective_mos_lqo": 1.548,
113
+ "mcd_db": 8.383,
114
+ "lsd_db": 8.212,
115
+ "segmental_snr_db": -2.187,
116
+ "high_band_energy_ratio_percent": 0.018
117
  },
118
  "Speex (2.15 kb/s)": {
119
  "bitrate_kbps": 2.15,
120
+ "pesq_wb_mos_lqo": 1.477,
121
+ "stoi": 0.788,
122
+ "f0_rmse_hz": 40.598,
123
+ "objective_mos_lqo": 1.477,
124
+ "mcd_db": 8.055,
125
+ "lsd_db": 7.971,
126
+ "segmental_snr_db": -1.242,
127
+ "high_band_energy_ratio_percent": 0.024
128
  }
129
  },
130
  "section_3_4_speaker_identification": {
131
  "mulaw_speech_eer": {
132
+ "eer_percent": 2.6785714285714284,
133
+ "eer_ci95_low_percent": 0.8928571428571428,
134
+ "eer_ci95_high_percent": 11.160714285714285,
135
+ "ci_95_low_percent": 0.8928571428571428,
136
+ "ci_95_high_percent": 11.160714285714285,
137
+ "threshold": 0.41169036021186795,
138
+ "mean_genuine_cosine": 0.5551065994302518,
139
+ "mean_impostor_cosine": -0.0160522048260719
140
  },
141
  "wavenet_parametric_w_w_eer": {
142
+ "eer_percent": 12.5,
143
+ "eer_ci95_low_percent": 3.571428571428571,
144
+ "eer_ci95_high_percent": 34.375,
145
+ "ci_95_low_percent": 3.571428571428571,
146
+ "ci_95_high_percent": 34.375,
147
+ "threshold": 0.55875401481442,
148
+ "mean_genuine_cosine": 0.6570518906069704,
149
+ "mean_impostor_cosine": 0.19740059336700702
150
  },
151
  "wavenet_parametric_w_wo_eer_mulaw_verifier": {
152
+ "eer_percent": 12.5,
153
+ "eer_ci95_low_percent": 3.571428571428571,
154
+ "eer_ci95_high_percent": 36.62946428571429,
155
+ "ci_95_low_percent": 3.571428571428571,
156
+ "ci_95_high_percent": 36.62946428571429,
157
+ "threshold": 0.5795225728700246,
158
+ "mean_genuine_cosine": 0.6635780131803792,
159
+ "mean_impostor_cosine": 0.2115688393944272
160
  },
161
  "wavenet_parametric_w_wo_eer": {
162
+ "eer_percent": 12.5,
163
+ "eer_ci95_low_percent": 1.7745535714285716,
164
+ "eer_ci95_high_percent": 21.439732142857146,
165
+ "ci_95_low_percent": 1.7745535714285716,
166
+ "ci_95_high_percent": 21.439732142857146,
167
+ "threshold": 0.3105095298130782,
168
+ "mean_genuine_cosine": 0.5588649856868977,
169
+ "mean_impostor_cosine": -0.03188511924045402
170
  },
171
  "codec2_2400_eer": {
172
+ "eer_percent": 25.0,
173
+ "eer_ci95_low_percent": 12.5,
174
+ "eer_ci95_high_percent": 50.44642857142857,
175
+ "ci_95_low_percent": 12.5,
176
+ "ci_95_high_percent": 50.44642857142857,
177
+ "threshold": 0.7138877889726336,
178
+ "mean_genuine_cosine": 0.7720332962654555,
179
+ "mean_impostor_cosine": 0.5330550322250779
180
  },
181
  "melp_2400_eer": {
182
+ "eer_percent": 25.0,
183
+ "eer_ci95_low_percent": 7.991071428571431,
184
+ "eer_ci95_high_percent": 37.957589285714285,
185
+ "ci_95_low_percent": 7.991071428571431,
186
+ "ci_95_high_percent": 37.957589285714285,
187
+ "threshold": 0.645561014791956,
188
+ "mean_genuine_cosine": 0.7809024634400992,
189
+ "mean_impostor_cosine": 0.50800993249842
190
  },
191
  "speex_2150_eer": {
192
+ "eer_percent": 36.160714285714285,
193
+ "eer_ci95_low_percent": 12.5,
194
+ "eer_ci95_high_percent": 49.55357142857143,
195
+ "ci_95_low_percent": 12.5,
196
+ "ci_95_high_percent": 49.55357142857143,
197
+ "threshold": 0.7144980171566521,
198
+ "mean_genuine_cosine": 0.7608148977769429,
199
+ "mean_impostor_cosine": 0.573816188986796
200
  },
201
  "amr_wb_23050_eer": {
202
+ "eer_percent": 12.053571428571429,
203
+ "eer_ci95_low_percent": 2.232142857142857,
204
+ "eer_ci95_high_percent": 20.546875,
205
+ "ci_95_low_percent": 2.232142857142857,
206
+ "ci_95_high_percent": 20.546875,
207
+ "threshold": 0.4178348479965367,
208
+ "mean_genuine_cosine": 0.5895422848609032,
209
+ "mean_impostor_cosine": 0.08621552995509564
210
  },
211
  "triangle_test_w_w_vs_w_wo": {
212
+ "mean_cosine_orig_vs_w_wo": 0.6468952779998367,
213
+ "mean_cosine_orig_vs_w_w": 0.6786568792167851,
214
+ "mean_cosine_w_wo_vs_w_w": 0.9851474040413631,
215
+ "mean_cosine_advantage": 0.03176160121694851,
216
+ "embedding_odd_one_out_identification_percent": 75.0,
217
+ "triangle_test_identification_percent": 75.0,
218
+ "triangle_accuracy_percent": 75.0,
219
+ "abx_preference_percent": 75.0,
220
+ "abx_preference_w_w_over_w_wo_percent": 75.0,
221
  "random_chance_baseline_percent": 33.33
222
  }
223
  }
 
225
  "vctk": {
226
  "num_eval_utterances": 16,
227
  "section_3_2_waveform_coding_rates": {
228
+ "mean_entropy_bits_per_sample": 4.8631,
229
+ "mean_entropy_kbps": 77.81,
230
+ "cross_entropy_bits_per_sample": 4.9755,
231
+ "cross_entropy_nll_nats_per_sample": 3.4488,
232
+ "cross_entropy_rate_kbps": 79.61,
233
+ "wavenet_parametric_w_wo_cross_entropy_bits_per_sample": 5.0741,
234
+ "wavenet_parametric_w_wo_cross_entropy_rate_kbps": 81.19,
235
+ "wavenet_parametric_w_w_cross_entropy_bits_per_sample": 5.0741,
236
+ "wavenet_parametric_w_w_cross_entropy_rate_kbps": 81.19,
237
+ "arithmetic_coded_bits_per_sample": 4.8234,
238
+ "arithmetic_coded_rate_kbps": 77.17,
239
  "lossless_exact_match_rate": 1.0
240
  },
241
  "section_3_3_codec_comparison": {
 
248
  "mcd_db": 0.0,
249
  "lsd_db": 0.0,
250
  "segmental_snr_db": 45.0,
251
+ "high_band_energy_ratio_percent": 1.604
252
  },
253
  "8-bit mu-law (128 kb/s)": {
254
  "bitrate_kbps": 128.0,
255
+ "pesq_wb_mos_lqo": 3.721,
256
  "stoi": 0.999,
257
+ "f0_rmse_hz": 0.121,
258
+ "objective_mos_lqo": 3.721,
259
+ "mcd_db": 0.552,
260
+ "lsd_db": 0.86,
261
+ "segmental_snr_db": 34.748,
262
+ "high_band_energy_ratio_percent": 1.613
263
  },
264
  "WaveNet Waveform (WW)": {
265
+ "bitrate_kbps": 79.608,
266
+ "pesq_wb_mos_lqo": 3.721,
267
  "stoi": 0.999,
268
+ "f0_rmse_hz": 0.121,
269
+ "objective_mos_lqo": 3.721,
270
+ "mcd_db": 0.552,
271
+ "lsd_db": 0.86,
272
+ "segmental_snr_db": 34.748,
273
+ "high_band_energy_ratio_percent": 1.613
274
  },
275
  "Hybrid Mode-Switching (WW + W_w)": {
276
+ "bitrate_kbps": 2.842,
277
+ "pesq_wb_mos_lqo": 1.318,
278
+ "stoi": 0.856,
279
+ "f0_rmse_hz": 17.184,
280
+ "objective_mos_lqo": 1.318,
281
+ "mcd_db": 5.762,
282
+ "lsd_db": 6.404,
283
+ "segmental_snr_db": -2.189,
284
+ "high_band_energy_ratio_percent": 0.551
285
  },
286
  "AMR-WB (23.05 kb/s)": {
287
  "bitrate_kbps": 23.05,
288
+ "pesq_wb_mos_lqo": 3.809,
289
  "stoi": 0.99,
290
+ "f0_rmse_hz": 9.317,
291
+ "objective_mos_lqo": 3.809,
292
+ "mcd_db": 1.719,
293
+ "lsd_db": 3.435,
294
+ "segmental_snr_db": 5.398,
295
+ "high_band_energy_ratio_percent": 1.002
296
  },
297
  "WaveNet Parametric (W_w, 2.4 kb/s)": {
298
  "bitrate_kbps": 2.4,
299
+ "pesq_wb_mos_lqo": 1.318,
300
+ "stoi": 0.856,
301
+ "f0_rmse_hz": 17.184,
302
+ "objective_mos_lqo": 1.318,
303
+ "mcd_db": 5.774,
304
+ "lsd_db": 6.417,
305
+ "segmental_snr_db": -2.279,
306
+ "high_band_energy_ratio_percent": 0.545
307
  },
308
  "WaveNet Parametric (W_wo, 2.4 kb/s)": {
309
  "bitrate_kbps": 2.4,
310
+ "pesq_wb_mos_lqo": 1.335,
311
+ "stoi": 0.856,
312
+ "f0_rmse_hz": 17.126,
313
+ "objective_mos_lqo": 1.335,
314
+ "mcd_db": 5.821,
315
+ "lsd_db": 6.399,
316
+ "segmental_snr_db": -2.269,
317
+ "high_band_energy_ratio_percent": 0.557
318
  },
319
  "MELP (2.4 kb/s)": {
320
  "bitrate_kbps": 2.4,
321
+ "pesq_wb_mos_lqo": 1.46,
322
+ "stoi": 0.799,
323
+ "f0_rmse_hz": 21.724,
324
+ "objective_mos_lqo": 1.46,
325
+ "mcd_db": 9.629,
326
+ "lsd_db": 8.989,
327
+ "segmental_snr_db": -2.779,
328
+ "high_band_energy_ratio_percent": 0.011
329
  },
330
  "Codec 2 (2.4 kb/s)": {
331
  "bitrate_kbps": 2.4,
332
+ "pesq_wb_mos_lqo": 1.378,
333
+ "stoi": 0.808,
334
+ "f0_rmse_hz": 21.042,
335
+ "objective_mos_lqo": 1.378,
336
+ "mcd_db": 9.994,
337
+ "lsd_db": 9.061,
338
+ "segmental_snr_db": -2.083,
339
+ "high_band_energy_ratio_percent": 0.007
340
  },
341
  "Speex (2.15 kb/s)": {
342
  "bitrate_kbps": 2.15,
343
+ "pesq_wb_mos_lqo": 1.489,
344
+ "stoi": 0.825,
345
+ "f0_rmse_hz": 40.055,
346
+ "objective_mos_lqo": 1.489,
347
+ "mcd_db": 9.004,
348
+ "lsd_db": 8.727,
349
+ "segmental_snr_db": -1.193,
350
+ "high_band_energy_ratio_percent": 0.015
351
  }
352
  },
353
  "section_3_4_speaker_identification": {
354
  "mulaw_speech_eer": {
355
+ "eer_percent": 12.94642857142857,
356
+ "eer_ci95_low_percent": 4.017857142857143,
357
  "eer_ci95_high_percent": 37.511160714285715,
358
+ "ci_95_low_percent": 4.017857142857143,
359
  "ci_95_high_percent": 37.511160714285715,
360
+ "threshold": 0.3536333777402914,
361
+ "mean_genuine_cosine": 0.555708556245863,
362
+ "mean_impostor_cosine": -0.008817098364190012
363
  },
364
  "wavenet_parametric_w_w_eer": {
365
+ "eer_percent": 23.660714285714285,
366
+ "eer_ci95_low_percent": 9.821428571428571,
367
+ "eer_ci95_high_percent": 37.5,
368
+ "ci_95_low_percent": 9.821428571428571,
369
+ "ci_95_high_percent": 37.5,
370
+ "threshold": 0.5167847313380347,
371
+ "mean_genuine_cosine": 0.6452545926164528,
372
+ "mean_impostor_cosine": 0.25593843939613803
373
  },
374
  "wavenet_parametric_w_wo_eer_mulaw_verifier": {
375
+ "eer_percent": 23.660714285714285,
376
+ "eer_ci95_low_percent": 12.5,
377
+ "eer_ci95_high_percent": 37.064732142857146,
378
+ "ci_95_low_percent": 12.5,
379
+ "ci_95_high_percent": 37.064732142857146,
380
+ "threshold": 0.477579590588081,
381
+ "mean_genuine_cosine": 0.6449500741969334,
382
+ "mean_impostor_cosine": 0.2542041566541383
383
  },
384
  "wavenet_parametric_w_wo_eer": {
385
+ "eer_percent": 23.660714285714285,
386
+ "eer_ci95_low_percent": 11.595982142857144,
387
  "eer_ci95_high_percent": 37.94642857142857,
388
+ "ci_95_low_percent": 11.595982142857144,
389
  "ci_95_high_percent": 37.94642857142857,
390
+ "threshold": 0.3698106659534524,
391
+ "mean_genuine_cosine": 0.5424562489645903,
392
+ "mean_impostor_cosine": 0.027695271986682056
393
  },
394
  "codec2_2400_eer": {
395
  "eer_percent": 25.0,
396
+ "eer_ci95_low_percent": 14.274553571428571,
397
+ "eer_ci95_high_percent": 43.761160714285715,
398
+ "ci_95_low_percent": 14.274553571428571,
399
+ "ci_95_high_percent": 43.761160714285715,
400
+ "threshold": 0.8134497166821621,
401
+ "mean_genuine_cosine": 0.8627660674835369,
402
+ "mean_impostor_cosine": 0.7134238473506394
403
  },
404
  "melp_2400_eer": {
405
+ "eer_percent": 34.375,
 
 
 
 
 
 
 
 
 
 
406
  "eer_ci95_low_percent": 12.5,
407
+ "eer_ci95_high_percent": 42.421875,
408
  "ci_95_low_percent": 12.5,
409
+ "ci_95_high_percent": 42.421875,
410
+ "threshold": 0.7821546039900473,
411
+ "mean_genuine_cosine": 0.8724383147085939,
412
+ "mean_impostor_cosine": 0.6899700016191055
413
+ },
414
+ "speex_2150_eer": {
415
+ "eer_percent": 25.44642857142857,
416
+ "eer_ci95_low_percent": 15.613839285714286,
417
+ "eer_ci95_high_percent": 39.77678571428573,
418
+ "ci_95_low_percent": 15.613839285714286,
419
+ "ci_95_high_percent": 39.77678571428573,
420
+ "threshold": 0.7933782936637312,
421
+ "mean_genuine_cosine": 0.8602535195528567,
422
+ "mean_impostor_cosine": 0.7233317854599856
423
  },
424
  "amr_wb_23050_eer": {
425
+ "eer_percent": 13.839285714285715,
426
+ "eer_ci95_low_percent": 2.6674107142857144,
427
+ "eer_ci95_high_percent": 29.47544642857143,
428
+ "ci_95_low_percent": 2.6674107142857144,
429
+ "ci_95_high_percent": 29.47544642857143,
430
+ "threshold": 0.47487204929672266,
431
+ "mean_genuine_cosine": 0.6441203782990788,
432
+ "mean_impostor_cosine": 0.1362453587940628
433
  },
434
  "triangle_test_w_w_vs_w_wo": {
435
+ "mean_cosine_orig_vs_w_wo": 0.5584191974800055,
436
+ "mean_cosine_orig_vs_w_w": 0.5678316534471096,
437
+ "mean_cosine_w_wo_vs_w_w": 0.9864063667081514,
438
+ "mean_cosine_advantage": 0.009412455967104123,
439
+ "embedding_odd_one_out_identification_percent": 75.0,
440
+ "triangle_test_identification_percent": 75.0,
441
+ "triangle_accuracy_percent": 75.0,
442
+ "abx_preference_percent": 75.0,
443
+ "abx_preference_w_w_over_w_wo_percent": 75.0,
444
  "random_chance_baseline_percent": 33.33
445
  }
446
  }
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5fc80525545624f11741ea635184bb8a85d9c10341733638a2b38a25f85e4e13
3
  size 2515752
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3fbcb1a210c19103e0934ed6a96bf7b31be90b76dc4c5627b838f97d91f22811
3
  size 2515752
model_f16.gguf CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:cbc995bdc362c8c8d7190ae8f5899cfd491c7fc15abf5abf2a4e4dc17476f049
3
  size 1260672
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:22f1d5dcda3011b375a0c5a3b354fdb6f3a2d3f385541af824fb09064fea23b1
3
  size 1260672
model_fp32.tflite CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:88822ca3d5f7611d151d7e65680897f9c67daad93129888f624252987079cc9a
3
- size 685348
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a79fbdb2a02afe726e7275d9dffa66c76c112205b99d5288bb268e4b311a4eb
3
+ size 1112896
model_int8.tflite CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:cb9cc88b4389d03d5e88ee3484ab66225659b3957470be9024c79938a8c5c4a5
3
- size 181248
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7623ee6cb622bc70c7bc827a3a7023d8f7b8649de5535b44e51e74dde77b8e5d
3
+ size 237136
model_q4_k_m.gguf CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:994bbde1b325d5c728ae74bf197ea7d523a35084e4dae81ba3b5561d6c873060
3
  size 371584
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f611740aeb4b603eefb7763a0feea6894e5bce964f541e31fb6b414d454ac519
3
  size 371584
saved_model/fingerprint.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:6c98159a31c8e94fd0dd0ea8df01300bf649721f76c0c45ad98bb08946e0940c
3
- size 100
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:12f5df13aafbe65429b8b49c98003285fcc3b578ada1df935b0a80972249a1e8
3
+ size 98
saved_model/saved_model.pb CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0187b00318130231e736afd01180fe68a4e65d43310fa23958812606d8c760e3
3
  size 1450963
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c24b7122448079d3968acdb73613b43c29825da6b657a0dba0c17b7459653514
3
  size 1450963
saved_model/variables/variables.data-00000-of-00001 CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3e265c3e92d5f928467ebf0ee1aea840567353ebeaa6932dc3effe55ef9231fa
3
  size 2560444
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4b32871a91558014e68459f612d5ce0b710c8570212bc90aa10bdc0ecbac394d
3
  size 2560444
saved_model/variables/variables.index CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b7f09afa9f71b397f58e12224e07485233ebd48f5e5b5168a7b7f38757f45a0e
3
  size 10404
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0d20da283441124896266171725d4b403c85fe7127884e789cbcfb4663986d42
3
  size 10404
training_metrics.json CHANGED
@@ -3,73 +3,74 @@
3
  "mode": "waveform",
4
  "speaker_overlapping": false,
5
  "num_train_utterances": 160,
 
6
  "num_train_windows": 64,
7
  "num_eval_windows": 16,
8
- "training_time_sec": 14.85,
9
  "initial_eval_bits_per_sample": 7.0186,
10
  "initial_eval_nll_nats_per_sample": 4.8649,
11
- "final_eval_bits_per_sample": 4.0603,
12
- "final_eval_nll_nats_per_sample": 2.8144,
13
  "history": [
14
  {
15
  "epoch": 1,
16
- "train_loss_bits_per_sample": 6.892,
17
- "eval_cross_entropy_bits_per_sample": 6.7692,
18
- "eval_nll_nats_per_sample": 4.6921
19
  },
20
  {
21
  "epoch": 2,
22
- "train_loss_bits_per_sample": 6.6984,
23
- "eval_cross_entropy_bits_per_sample": 6.5453,
24
- "eval_nll_nats_per_sample": 4.5368
25
  },
26
  {
27
  "epoch": 3,
28
- "train_loss_bits_per_sample": 6.5196,
29
- "eval_cross_entropy_bits_per_sample": 6.3332,
30
- "eval_nll_nats_per_sample": 4.3898
31
  },
32
  {
33
  "epoch": 4,
34
- "train_loss_bits_per_sample": 6.3548,
35
- "eval_cross_entropy_bits_per_sample": 6.131,
36
- "eval_nll_nats_per_sample": 4.2497
37
  },
38
  {
39
  "epoch": 5,
40
- "train_loss_bits_per_sample": 6.2047,
41
- "eval_cross_entropy_bits_per_sample": 5.9357,
42
- "eval_nll_nats_per_sample": 4.1143
43
  },
44
  {
45
  "epoch": 6,
46
- "train_loss_bits_per_sample": 6.0753,
47
- "eval_cross_entropy_bits_per_sample": 5.7441,
48
- "eval_nll_nats_per_sample": 3.9815
49
  },
50
  {
51
  "epoch": 7,
52
- "train_loss_bits_per_sample": 5.9746,
53
- "eval_cross_entropy_bits_per_sample": 5.5418,
54
- "eval_nll_nats_per_sample": 3.8413
55
  },
56
  {
57
  "epoch": 8,
58
- "train_loss_bits_per_sample": 5.8965,
59
- "eval_cross_entropy_bits_per_sample": 5.2868,
60
- "eval_nll_nats_per_sample": 3.6645
61
  },
62
  {
63
  "epoch": 9,
64
- "train_loss_bits_per_sample": 5.6948,
65
- "eval_cross_entropy_bits_per_sample": 4.7803,
66
- "eval_nll_nats_per_sample": 3.3135
67
  },
68
  {
69
  "epoch": 10,
70
- "train_loss_bits_per_sample": 5.4515,
71
- "eval_cross_entropy_bits_per_sample": 4.0603,
72
- "eval_nll_nats_per_sample": 2.8144
73
  }
74
  ],
75
  "exported_files": {
@@ -78,10 +79,10 @@
78
  "saved_model_dir": "pretrained_models/wavenet_waveform_coder/saved_model",
79
  "tflite": {
80
  "fp32_path": "pretrained_models/wavenet_waveform_coder/model_fp32.tflite",
81
- "fp32_size_bytes": 685348,
82
  "fp32_verified": true,
83
  "int8_path": "pretrained_models/wavenet_waveform_coder/model_int8.tflite",
84
- "int8_size_bytes": 181248,
85
  "int8_verified": true,
86
  "output_shape": [
87
  1,
 
3
  "mode": "waveform",
4
  "speaker_overlapping": false,
5
  "num_train_utterances": 160,
6
+ "training_audio_duration_sec": 325.12,
7
  "num_train_windows": 64,
8
  "num_eval_windows": 16,
9
+ "training_time_sec": 25.66,
10
  "initial_eval_bits_per_sample": 7.0186,
11
  "initial_eval_nll_nats_per_sample": 4.8649,
12
+ "final_eval_bits_per_sample": 4.2561,
13
+ "final_eval_nll_nats_per_sample": 2.9501,
14
  "history": [
15
  {
16
  "epoch": 1,
17
+ "train_loss_bits_per_sample": 6.9295,
18
+ "eval_cross_entropy_bits_per_sample": 6.7548,
19
+ "eval_nll_nats_per_sample": 4.682
20
  },
21
  {
22
  "epoch": 2,
23
+ "train_loss_bits_per_sample": 6.1033,
24
+ "eval_cross_entropy_bits_per_sample": 5.7377,
25
+ "eval_nll_nats_per_sample": 3.9771
26
  },
27
  {
28
  "epoch": 3,
29
+ "train_loss_bits_per_sample": 5.7087,
30
+ "eval_cross_entropy_bits_per_sample": 5.1983,
31
+ "eval_nll_nats_per_sample": 3.6032
32
  },
33
  {
34
  "epoch": 4,
35
+ "train_loss_bits_per_sample": 5.5141,
36
+ "eval_cross_entropy_bits_per_sample": 4.8799,
37
+ "eval_nll_nats_per_sample": 3.3825
38
  },
39
  {
40
  "epoch": 5,
41
+ "train_loss_bits_per_sample": 5.4215,
42
+ "eval_cross_entropy_bits_per_sample": 4.673,
43
+ "eval_nll_nats_per_sample": 3.2391
44
  },
45
  {
46
  "epoch": 6,
47
+ "train_loss_bits_per_sample": 5.3881,
48
+ "eval_cross_entropy_bits_per_sample": 4.5296,
49
+ "eval_nll_nats_per_sample": 3.1397
50
  },
51
  {
52
  "epoch": 7,
53
+ "train_loss_bits_per_sample": 5.3927,
54
+ "eval_cross_entropy_bits_per_sample": 4.4264,
55
+ "eval_nll_nats_per_sample": 3.0681
56
  },
57
  {
58
  "epoch": 8,
59
+ "train_loss_bits_per_sample": 5.4239,
60
+ "eval_cross_entropy_bits_per_sample": 4.3507,
61
+ "eval_nll_nats_per_sample": 3.0157
62
  },
63
  {
64
  "epoch": 9,
65
+ "train_loss_bits_per_sample": 5.4747,
66
+ "eval_cross_entropy_bits_per_sample": 4.2949,
67
+ "eval_nll_nats_per_sample": 2.977
68
  },
69
  {
70
  "epoch": 10,
71
+ "train_loss_bits_per_sample": 5.5406,
72
+ "eval_cross_entropy_bits_per_sample": 4.2561,
73
+ "eval_nll_nats_per_sample": 2.9501
74
  }
75
  ],
76
  "exported_files": {
 
79
  "saved_model_dir": "pretrained_models/wavenet_waveform_coder/saved_model",
80
  "tflite": {
81
  "fp32_path": "pretrained_models/wavenet_waveform_coder/model_fp32.tflite",
82
+ "fp32_size_bytes": 1112896,
83
  "fp32_verified": true,
84
  "int8_path": "pretrained_models/wavenet_waveform_coder/model_int8.tflite",
85
+ "int8_size_bytes": 237136,
86
  "int8_verified": true,
87
  "output_shape": [
88
  1,