Jiyang Tang
commited on
Commit
•
9db65ab
1
Parent(s):
50cc396
Update model
Browse files- README.md +385 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/RESULTS.md +27 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/config.yaml +286 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/acc.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/backward_time.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/cer.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/cer_ctc.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/forward_time.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/gpu_max_cached_mem_GB.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/iter_time.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/loss.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/loss_att.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/loss_ctc.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/optim0_lr0.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/optim_step_time.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/train_time.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/images/wer.png +0 -0
- exp/asr_ebranchformer_wavlm_aph_en_both/valid.acc.ave_10best.pth +3 -0
- meta.yaml +8 -0
README.md
ADDED
@@ -0,0 +1,385 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
tags:
|
3 |
+
- espnet
|
4 |
+
- audio
|
5 |
+
- automatic-speech-recognition
|
6 |
+
language: en
|
7 |
+
datasets:
|
8 |
+
- A
|
9 |
+
license: cc-by-4.0
|
10 |
+
---
|
11 |
+
|
12 |
+
## ESPnet2 ASR model
|
13 |
+
|
14 |
+
### `espnet/jiyang_tang_aphsiabank_english_asr_ebranchformer_wavlm_aph_en_both`
|
15 |
+
|
16 |
+
This model was trained by Jiyang Tang using A recipe in [espnet](https://github.com/espnet/espnet/).
|
17 |
+
|
18 |
+
### Demo: How to use in ESPnet2
|
19 |
+
|
20 |
+
Follow the [ESPnet installation instructions](https://espnet.github.io/espnet/installation.html)
|
21 |
+
if you haven't done that already.
|
22 |
+
|
23 |
+
```bash
|
24 |
+
cd espnet
|
25 |
+
git checkout edf949f535938da8c705c1d26cc561b2d4cb4778
|
26 |
+
pip install -e .
|
27 |
+
cd jtang1/A/asr1
|
28 |
+
./run.sh --skip_data_prep false --skip_train true --download_model espnet/jiyang_tang_aphsiabank_english_asr_ebranchformer_wavlm_aph_en_both
|
29 |
+
```
|
30 |
+
|
31 |
+
<!-- Generated by scripts/utils/show_asr_result.sh -->
|
32 |
+
# RESULTS
|
33 |
+
## Environments
|
34 |
+
- date: `Tue Mar 7 12:06:32 EST 2023`
|
35 |
+
- python version: `3.9.12 (main, Apr 5 2022, 06:56:58) [GCC 7.5.0]`
|
36 |
+
- espnet version: `espnet 202301`
|
37 |
+
- pytorch version: `pytorch 1.8.1`
|
38 |
+
- Git hash: `b0b2a0aa9c335267046e83036b87e88af30698da`
|
39 |
+
- Commit date: `Tue Feb 7 14:56:31 2023 -0500`
|
40 |
+
|
41 |
+
## asr_ebranchformer_wavlm_aph_en_both
|
42 |
+
### WER
|
43 |
+
|
44 |
+
|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|
45 |
+
|---|---|---|---|---|---|---|---|---|
|
46 |
+
|decode_asr_model_valid.acc.ave/test|28424|296887|83.0|12.4|4.6|2.7|19.6|71.1|
|
47 |
+
|
48 |
+
### CER
|
49 |
+
|
50 |
+
|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|
51 |
+
|---|---|---|---|---|---|---|---|---|
|
52 |
+
|decode_asr_model_valid.acc.ave/test|28424|1507391|91.9|3.0|5.1|3.0|11.1|71.1|
|
53 |
+
|
54 |
+
### TER
|
55 |
+
|
56 |
+
|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|
57 |
+
|---|---|---|---|---|---|---|---|---|
|
58 |
+
|
59 |
+
## ASR config
|
60 |
+
|
61 |
+
<details><summary>expand</summary>
|
62 |
+
|
63 |
+
```
|
64 |
+
config: conf/tuning/train_asr_ebranchformer_small_wavlm_large1.yaml
|
65 |
+
print_config: false
|
66 |
+
log_level: INFO
|
67 |
+
dry_run: false
|
68 |
+
iterator_type: sequence
|
69 |
+
output_dir: exp/asr_ebranchformer_wavlm_aph_en_both
|
70 |
+
ngpu: 1
|
71 |
+
seed: 2022
|
72 |
+
num_workers: 4
|
73 |
+
num_att_plot: 0
|
74 |
+
dist_backend: nccl
|
75 |
+
dist_init_method: env://
|
76 |
+
dist_world_size: 2
|
77 |
+
dist_rank: 0
|
78 |
+
local_rank: 0
|
79 |
+
dist_master_addr: localhost
|
80 |
+
dist_master_port: 44175
|
81 |
+
dist_launcher: null
|
82 |
+
multiprocessing_distributed: true
|
83 |
+
unused_parameters: true
|
84 |
+
sharded_ddp: false
|
85 |
+
cudnn_enabled: true
|
86 |
+
cudnn_benchmark: false
|
87 |
+
cudnn_deterministic: false
|
88 |
+
collect_stats: false
|
89 |
+
write_collected_feats: false
|
90 |
+
max_epoch: 30
|
91 |
+
patience: null
|
92 |
+
val_scheduler_criterion:
|
93 |
+
- valid
|
94 |
+
- loss
|
95 |
+
early_stopping_criterion:
|
96 |
+
- valid
|
97 |
+
- loss
|
98 |
+
- min
|
99 |
+
best_model_criterion:
|
100 |
+
- - valid
|
101 |
+
- acc
|
102 |
+
- max
|
103 |
+
keep_nbest_models: 10
|
104 |
+
nbest_averaging_interval: 0
|
105 |
+
grad_clip: 5
|
106 |
+
grad_clip_type: 2.0
|
107 |
+
grad_noise: false
|
108 |
+
accum_grad: 8
|
109 |
+
no_forward_run: false
|
110 |
+
resume: true
|
111 |
+
train_dtype: float32
|
112 |
+
use_amp: true
|
113 |
+
log_interval: 200
|
114 |
+
use_matplotlib: true
|
115 |
+
use_tensorboard: true
|
116 |
+
create_graph_in_tensorboard: false
|
117 |
+
use_wandb: false
|
118 |
+
wandb_project: null
|
119 |
+
wandb_id: null
|
120 |
+
wandb_entity: null
|
121 |
+
wandb_name: null
|
122 |
+
wandb_model_log_interval: -1
|
123 |
+
detect_anomaly: false
|
124 |
+
pretrain_path: null
|
125 |
+
init_param: []
|
126 |
+
ignore_init_mismatch: false
|
127 |
+
freeze_param:
|
128 |
+
- frontend.upstream
|
129 |
+
num_iters_per_epoch: null
|
130 |
+
batch_size: 20
|
131 |
+
valid_batch_size: null
|
132 |
+
batch_bins: 6000000
|
133 |
+
valid_batch_bins: null
|
134 |
+
train_shape_file:
|
135 |
+
- exp/asr_stats_raw_en_char_sp/train/speech_shape
|
136 |
+
- exp/asr_stats_raw_en_char_sp/train/text_shape.char
|
137 |
+
valid_shape_file:
|
138 |
+
- exp/asr_stats_raw_en_char_sp/valid/speech_shape
|
139 |
+
- exp/asr_stats_raw_en_char_sp/valid/text_shape.char
|
140 |
+
batch_type: numel
|
141 |
+
valid_batch_type: null
|
142 |
+
fold_length:
|
143 |
+
- 80000
|
144 |
+
- 150
|
145 |
+
sort_in_batch: descending
|
146 |
+
sort_batch: descending
|
147 |
+
multiple_iterator: false
|
148 |
+
chunk_length: 500
|
149 |
+
chunk_shift_ratio: 0.5
|
150 |
+
num_cache_chunks: 1024
|
151 |
+
train_data_path_and_name_and_type:
|
152 |
+
- - dump/raw/train_sp/wav.scp
|
153 |
+
- speech
|
154 |
+
- sound
|
155 |
+
- - dump/raw/train_sp/text
|
156 |
+
- text
|
157 |
+
- text
|
158 |
+
valid_data_path_and_name_and_type:
|
159 |
+
- - dump/raw/val/wav.scp
|
160 |
+
- speech
|
161 |
+
- sound
|
162 |
+
- - dump/raw/val/text
|
163 |
+
- text
|
164 |
+
- text
|
165 |
+
allow_variable_data_keys: false
|
166 |
+
max_cache_size: 0.0
|
167 |
+
max_cache_fd: 32
|
168 |
+
valid_max_cache_size: null
|
169 |
+
exclude_weight_decay: false
|
170 |
+
exclude_weight_decay_conf: {}
|
171 |
+
optim: adam
|
172 |
+
optim_conf:
|
173 |
+
lr: 0.001
|
174 |
+
weight_decay: 1.0e-06
|
175 |
+
scheduler: warmuplr
|
176 |
+
scheduler_conf:
|
177 |
+
warmup_steps: 2500
|
178 |
+
token_list:
|
179 |
+
- <blank>
|
180 |
+
- <unk>
|
181 |
+
- '[APH]'
|
182 |
+
- '[NONAPH]'
|
183 |
+
- <space>
|
184 |
+
- e
|
185 |
+
- t
|
186 |
+
- a
|
187 |
+
- h
|
188 |
+
- o
|
189 |
+
- A
|
190 |
+
- n
|
191 |
+
- '['
|
192 |
+
- P
|
193 |
+
- H
|
194 |
+
- ']'
|
195 |
+
- i
|
196 |
+
- s
|
197 |
+
- N
|
198 |
+
- d
|
199 |
+
- r
|
200 |
+
- u
|
201 |
+
- l
|
202 |
+
- m
|
203 |
+
- w
|
204 |
+
- O
|
205 |
+
- y
|
206 |
+
- g
|
207 |
+
- c
|
208 |
+
- b
|
209 |
+
- f
|
210 |
+
- p
|
211 |
+
- k
|
212 |
+
- ''''
|
213 |
+
- v
|
214 |
+
- j
|
215 |
+
- <
|
216 |
+
- L
|
217 |
+
- U
|
218 |
+
- '>'
|
219 |
+
- ɪ
|
220 |
+
- x
|
221 |
+
- ə
|
222 |
+
- z
|
223 |
+
- ɛ
|
224 |
+
- ɑ
|
225 |
+
- q
|
226 |
+
- ɹ
|
227 |
+
- æ
|
228 |
+
- ˞
|
229 |
+
- ʌ
|
230 |
+
- ʃ
|
231 |
+
- ʊ
|
232 |
+
- ɔ
|
233 |
+
- ŋ
|
234 |
+
- ɚ
|
235 |
+
- ɾ
|
236 |
+
- ʒ
|
237 |
+
- ð
|
238 |
+
- θ
|
239 |
+
- ɜ
|
240 |
+
- ɝ
|
241 |
+
- ɡ
|
242 |
+
- '0'
|
243 |
+
- ː
|
244 |
+
- ʔ
|
245 |
+
- ɒ
|
246 |
+
- é
|
247 |
+
- ɸ
|
248 |
+
- ̩
|
249 |
+
- ʤ
|
250 |
+
- ʧ
|
251 |
+
- <sos/eos>
|
252 |
+
init: null
|
253 |
+
input_size: null
|
254 |
+
ctc_conf:
|
255 |
+
dropout_rate: 0.0
|
256 |
+
ctc_type: builtin
|
257 |
+
reduce: true
|
258 |
+
ignore_nan_grad: null
|
259 |
+
zero_infinity: true
|
260 |
+
joint_net_conf: null
|
261 |
+
use_preprocessor: true
|
262 |
+
token_type: char
|
263 |
+
bpemodel: null
|
264 |
+
non_linguistic_symbols: local/nlsyms.txt
|
265 |
+
cleaner: null
|
266 |
+
g2p: null
|
267 |
+
speech_volume_normalize: null
|
268 |
+
rir_scp: null
|
269 |
+
rir_apply_prob: 1.0
|
270 |
+
noise_scp: null
|
271 |
+
noise_apply_prob: 1.0
|
272 |
+
noise_db_range: '13_15'
|
273 |
+
short_noise_thres: 0.5
|
274 |
+
aux_ctc_tasks: []
|
275 |
+
frontend: s3prl
|
276 |
+
frontend_conf:
|
277 |
+
frontend_conf:
|
278 |
+
upstream: wavlm_large
|
279 |
+
download_dir: ./hub
|
280 |
+
multilayer_feature: true
|
281 |
+
fs: 16k
|
282 |
+
specaug: specaug
|
283 |
+
specaug_conf:
|
284 |
+
apply_time_warp: true
|
285 |
+
time_warp_window: 5
|
286 |
+
time_warp_mode: bicubic
|
287 |
+
apply_freq_mask: true
|
288 |
+
freq_mask_width_range:
|
289 |
+
- 0
|
290 |
+
- 27
|
291 |
+
num_freq_mask: 2
|
292 |
+
apply_time_mask: true
|
293 |
+
time_mask_width_ratio_range:
|
294 |
+
- 0.0
|
295 |
+
- 0.05
|
296 |
+
num_time_mask: 5
|
297 |
+
normalize: utterance_mvn
|
298 |
+
normalize_conf: {}
|
299 |
+
model: espnet
|
300 |
+
model_conf:
|
301 |
+
ctc_weight: 0.3
|
302 |
+
lsm_weight: 0.1
|
303 |
+
length_normalized_loss: false
|
304 |
+
extract_feats_in_collect_stats: false
|
305 |
+
preencoder: linear
|
306 |
+
preencoder_conf:
|
307 |
+
input_size: 1024
|
308 |
+
output_size: 80
|
309 |
+
encoder: e_branchformer
|
310 |
+
encoder_conf:
|
311 |
+
output_size: 256
|
312 |
+
attention_heads: 4
|
313 |
+
linear_units: 1024
|
314 |
+
num_blocks: 12
|
315 |
+
dropout_rate: 0.1
|
316 |
+
positional_dropout_rate: 0.1
|
317 |
+
attention_dropout_rate: 0.1
|
318 |
+
layer_drop_rate: 0.1
|
319 |
+
input_layer: conv2d1
|
320 |
+
macaron_ffn: true
|
321 |
+
pos_enc_layer_type: rel_pos
|
322 |
+
attention_layer_type: rel_selfattn
|
323 |
+
rel_pos_type: latest
|
324 |
+
cgmlp_linear_units: 3072
|
325 |
+
cgmlp_conv_kernel: 31
|
326 |
+
use_linear_after_conv: false
|
327 |
+
gate_activation: identity
|
328 |
+
positionwise_layer_type: linear
|
329 |
+
use_ffn: true
|
330 |
+
merge_conv_kernel: 31
|
331 |
+
postencoder: null
|
332 |
+
postencoder_conf: {}
|
333 |
+
decoder: transformer
|
334 |
+
decoder_conf:
|
335 |
+
attention_heads: 4
|
336 |
+
linear_units: 2048
|
337 |
+
num_blocks: 6
|
338 |
+
dropout_rate: 0.1
|
339 |
+
positional_dropout_rate: 0.1
|
340 |
+
self_attention_dropout_rate: 0.1
|
341 |
+
src_attention_dropout_rate: 0.1
|
342 |
+
layer_drop_rate: 0.2
|
343 |
+
preprocessor: default
|
344 |
+
preprocessor_conf: {}
|
345 |
+
required:
|
346 |
+
- output_dir
|
347 |
+
- token_list
|
348 |
+
version: '202301'
|
349 |
+
distributed: true
|
350 |
+
```
|
351 |
+
|
352 |
+
</details>
|
353 |
+
|
354 |
+
|
355 |
+
|
356 |
+
### Citing ESPnet
|
357 |
+
|
358 |
+
```BibTex
|
359 |
+
@inproceedings{watanabe2018espnet,
|
360 |
+
author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
|
361 |
+
title={{ESPnet}: End-to-End Speech Processing Toolkit},
|
362 |
+
year={2018},
|
363 |
+
booktitle={Proceedings of Interspeech},
|
364 |
+
pages={2207--2211},
|
365 |
+
doi={10.21437/Interspeech.2018-1456},
|
366 |
+
url={http://dx.doi.org/10.21437/Interspeech.2018-1456}
|
367 |
+
}
|
368 |
+
|
369 |
+
|
370 |
+
|
371 |
+
|
372 |
+
```
|
373 |
+
|
374 |
+
or arXiv:
|
375 |
+
|
376 |
+
```bibtex
|
377 |
+
@misc{watanabe2018espnet,
|
378 |
+
title={ESPnet: End-to-End Speech Processing Toolkit},
|
379 |
+
author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai},
|
380 |
+
year={2018},
|
381 |
+
eprint={1804.00015},
|
382 |
+
archivePrefix={arXiv},
|
383 |
+
primaryClass={cs.CL}
|
384 |
+
}
|
385 |
+
```
|
exp/asr_ebranchformer_wavlm_aph_en_both/RESULTS.md
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<!-- Generated by scripts/utils/show_asr_result.sh -->
|
2 |
+
# RESULTS
|
3 |
+
## Environments
|
4 |
+
- date: `Tue Mar 7 12:06:32 EST 2023`
|
5 |
+
- python version: `3.9.12 (main, Apr 5 2022, 06:56:58) [GCC 7.5.0]`
|
6 |
+
- espnet version: `espnet 202301`
|
7 |
+
- pytorch version: `pytorch 1.8.1`
|
8 |
+
- Git hash: `b0b2a0aa9c335267046e83036b87e88af30698da`
|
9 |
+
- Commit date: `Tue Feb 7 14:56:31 2023 -0500`
|
10 |
+
|
11 |
+
## asr_ebranchformer_wavlm_aph_en_both
|
12 |
+
### WER
|
13 |
+
|
14 |
+
|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|
15 |
+
|---|---|---|---|---|---|---|---|---|
|
16 |
+
|decode_asr_model_valid.acc.ave/test|28424|296887|83.0|12.4|4.6|2.7|19.6|71.1|
|
17 |
+
|
18 |
+
### CER
|
19 |
+
|
20 |
+
|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|
21 |
+
|---|---|---|---|---|---|---|---|---|
|
22 |
+
|decode_asr_model_valid.acc.ave/test|28424|1507391|91.9|3.0|5.1|3.0|11.1|71.1|
|
23 |
+
|
24 |
+
### TER
|
25 |
+
|
26 |
+
|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|
27 |
+
|---|---|---|---|---|---|---|---|---|
|
exp/asr_ebranchformer_wavlm_aph_en_both/config.yaml
ADDED
@@ -0,0 +1,286 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
config: conf/tuning/train_asr_ebranchformer_small_wavlm_large1.yaml
|
2 |
+
print_config: false
|
3 |
+
log_level: INFO
|
4 |
+
dry_run: false
|
5 |
+
iterator_type: sequence
|
6 |
+
output_dir: exp/asr_ebranchformer_wavlm_aph_en_both
|
7 |
+
ngpu: 1
|
8 |
+
seed: 2022
|
9 |
+
num_workers: 4
|
10 |
+
num_att_plot: 0
|
11 |
+
dist_backend: nccl
|
12 |
+
dist_init_method: env://
|
13 |
+
dist_world_size: 2
|
14 |
+
dist_rank: 0
|
15 |
+
local_rank: 0
|
16 |
+
dist_master_addr: localhost
|
17 |
+
dist_master_port: 44175
|
18 |
+
dist_launcher: null
|
19 |
+
multiprocessing_distributed: true
|
20 |
+
unused_parameters: true
|
21 |
+
sharded_ddp: false
|
22 |
+
cudnn_enabled: true
|
23 |
+
cudnn_benchmark: false
|
24 |
+
cudnn_deterministic: false
|
25 |
+
collect_stats: false
|
26 |
+
write_collected_feats: false
|
27 |
+
max_epoch: 30
|
28 |
+
patience: null
|
29 |
+
val_scheduler_criterion:
|
30 |
+
- valid
|
31 |
+
- loss
|
32 |
+
early_stopping_criterion:
|
33 |
+
- valid
|
34 |
+
- loss
|
35 |
+
- min
|
36 |
+
best_model_criterion:
|
37 |
+
- - valid
|
38 |
+
- acc
|
39 |
+
- max
|
40 |
+
keep_nbest_models: 10
|
41 |
+
nbest_averaging_interval: 0
|
42 |
+
grad_clip: 5
|
43 |
+
grad_clip_type: 2.0
|
44 |
+
grad_noise: false
|
45 |
+
accum_grad: 8
|
46 |
+
no_forward_run: false
|
47 |
+
resume: true
|
48 |
+
train_dtype: float32
|
49 |
+
use_amp: true
|
50 |
+
log_interval: 200
|
51 |
+
use_matplotlib: true
|
52 |
+
use_tensorboard: true
|
53 |
+
create_graph_in_tensorboard: false
|
54 |
+
use_wandb: false
|
55 |
+
wandb_project: null
|
56 |
+
wandb_id: null
|
57 |
+
wandb_entity: null
|
58 |
+
wandb_name: null
|
59 |
+
wandb_model_log_interval: -1
|
60 |
+
detect_anomaly: false
|
61 |
+
pretrain_path: null
|
62 |
+
init_param: []
|
63 |
+
ignore_init_mismatch: false
|
64 |
+
freeze_param:
|
65 |
+
- frontend.upstream
|
66 |
+
num_iters_per_epoch: null
|
67 |
+
batch_size: 20
|
68 |
+
valid_batch_size: null
|
69 |
+
batch_bins: 6000000
|
70 |
+
valid_batch_bins: null
|
71 |
+
train_shape_file:
|
72 |
+
- exp/asr_stats_raw_en_char_sp/train/speech_shape
|
73 |
+
- exp/asr_stats_raw_en_char_sp/train/text_shape.char
|
74 |
+
valid_shape_file:
|
75 |
+
- exp/asr_stats_raw_en_char_sp/valid/speech_shape
|
76 |
+
- exp/asr_stats_raw_en_char_sp/valid/text_shape.char
|
77 |
+
batch_type: numel
|
78 |
+
valid_batch_type: null
|
79 |
+
fold_length:
|
80 |
+
- 80000
|
81 |
+
- 150
|
82 |
+
sort_in_batch: descending
|
83 |
+
sort_batch: descending
|
84 |
+
multiple_iterator: false
|
85 |
+
chunk_length: 500
|
86 |
+
chunk_shift_ratio: 0.5
|
87 |
+
num_cache_chunks: 1024
|
88 |
+
train_data_path_and_name_and_type:
|
89 |
+
- - dump/raw/train_sp/wav.scp
|
90 |
+
- speech
|
91 |
+
- sound
|
92 |
+
- - dump/raw/train_sp/text
|
93 |
+
- text
|
94 |
+
- text
|
95 |
+
valid_data_path_and_name_and_type:
|
96 |
+
- - dump/raw/val/wav.scp
|
97 |
+
- speech
|
98 |
+
- sound
|
99 |
+
- - dump/raw/val/text
|
100 |
+
- text
|
101 |
+
- text
|
102 |
+
allow_variable_data_keys: false
|
103 |
+
max_cache_size: 0.0
|
104 |
+
max_cache_fd: 32
|
105 |
+
valid_max_cache_size: null
|
106 |
+
exclude_weight_decay: false
|
107 |
+
exclude_weight_decay_conf: {}
|
108 |
+
optim: adam
|
109 |
+
optim_conf:
|
110 |
+
lr: 0.001
|
111 |
+
weight_decay: 1.0e-06
|
112 |
+
scheduler: warmuplr
|
113 |
+
scheduler_conf:
|
114 |
+
warmup_steps: 2500
|
115 |
+
token_list:
|
116 |
+
- <blank>
|
117 |
+
- <unk>
|
118 |
+
- '[APH]'
|
119 |
+
- '[NONAPH]'
|
120 |
+
- <space>
|
121 |
+
- e
|
122 |
+
- t
|
123 |
+
- a
|
124 |
+
- h
|
125 |
+
- o
|
126 |
+
- A
|
127 |
+
- n
|
128 |
+
- '['
|
129 |
+
- P
|
130 |
+
- H
|
131 |
+
- ']'
|
132 |
+
- i
|
133 |
+
- s
|
134 |
+
- N
|
135 |
+
- d
|
136 |
+
- r
|
137 |
+
- u
|
138 |
+
- l
|
139 |
+
- m
|
140 |
+
- w
|
141 |
+
- O
|
142 |
+
- y
|
143 |
+
- g
|
144 |
+
- c
|
145 |
+
- b
|
146 |
+
- f
|
147 |
+
- p
|
148 |
+
- k
|
149 |
+
- ''''
|
150 |
+
- v
|
151 |
+
- j
|
152 |
+
- <
|
153 |
+
- L
|
154 |
+
- U
|
155 |
+
- '>'
|
156 |
+
- ɪ
|
157 |
+
- x
|
158 |
+
- ə
|
159 |
+
- z
|
160 |
+
- ɛ
|
161 |
+
- ɑ
|
162 |
+
- q
|
163 |
+
- ɹ
|
164 |
+
- æ
|
165 |
+
- ˞
|
166 |
+
- ʌ
|
167 |
+
- ʃ
|
168 |
+
- ʊ
|
169 |
+
- ɔ
|
170 |
+
- ŋ
|
171 |
+
- ɚ
|
172 |
+
- ɾ
|
173 |
+
- ʒ
|
174 |
+
- ð
|
175 |
+
- θ
|
176 |
+
- ɜ
|
177 |
+
- ɝ
|
178 |
+
- ɡ
|
179 |
+
- '0'
|
180 |
+
- ː
|
181 |
+
- ʔ
|
182 |
+
- ɒ
|
183 |
+
- é
|
184 |
+
- ɸ
|
185 |
+
- ̩
|
186 |
+
- ʤ
|
187 |
+
- ʧ
|
188 |
+
- <sos/eos>
|
189 |
+
init: null
|
190 |
+
input_size: null
|
191 |
+
ctc_conf:
|
192 |
+
dropout_rate: 0.0
|
193 |
+
ctc_type: builtin
|
194 |
+
reduce: true
|
195 |
+
ignore_nan_grad: null
|
196 |
+
zero_infinity: true
|
197 |
+
joint_net_conf: null
|
198 |
+
use_preprocessor: true
|
199 |
+
token_type: char
|
200 |
+
bpemodel: null
|
201 |
+
non_linguistic_symbols: local/nlsyms.txt
|
202 |
+
cleaner: null
|
203 |
+
g2p: null
|
204 |
+
speech_volume_normalize: null
|
205 |
+
rir_scp: null
|
206 |
+
rir_apply_prob: 1.0
|
207 |
+
noise_scp: null
|
208 |
+
noise_apply_prob: 1.0
|
209 |
+
noise_db_range: '13_15'
|
210 |
+
short_noise_thres: 0.5
|
211 |
+
aux_ctc_tasks: []
|
212 |
+
frontend: s3prl
|
213 |
+
frontend_conf:
|
214 |
+
frontend_conf:
|
215 |
+
upstream: wavlm_large
|
216 |
+
download_dir: ./hub
|
217 |
+
multilayer_feature: true
|
218 |
+
fs: 16k
|
219 |
+
specaug: specaug
|
220 |
+
specaug_conf:
|
221 |
+
apply_time_warp: true
|
222 |
+
time_warp_window: 5
|
223 |
+
time_warp_mode: bicubic
|
224 |
+
apply_freq_mask: true
|
225 |
+
freq_mask_width_range:
|
226 |
+
- 0
|
227 |
+
- 27
|
228 |
+
num_freq_mask: 2
|
229 |
+
apply_time_mask: true
|
230 |
+
time_mask_width_ratio_range:
|
231 |
+
- 0.0
|
232 |
+
- 0.05
|
233 |
+
num_time_mask: 5
|
234 |
+
normalize: utterance_mvn
|
235 |
+
normalize_conf: {}
|
236 |
+
model: espnet
|
237 |
+
model_conf:
|
238 |
+
ctc_weight: 0.3
|
239 |
+
lsm_weight: 0.1
|
240 |
+
length_normalized_loss: false
|
241 |
+
extract_feats_in_collect_stats: false
|
242 |
+
preencoder: linear
|
243 |
+
preencoder_conf:
|
244 |
+
input_size: 1024
|
245 |
+
output_size: 80
|
246 |
+
encoder: e_branchformer
|
247 |
+
encoder_conf:
|
248 |
+
output_size: 256
|
249 |
+
attention_heads: 4
|
250 |
+
linear_units: 1024
|
251 |
+
num_blocks: 12
|
252 |
+
dropout_rate: 0.1
|
253 |
+
positional_dropout_rate: 0.1
|
254 |
+
attention_dropout_rate: 0.1
|
255 |
+
layer_drop_rate: 0.1
|
256 |
+
input_layer: conv2d1
|
257 |
+
macaron_ffn: true
|
258 |
+
pos_enc_layer_type: rel_pos
|
259 |
+
attention_layer_type: rel_selfattn
|
260 |
+
rel_pos_type: latest
|
261 |
+
cgmlp_linear_units: 3072
|
262 |
+
cgmlp_conv_kernel: 31
|
263 |
+
use_linear_after_conv: false
|
264 |
+
gate_activation: identity
|
265 |
+
positionwise_layer_type: linear
|
266 |
+
use_ffn: true
|
267 |
+
merge_conv_kernel: 31
|
268 |
+
postencoder: null
|
269 |
+
postencoder_conf: {}
|
270 |
+
decoder: transformer
|
271 |
+
decoder_conf:
|
272 |
+
attention_heads: 4
|
273 |
+
linear_units: 2048
|
274 |
+
num_blocks: 6
|
275 |
+
dropout_rate: 0.1
|
276 |
+
positional_dropout_rate: 0.1
|
277 |
+
self_attention_dropout_rate: 0.1
|
278 |
+
src_attention_dropout_rate: 0.1
|
279 |
+
layer_drop_rate: 0.2
|
280 |
+
preprocessor: default
|
281 |
+
preprocessor_conf: {}
|
282 |
+
required:
|
283 |
+
- output_dir
|
284 |
+
- token_list
|
285 |
+
version: '202301'
|
286 |
+
distributed: true
|
exp/asr_ebranchformer_wavlm_aph_en_both/images/acc.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/backward_time.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/cer.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/cer_ctc.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/forward_time.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/gpu_max_cached_mem_GB.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/iter_time.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/loss.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/loss_att.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/loss_ctc.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/optim0_lr0.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/optim_step_time.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/train_time.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/images/wer.png
ADDED
exp/asr_ebranchformer_wavlm_aph_en_both/valid.acc.ave_10best.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:cac500148ed6b9f9b0be80e21e56dc57b285b58ce577787d2a3b8daccbaee0dc
|
3 |
+
size 1455789953
|
meta.yaml
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
espnet: '202301'
|
2 |
+
files:
|
3 |
+
asr_model_file: exp/asr_ebranchformer_wavlm_aph_en_both/valid.acc.ave_10best.pth
|
4 |
+
python: "3.9.12 (main, Apr 5 2022, 06:56:58) \n[GCC 7.5.0]"
|
5 |
+
timestamp: 1678208795.72663
|
6 |
+
torch: 1.8.1
|
7 |
+
yaml_files:
|
8 |
+
asr_train_config: exp/asr_ebranchformer_wavlm_aph_en_both/config.yaml
|