aapot commited on
Commit
84d20ee
1 Parent(s): 76703b0

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +227 -207
README.md CHANGED
@@ -1,208 +1,228 @@
1
- ---
2
- license: apache-2.0
3
- language: fi
4
- metrics:
5
- - wer
6
- - cer
7
- tags:
8
- - automatic-speech-recognition
9
- - fi
10
- - finnish
11
- - generated_from_trainer
12
- - hf-asr-leaderboard
13
- - robust-speech-event
14
- datasets:
15
- - mozilla-foundation/common_voice_7_0
16
- model-index:
17
- - name: wav2vec2-xlsr-1b-finnish-lm-v2
18
- results:
19
- - task:
20
- name: Automatic Speech Recognition
21
- type: automatic-speech-recognition
22
- dataset:
23
- name: Common Voice 7
24
- type: mozilla-foundation/common_voice_7_0
25
- args: fi
26
- metrics:
27
- - name: Test WER
28
- type: wer
29
- value: 4.09
30
- - name: Test CER
31
- type: cer
32
- value: 0.88
33
- ---
34
-
35
- # Wav2Vec2 XLS-R for Finnish ASR
36
-
37
- This acoustic model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) for Finnish ASR. The model has been fine-tuned with 275.6 hours of Finnish transcribed speech data. Wav2Vec2 XLS-R was introduced in
38
- [this paper](https://arxiv.org/abs/2111.09296) and first released at [this page](https://github.com/pytorch/fairseq/tree/main/examples/wav2vec#wav2vec-20).
39
-
40
- This repository also includes Finnish KenLM language model used in the decoding phase with the acoustic model.
41
-
42
- **Note**: this model is exactly the same as the [aapot/wav2vec2-xlsr-1b-finnish-lm-v2](https://huggingface.co/aapot/wav2vec2-xlsr-1b-finnish-lm-v2) model so that model has just been copied/moved to this `Finnish-NLP` Hugging Face organization.
43
-
44
- ## Model description
45
-
46
- Wav2Vec2 XLS-R is Facebook AI's large-scale multilingual pretrained model for speech. It is pretrained on 436k hours of unlabeled speech, including VoxPopuli, MLS, CommonVoice, BABEL, and VoxLingua107. It uses the wav2vec 2.0 objective, in 128 languages.
47
-
48
- You can read more about the pretrained model from [this blog](https://ai.facebook.com/blog/xls-r-self-supervised-speech-processing-for-128-languages) and [this paper](https://arxiv.org/abs/2111.09296).
49
-
50
- This model is fine-tuned version of the pretrained model (1 billion parameter variant) for Finnish ASR.
51
-
52
- ## Intended uses & limitations
53
-
54
- You can use this model for Finnish ASR (speech-to-text) task.
55
-
56
- ### How to use
57
-
58
- Check the [run-finnish-asr-models.ipynb](https://huggingface.co/Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2/blob/main/run-finnish-asr-models.ipynb) notebook in this repository for an detailed example on how to use this model.
59
-
60
- ### Limitations and bias
61
-
62
- This model was fine-tuned with audio samples which maximum length was 20 seconds so this model most likely works the best for quite short audios of similar length. However, you can try this model with a lot longer audios too and see how it works. If you encounter out of memory errors with very long audio files you can use the audio chunking method introduced in [this blog post](https://huggingface.co/blog/asr-chunking).
63
-
64
- A vast majority of the data used for fine-tuning was from the Finnish Parliament dataset so this model may not generalize so well to very different domains like common daily spoken Finnish with dialects etc. In addition, audios of the datasets tend to be adult male dominated so this model may not work as well for speeches of children and women, for example.
65
-
66
- The Finnish KenLM language model used in the decoding phase has been trained with text data from the audio transcriptions and from a subset of Finnish Wikipedia. Thus, the decoder's language model may not generalize to very different language, for example to spoken daily language with dialects (because especially the Wikipedia contains mostly formal Finnish language). It may be beneficial to train your own KenLM language model for your domain language and use that in the decoding.
67
-
68
- ## Training data
69
-
70
- This model was fine-tuned with 275.6 hours of Finnish transcribed speech data from following datasets:
71
-
72
- | Dataset | Hours | % of total hours |
73
- |:------------------------------------------------------------------------------------------------------------------------------ |:--------:|:----------------:|
74
- | [Common Voice 7.0 Finnish train + evaluation + other splits](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) | 9.70 h | 3.52 % |
75
- | [Finnish parliament session 2](https://b2share.eudat.eu/records/4df422d631544ce682d6af1d4714b2d4) | 0.24 h | 0.09 % |
76
- | [VoxPopuli Finnish](https://github.com/facebookresearch/voxpopuli) | 21.97 h | 7.97 % |
77
- | [CSS10 Finnish](https://github.com/kyubyong/css10) | 10.32 h | 3.74 % |
78
- | [Aalto Finnish Parliament ASR Corpus](http://urn.fi/urn:nbn:fi:lb-2021051903) | 228.00 h | 82.73 % |
79
- | [Finnish Broadcast Corpus](http://urn.fi/urn:nbn:fi:lb-2016042502) | 5.37 h | 1.95 % |
80
-
81
- Datasets were filtered to include maximum length of 20 seconds long audio samples.
82
-
83
- ## Training procedure
84
-
85
- This model was trained during [Robust Speech Challenge Event](https://discuss.huggingface.co/t/open-to-the-community-robust-speech-recognition-challenge/13614) organized by Hugging Face. Training was done on a Tesla V100 GPU, sponsored by OVHcloud.
86
-
87
- Training script was provided by Hugging Face and it is available [here](https://github.com/huggingface/transformers/blob/main/examples/research_projects/robust-speech-event/run_speech_recognition_ctc_bnb.py). We only modified its data loading for our custom datasets.
88
-
89
- For the KenLM language model training, we followed the [blog post tutorial](https://huggingface.co/blog/wav2vec2-with-ngram) provided by Hugging Face. Training data for the 5-gram KenLM were text transcriptions of the audio training data and 100k random samples of cleaned [Finnish Wikipedia](https://huggingface.co/datasets/wikipedia) (August 2021) dataset.
90
-
91
- ### Training hyperparameters
92
-
93
- The following hyperparameters were used during training:
94
- - learning_rate: 5e-05
95
- - train_batch_size: 32
96
- - eval_batch_size: 8
97
- - seed: 42
98
- - optimizer: [8-bit Adam](https://github.com/facebookresearch/bitsandbytes) with betas=(0.9,0.999) and epsilon=1e-08
99
- - lr_scheduler_type: linear
100
- - lr_scheduler_warmup_steps: 500
101
- - num_epochs: 10
102
- - mixed_precision_training: Native AMP
103
-
104
- The pretrained `facebook/wav2vec2-xls-r-1b` model was initialized with following hyperparameters:
105
- - attention_dropout: 0.094
106
- - hidden_dropout: 0.047
107
- - feat_proj_dropout: 0.04
108
- - mask_time_prob: 0.082
109
- - layerdrop: 0.041
110
- - activation_dropout: 0.055
111
- - ctc_loss_reduction: "mean"
112
-
113
- ### Training results
114
-
115
- | Training Loss | Epoch | Step | Validation Loss | Wer |
116
- |:-------------:|:-----:|:-----:|:---------------:|:------:|
117
- | 0.7778 | 0.17 | 500 | 0.2851 | 0.3572 |
118
- | 0.5506 | 0.34 | 1000 | 0.1595 | 0.2130 |
119
- | 0.6569 | 0.5 | 1500 | 0.1458 | 0.2046 |
120
- | 0.5997 | 0.67 | 2000 | 0.1374 | 0.1975 |
121
- | 0.542 | 0.84 | 2500 | 0.1390 | 0.1956 |
122
- | 0.4815 | 1.01 | 3000 | 0.1266 | 0.1813 |
123
- | 0.6982 | 1.17 | 3500 | 0.1441 | 0.1965 |
124
- | 0.4522 | 1.34 | 4000 | 0.1232 | 0.1822 |
125
- | 0.4655 | 1.51 | 4500 | 0.1209 | 0.1702 |
126
- | 0.4069 | 1.68 | 5000 | 0.1149 | 0.1688 |
127
- | 0.4226 | 1.84 | 5500 | 0.1121 | 0.1560 |
128
- | 0.3993 | 2.01 | 6000 | 0.1091 | 0.1557 |
129
- | 0.406 | 2.18 | 6500 | 0.1115 | 0.1553 |
130
- | 0.4098 | 2.35 | 7000 | 0.1144 | 0.1560 |
131
- | 0.3995 | 2.51 | 7500 | 0.1028 | 0.1476 |
132
- | 0.4101 | 2.68 | 8000 | 0.1129 | 0.1511 |
133
- | 0.3636 | 2.85 | 8500 | 0.1025 | 0.1517 |
134
- | 0.3534 | 3.02 | 9000 | 0.1068 | 0.1480 |
135
- | 0.3836 | 3.18 | 9500 | 0.1072 | 0.1459 |
136
- | 0.3531 | 3.35 | 10000 | 0.0928 | 0.1367 |
137
- | 0.3649 | 3.52 | 10500 | 0.1042 | 0.1426 |
138
- | 0.3645 | 3.69 | 11000 | 0.0979 | 0.1433 |
139
- | 0.3685 | 3.85 | 11500 | 0.0947 | 0.1346 |
140
- | 0.3325 | 4.02 | 12000 | 0.0991 | 0.1352 |
141
- | 0.3497 | 4.19 | 12500 | 0.0919 | 0.1358 |
142
- | 0.3303 | 4.36 | 13000 | 0.0888 | 0.1272 |
143
- | 0.3323 | 4.52 | 13500 | 0.0888 | 0.1277 |
144
- | 0.3452 | 4.69 | 14000 | 0.0894 | 0.1279 |
145
- | 0.337 | 4.86 | 14500 | 0.0917 | 0.1289 |
146
- | 0.3114 | 5.03 | 15000 | 0.0942 | 0.1313 |
147
- | 0.3099 | 5.19 | 15500 | 0.0902 | 0.1239 |
148
- | 0.3079 | 5.36 | 16000 | 0.0871 | 0.1256 |
149
- | 0.3293 | 5.53 | 16500 | 0.0861 | 0.1263 |
150
- | 0.3123 | 5.7 | 17000 | 0.0876 | 0.1203 |
151
- | 0.3093 | 5.86 | 17500 | 0.0848 | 0.1226 |
152
- | 0.2903 | 6.03 | 18000 | 0.0914 | 0.1221 |
153
- | 0.297 | 6.2 | 18500 | 0.0841 | 0.1185 |
154
- | 0.2797 | 6.37 | 19000 | 0.0858 | 0.1165 |
155
- | 0.2878 | 6.53 | 19500 | 0.0874 | 0.1161 |
156
- | 0.2974 | 6.7 | 20000 | 0.0835 | 0.1173 |
157
- | 0.3051 | 6.87 | 20500 | 0.0835 | 0.1178 |
158
- | 0.2941 | 7.04 | 21000 | 0.0852 | 0.1155 |
159
- | 0.258 | 7.21 | 21500 | 0.0832 | 0.1132 |
160
- | 0.2778 | 7.37 | 22000 | 0.0829 | 0.1110 |
161
- | 0.2751 | 7.54 | 22500 | 0.0822 | 0.1069 |
162
- | 0.2887 | 7.71 | 23000 | 0.0819 | 0.1103 |
163
- | 0.2509 | 7.88 | 23500 | 0.0787 | 0.1055 |
164
- | 0.2501 | 8.04 | 24000 | 0.0807 | 0.1076 |
165
- | 0.2399 | 8.21 | 24500 | 0.0784 | 0.1052 |
166
- | 0.2539 | 8.38 | 25000 | 0.0772 | 0.1075 |
167
- | 0.248 | 8.55 | 25500 | 0.0772 | 0.1055 |
168
- | 0.2689 | 8.71 | 26000 | 0.0763 | 0.1027 |
169
- | 0.2855 | 8.88 | 26500 | 0.0756 | 0.1035 |
170
- | 0.2421 | 9.05 | 27000 | 0.0771 | 0.0998 |
171
- | 0.2497 | 9.22 | 27500 | 0.0756 | 0.0971 |
172
- | 0.2367 | 9.38 | 28000 | 0.0741 | 0.0974 |
173
- | 0.2473 | 9.55 | 28500 | 0.0739 | 0.0982 |
174
- | 0.2396 | 9.72 | 29000 | 0.0756 | 0.0991 |
175
- | 0.2602 | 9.89 | 29500 | 0.0737 | 0.0975 |
176
-
177
-
178
- ### Framework versions
179
-
180
- - Transformers 4.17.0.dev0
181
- - Pytorch 1.10.2+cu102
182
- - Datasets 1.18.3
183
- - Tokenizers 0.11.0
184
-
185
- ## Evaluation results
186
-
187
- Evaluation was done with the [Common Voice 7.0 Finnish test split](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0).
188
-
189
- To evaluate this model, run the `eval.py` script in this repository:
190
-
191
- ```bash
192
- python3 eval.py --model_id Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2 --dataset mozilla-foundation/common_voice_7_0 --config fi --split test
193
- ```
194
-
195
- This model (the first row of the table) achieves the following WER (Word Error Rate) and CER (Character Error Rate) results compared to our other models:
196
-
197
- | | WER (with LM) | WER (without LM) | CER (with LM) | CER (without LM) |
198
- |-----------------------------------------------|---------------|------------------|---------------|------------------|
199
- |Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2 |**4.09** |**9.73** |**0.88** |**1.65** |
200
- |Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm |5.65 |13.11 |1.20 |2.23 |
201
- |Finnish-NLP/wav2vec2-xlsr-300m-finnish-lm |8.16 |17.92 |1.97 |3.36 |
202
-
203
- ## Team Members
204
-
205
- - Aapo Tanskanen, [Hugging Face profile](https://huggingface.co/aapot), [LinkedIn profile](https://www.linkedin.com/in/aapotanskanen/)
206
- - Rasmus Toivanen, [Hugging Face profile](https://huggingface.co/RASMUS), [LinkedIn profile](https://www.linkedin.com/in/rasmustoivanen/)
207
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
208
  Feel free to contact us for more details 🤗
 
1
+ ---
2
+ license: apache-2.0
3
+ language: fi
4
+ metrics:
5
+ - wer
6
+ - cer
7
+ tags:
8
+ - automatic-speech-recognition
9
+ - fi
10
+ - finnish
11
+ - generated_from_trainer
12
+ - hf-asr-leaderboard
13
+ - robust-speech-event
14
+ datasets:
15
+ - mozilla-foundation/common_voice_7_0
16
+ model-index:
17
+ - name: wav2vec2-xlsr-1b-finnish-lm-v2
18
+ results:
19
+ - task:
20
+ name: Automatic Speech Recognition
21
+ type: automatic-speech-recognition
22
+ dataset:
23
+ name: Common Voice 7
24
+ type: mozilla-foundation/common_voice_7_0
25
+ args: fi
26
+ metrics:
27
+ - name: Test WER
28
+ type: wer
29
+ value: 4.09
30
+ - name: Test CER
31
+ type: cer
32
+ value: 0.88
33
+ ---
34
+
35
+ # Wav2Vec2 XLS-R for Finnish ASR
36
+
37
+ This acoustic model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) for Finnish ASR. The model has been fine-tuned with 275.6 hours of Finnish transcribed speech data. Wav2Vec2 XLS-R was introduced in
38
+ [this paper](https://arxiv.org/abs/2111.09296) and first released at [this page](https://github.com/pytorch/fairseq/tree/main/examples/wav2vec#wav2vec-20).
39
+
40
+ This repository also includes Finnish KenLM language model used in the decoding phase with the acoustic model.
41
+
42
+ **Note**: this model is exactly the same as the [aapot/wav2vec2-xlsr-1b-finnish-lm-v2](https://huggingface.co/aapot/wav2vec2-xlsr-1b-finnish-lm-v2) model so that model has just been copied/moved to this `Finnish-NLP` Hugging Face organization.
43
+
44
+ ## Model description
45
+
46
+ Wav2Vec2 XLS-R is Facebook AI's large-scale multilingual pretrained model for speech. It is pretrained on 436k hours of unlabeled speech, including VoxPopuli, MLS, CommonVoice, BABEL, and VoxLingua107. It uses the wav2vec 2.0 objective, in 128 languages.
47
+
48
+ You can read more about the pretrained model from [this blog](https://ai.facebook.com/blog/xls-r-self-supervised-speech-processing-for-128-languages) and [this paper](https://arxiv.org/abs/2111.09296).
49
+
50
+ This model is fine-tuned version of the pretrained model (1 billion parameter variant) for Finnish ASR.
51
+
52
+ ## Intended uses & limitations
53
+
54
+ You can use this model for Finnish ASR (speech-to-text) task.
55
+
56
+ ### How to use
57
+
58
+ Check the [run-finnish-asr-models.ipynb](https://huggingface.co/Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2/blob/main/run-finnish-asr-models.ipynb) notebook in this repository for an detailed example on how to use this model.
59
+
60
+ ### Limitations and bias
61
+
62
+ This model was fine-tuned with audio samples which maximum length was 20 seconds so this model most likely works the best for quite short audios of similar length. However, you can try this model with a lot longer audios too and see how it works. If you encounter out of memory errors with very long audio files you can use the audio chunking method introduced in [this blog post](https://huggingface.co/blog/asr-chunking).
63
+
64
+ A vast majority of the data used for fine-tuning was from the Finnish Parliament dataset so this model may not generalize so well to very different domains like common daily spoken Finnish with dialects etc. In addition, audios of the datasets tend to be adult male dominated so this model may not work as well for speeches of children and women, for example.
65
+
66
+ The Finnish KenLM language model used in the decoding phase has been trained with text data from the audio transcriptions and from a subset of Finnish Wikipedia. Thus, the decoder's language model may not generalize to very different language, for example to spoken daily language with dialects (because especially the Wikipedia contains mostly formal Finnish language). It may be beneficial to train your own KenLM language model for your domain language and use that in the decoding.
67
+
68
+ ## Training data
69
+
70
+ This model was fine-tuned with 275.6 hours of Finnish transcribed speech data from following datasets:
71
+
72
+ | Dataset | Hours | % of total hours |
73
+ |:------------------------------------------------------------------------------------------------------------------------------ |:--------:|:----------------:|
74
+ | [Common Voice 7.0 Finnish train + evaluation + other splits](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) | 9.70 h | 3.52 % |
75
+ | [Finnish parliament session 2](https://b2share.eudat.eu/records/4df422d631544ce682d6af1d4714b2d4) | 0.24 h | 0.09 % |
76
+ | [VoxPopuli Finnish](https://github.com/facebookresearch/voxpopuli) | 21.97 h | 7.97 % |
77
+ | [CSS10 Finnish](https://github.com/kyubyong/css10) | 10.32 h | 3.74 % |
78
+ | [Aalto Finnish Parliament ASR Corpus](http://urn.fi/urn:nbn:fi:lb-2021051903) | 228.00 h | 82.73 % |
79
+ | [Finnish Broadcast Corpus](http://urn.fi/urn:nbn:fi:lb-2016042502) | 5.37 h | 1.95 % |
80
+
81
+ Datasets were filtered to include maximum length of 20 seconds long audio samples.
82
+
83
+ ## Training procedure
84
+
85
+ This model was trained during [Robust Speech Challenge Event](https://discuss.huggingface.co/t/open-to-the-community-robust-speech-recognition-challenge/13614) organized by Hugging Face. Training was done on a Tesla V100 GPU, sponsored by OVHcloud.
86
+
87
+ Training script was provided by Hugging Face and it is available [here](https://github.com/huggingface/transformers/blob/main/examples/research_projects/robust-speech-event/run_speech_recognition_ctc_bnb.py). We only modified its data loading for our custom datasets.
88
+
89
+ For the KenLM language model training, we followed the [blog post tutorial](https://huggingface.co/blog/wav2vec2-with-ngram) provided by Hugging Face. Training data for the 5-gram KenLM were text transcriptions of the audio training data and 100k random samples of cleaned [Finnish Wikipedia](https://huggingface.co/datasets/wikipedia) (August 2021) dataset.
90
+
91
+ ### Training hyperparameters
92
+
93
+ The following hyperparameters were used during training:
94
+ - learning_rate: 5e-05
95
+ - train_batch_size: 32
96
+ - eval_batch_size: 8
97
+ - seed: 42
98
+ - optimizer: [8-bit Adam](https://github.com/facebookresearch/bitsandbytes) with betas=(0.9,0.999) and epsilon=1e-08
99
+ - lr_scheduler_type: linear
100
+ - lr_scheduler_warmup_steps: 500
101
+ - num_epochs: 10
102
+ - mixed_precision_training: Native AMP
103
+
104
+ The pretrained `facebook/wav2vec2-xls-r-1b` model was initialized with following hyperparameters:
105
+ - attention_dropout: 0.094
106
+ - hidden_dropout: 0.047
107
+ - feat_proj_dropout: 0.04
108
+ - mask_time_prob: 0.082
109
+ - layerdrop: 0.041
110
+ - activation_dropout: 0.055
111
+ - ctc_loss_reduction: "mean"
112
+
113
+ ### Training results
114
+
115
+ | Training Loss | Epoch | Step | Validation Loss | Wer |
116
+ |:-------------:|:-----:|:-----:|:---------------:|:------:|
117
+ | 0.7778 | 0.17 | 500 | 0.2851 | 0.3572 |
118
+ | 0.5506 | 0.34 | 1000 | 0.1595 | 0.2130 |
119
+ | 0.6569 | 0.5 | 1500 | 0.1458 | 0.2046 |
120
+ | 0.5997 | 0.67 | 2000 | 0.1374 | 0.1975 |
121
+ | 0.542 | 0.84 | 2500 | 0.1390 | 0.1956 |
122
+ | 0.4815 | 1.01 | 3000 | 0.1266 | 0.1813 |
123
+ | 0.6982 | 1.17 | 3500 | 0.1441 | 0.1965 |
124
+ | 0.4522 | 1.34 | 4000 | 0.1232 | 0.1822 |
125
+ | 0.4655 | 1.51 | 4500 | 0.1209 | 0.1702 |
126
+ | 0.4069 | 1.68 | 5000 | 0.1149 | 0.1688 |
127
+ | 0.4226 | 1.84 | 5500 | 0.1121 | 0.1560 |
128
+ | 0.3993 | 2.01 | 6000 | 0.1091 | 0.1557 |
129
+ | 0.406 | 2.18 | 6500 | 0.1115 | 0.1553 |
130
+ | 0.4098 | 2.35 | 7000 | 0.1144 | 0.1560 |
131
+ | 0.3995 | 2.51 | 7500 | 0.1028 | 0.1476 |
132
+ | 0.4101 | 2.68 | 8000 | 0.1129 | 0.1511 |
133
+ | 0.3636 | 2.85 | 8500 | 0.1025 | 0.1517 |
134
+ | 0.3534 | 3.02 | 9000 | 0.1068 | 0.1480 |
135
+ | 0.3836 | 3.18 | 9500 | 0.1072 | 0.1459 |
136
+ | 0.3531 | 3.35 | 10000 | 0.0928 | 0.1367 |
137
+ | 0.3649 | 3.52 | 10500 | 0.1042 | 0.1426 |
138
+ | 0.3645 | 3.69 | 11000 | 0.0979 | 0.1433 |
139
+ | 0.3685 | 3.85 | 11500 | 0.0947 | 0.1346 |
140
+ | 0.3325 | 4.02 | 12000 | 0.0991 | 0.1352 |
141
+ | 0.3497 | 4.19 | 12500 | 0.0919 | 0.1358 |
142
+ | 0.3303 | 4.36 | 13000 | 0.0888 | 0.1272 |
143
+ | 0.3323 | 4.52 | 13500 | 0.0888 | 0.1277 |
144
+ | 0.3452 | 4.69 | 14000 | 0.0894 | 0.1279 |
145
+ | 0.337 | 4.86 | 14500 | 0.0917 | 0.1289 |
146
+ | 0.3114 | 5.03 | 15000 | 0.0942 | 0.1313 |
147
+ | 0.3099 | 5.19 | 15500 | 0.0902 | 0.1239 |
148
+ | 0.3079 | 5.36 | 16000 | 0.0871 | 0.1256 |
149
+ | 0.3293 | 5.53 | 16500 | 0.0861 | 0.1263 |
150
+ | 0.3123 | 5.7 | 17000 | 0.0876 | 0.1203 |
151
+ | 0.3093 | 5.86 | 17500 | 0.0848 | 0.1226 |
152
+ | 0.2903 | 6.03 | 18000 | 0.0914 | 0.1221 |
153
+ | 0.297 | 6.2 | 18500 | 0.0841 | 0.1185 |
154
+ | 0.2797 | 6.37 | 19000 | 0.0858 | 0.1165 |
155
+ | 0.2878 | 6.53 | 19500 | 0.0874 | 0.1161 |
156
+ | 0.2974 | 6.7 | 20000 | 0.0835 | 0.1173 |
157
+ | 0.3051 | 6.87 | 20500 | 0.0835 | 0.1178 |
158
+ | 0.2941 | 7.04 | 21000 | 0.0852 | 0.1155 |
159
+ | 0.258 | 7.21 | 21500 | 0.0832 | 0.1132 |
160
+ | 0.2778 | 7.37 | 22000 | 0.0829 | 0.1110 |
161
+ | 0.2751 | 7.54 | 22500 | 0.0822 | 0.1069 |
162
+ | 0.2887 | 7.71 | 23000 | 0.0819 | 0.1103 |
163
+ | 0.2509 | 7.88 | 23500 | 0.0787 | 0.1055 |
164
+ | 0.2501 | 8.04 | 24000 | 0.0807 | 0.1076 |
165
+ | 0.2399 | 8.21 | 24500 | 0.0784 | 0.1052 |
166
+ | 0.2539 | 8.38 | 25000 | 0.0772 | 0.1075 |
167
+ | 0.248 | 8.55 | 25500 | 0.0772 | 0.1055 |
168
+ | 0.2689 | 8.71 | 26000 | 0.0763 | 0.1027 |
169
+ | 0.2855 | 8.88 | 26500 | 0.0756 | 0.1035 |
170
+ | 0.2421 | 9.05 | 27000 | 0.0771 | 0.0998 |
171
+ | 0.2497 | 9.22 | 27500 | 0.0756 | 0.0971 |
172
+ | 0.2367 | 9.38 | 28000 | 0.0741 | 0.0974 |
173
+ | 0.2473 | 9.55 | 28500 | 0.0739 | 0.0982 |
174
+ | 0.2396 | 9.72 | 29000 | 0.0756 | 0.0991 |
175
+ | 0.2602 | 9.89 | 29500 | 0.0737 | 0.0975 |
176
+
177
+
178
+ ### Framework versions
179
+
180
+ - Transformers 4.17.0.dev0
181
+ - Pytorch 1.10.2+cu102
182
+ - Datasets 1.18.3
183
+ - Tokenizers 0.11.0
184
+
185
+ ## Evaluation results
186
+
187
+ Evaluation was done with the [Common Voice 7.0 Finnish test split](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) and with the [Common Voice 9.0 Finnish test split](https://huggingface.co/datasets/mozilla-foundation/common_voice_9_0). This model's training data includes the training splits of Common Voice 7.0 but our newest `Finnish-NLP/wav2vec2-base-fi-voxpopuli-v2-finetuned` model includes the Common Voice 9.0 so we ran tests for both versions. Note: Common Voice doesn't seem to fully preserve the test split as fixed between the dataset versions so it is possible that some of the training examples of Common Voice 9.0 are in the test split of the Common Voice 7.0 and vice versa. Thus, test result comparisons are not fully accurate between the models trained with different Common Voice versions but the comparison should still be meaningful enough.
188
+
189
+ ### Common Voice 7.0 testing
190
+
191
+ To evaluate this model, run the `eval.py` script in this repository:
192
+
193
+ ```bash
194
+ python3 eval.py --model_id Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2 --dataset mozilla-foundation/common_voice_7_0 --config fi --split test
195
+ ```
196
+
197
+ This model (the first row of the table) achieves the following WER (Word Error Rate) and CER (Character Error Rate) results compared to our other models and their parameter counts:
198
+
199
+ | | Model parameters | WER (with LM) | WER (without LM) | CER (with LM) | CER (without LM) |
200
+ |----------------------------------------------------|------------------|---------------|------------------|---------------|------------------|
201
+ |Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2 | 1000 million |**4.09** |**9.73** |**0.88** |**1.65** |
202
+ |Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm | 1000 million |5.65 |13.11 |1.20 |2.23 |
203
+ |Finnish-NLP/wav2vec2-base-fi-voxpopuli-v2-finetuned | 95 million |5.85 |13.52 |1.35 |2.44 |
204
+ |Finnish-NLP/wav2vec2-xlsr-300m-finnish-lm | 300 million |8.16 |17.92 |1.97 |3.36 |
205
+
206
+ ### Common Voice 9.0 testing
207
+
208
+ To evaluate this model, run the `eval.py` script in this repository:
209
+
210
+ ```bash
211
+ python3 eval.py --model_id Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2 --dataset mozilla-foundation/common_voice_9_0 --config fi --split test
212
+ ```
213
+
214
+ This model (the first row of the table) achieves the following WER (Word Error Rate) and CER (Character Error Rate) results compared to our other models and their parameter counts:
215
+
216
+ | | Model parameters | WER (with LM) | WER (without LM) | CER (with LM) | CER (without LM) |
217
+ |----------------------------------------------------|------------------|---------------|------------------|---------------|------------------|
218
+ |Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2 | 1000 million |**3.72** |**8.96** |**0.80** |**1.52** |
219
+ |Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm | 1000 million |5.35 |13.00 |1.14 |2.20 |
220
+ |Finnish-NLP/wav2vec2-base-fi-voxpopuli-v2-finetuned | 95 million |5.93 |14.08 |1.40 |2.59 |
221
+ |Finnish-NLP/wav2vec2-xlsr-300m-finnish-lm | 300 million |7.42 |16.45 |1.79 |3.07 |
222
+
223
+ ## Team Members
224
+
225
+ - Aapo Tanskanen, [Hugging Face profile](https://huggingface.co/aapot), [LinkedIn profile](https://www.linkedin.com/in/aapotanskanen/)
226
+ - Rasmus Toivanen, [Hugging Face profile](https://huggingface.co/RASMUS), [LinkedIn profile](https://www.linkedin.com/in/rasmustoivanen/)
227
+
228
  Feel free to contact us for more details 🤗