Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- rm-sursilv
|
4 |
+
license: apache-2.0
|
5 |
+
tags:
|
6 |
+
- automatic-speech-recognition
|
7 |
+
- mozilla-foundation/common_voice_8_0
|
8 |
+
- generated_from_trainer
|
9 |
+
- rm-sursilv
|
10 |
+
- robust-speech-event
|
11 |
+
- model_for_talk
|
12 |
+
datasets:
|
13 |
+
- mozilla-foundation/common_voice_8_0
|
14 |
+
|
15 |
+
model-index:
|
16 |
+
- name: sammy786/wav2vec2-xlsr-romansh_sursilvan
|
17 |
+
results:
|
18 |
+
- task:
|
19 |
+
name: Automatic Speech Recognition
|
20 |
+
type: automatic-speech-recognition
|
21 |
+
dataset:
|
22 |
+
name: Common Voice 8
|
23 |
+
type: mozilla-foundation/common_voice_8_0
|
24 |
+
args: rm-sursilv
|
25 |
+
metrics:
|
26 |
+
- name: Test WER
|
27 |
+
type: wer
|
28 |
+
value: 13.82
|
29 |
+
- name: Test CER
|
30 |
+
type: cer
|
31 |
+
value: 3.02
|
32 |
+
---
|
33 |
+
# sammy786/wav2vec2-xlsr-romansh_sursilvan
|
34 |
+
|
35 |
+
This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - rm-sursilv dataset.
|
36 |
+
It achieves the following results on evaluation set (which is 10 percent of train data set merged with other and dev datasets):
|
37 |
+
- Loss: 16.38
|
38 |
+
- Wer: 21.25
|
39 |
+
|
40 |
+
## Model description
|
41 |
+
"facebook/wav2vec2-xls-r-1b" was finetuned.
|
42 |
+
|
43 |
+
## Intended uses & limitations
|
44 |
+
More information needed
|
45 |
+
## Training and evaluation data
|
46 |
+
Training data -
|
47 |
+
Common voice Finnish train.tsv, dev.tsv and other.tsv
|
48 |
+
|
49 |
+
## Training procedure
|
50 |
+
For creating the train dataset, all possible datasets were appended and 90-10 split was used.
|
51 |
+
|
52 |
+
### Training hyperparameters
|
53 |
+
|
54 |
+
The following hyperparameters were used during training:
|
55 |
+
|
56 |
+
- learning_rate: 0.000045637994662983496
|
57 |
+
- train_batch_size: 16
|
58 |
+
- eval_batch_size: 16
|
59 |
+
- seed: 13
|
60 |
+
- gradient_accumulation_steps: 2
|
61 |
+
- total_train_batch_size: 32
|
62 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
63 |
+
- lr_scheduler_type: cosine_with_restarts
|
64 |
+
- lr_scheduler_warmup_steps: 500
|
65 |
+
- num_epochs: 15
|
66 |
+
- mixed_precision_training: Native AMP
|
67 |
+
|
68 |
+
|
69 |
+
### Training results
|
70 |
+
|
71 |
+
|
72 |
+
| Step | Training Loss | Validation Loss | Wer |
|
73 |
+
|------|---------------|-----------------|----------|
|
74 |
+
| 200 | 4.825500 | 2.932350 | 1.000000 |
|
75 |
+
| 400 | 1.325600 | 0.292645 | 0.415436 |
|
76 |
+
| 600 | 0.709800 | 0.219167 | 0.324451 |
|
77 |
+
| 800 | 0.576800 | 0.174390 | 0.275477 |
|
78 |
+
| 1000 | 0.538100 | 0.183737 | 0.272116 |
|
79 |
+
| 1200 | 0.475200 | 0.159078 | 0.253871 |
|
80 |
+
| 1400 | 0.420400 | 0.167277 | 0.240907 |
|
81 |
+
| 1600 | 0.393500 | 0.167216 | 0.247269 |
|
82 |
+
| 1800 | 0.407500 | 0.178282 | 0.239827 |
|
83 |
+
| 2000 | 0.374400 | 0.184590 | 0.239467 |
|
84 |
+
| 2200 | 0.382600 | 0.164106 | 0.227824 |
|
85 |
+
| 2400 | 0.363100 | 0.162543 | 0.228544 |
|
86 |
+
| 2600 | 0.199000 | 0.172903 | 0.231665 |
|
87 |
+
| 2800 | 0.150800 | 0.160117 | 0.222662 |
|
88 |
+
| 3000 | 0.101100 | 0.169553 | 0.222662 |
|
89 |
+
| 3200 | 0.104200 | 0.161056 | 0.220622 |
|
90 |
+
| 3400 | 0.096900 | 0.161562 | 0.216781 |
|
91 |
+
| 3600 | 0.092200 | 0.163880 | 0.212580 |
|
92 |
+
| 3800 | 0.089200 | 0.162288 | 0.214140 |
|
93 |
+
| 4000 | 0.076200 | 0.160470 | 0.213540 |
|
94 |
+
| 4200 | 0.087900 | 0.162827 | 0.213060 |
|
95 |
+
| 4400 | 0.066200 | 0.161096 | 0.213300 |
|
96 |
+
| 4600 | 0.076000 | 0.162060 | 0.213660 |
|
97 |
+
| 4800 | 0.071400 | 0.162045 | 0.213300 |
|
98 |
+
|
99 |
+
|
100 |
+
|
101 |
+
### Framework versions
|
102 |
+
- Transformers 4.16.0.dev0
|
103 |
+
- Pytorch 1.10.0+cu102
|
104 |
+
- Datasets 1.17.1.dev0
|
105 |
+
- Tokenizers 0.10.3
|
106 |
+
|
107 |
+
#### Evaluation Commands
|
108 |
+
|
109 |
+
1. To evaluate on `mozilla-foundation/common_voice_8_0` with split `test`
|
110 |
+
|
111 |
+
```bash
|
112 |
+
python eval.py --model_id sammy786/wav2vec2-xlsr-romansh_sursilvan --dataset mozilla-foundation/common_voice_8_0 --config rm-sursilv --split test
|
113 |
+
```
|