aapot commited on
Commit
55587f6
1 Parent(s): 0292fff

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +69 -20
README.md CHANGED
@@ -32,41 +32,56 @@ model-index:
32
  value: 1.65
33
  ---
34
 
35
- <!-- This model card has been generated automatically according to the information the Trainer had access to. You
36
- should probably proofread and complete it, then remove this comment. -->
37
 
38
- # wav2vec2-xlsr-1b-finnish-v2
 
39
 
40
- **Note**: there is a language model version of this acoustic only model which achieves better test results thanks to the added language model: [aapot/wav2vec2-xlsr-1b-finnish-lm-v2](https://huggingface.co/aapot/wav2vec2-xlsr-1b-finnish-lm-v2)
41
-
42
- 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.
43
- It achieves the following results on the Common Voice 7 test set without language model:
44
- - Wer: 9.73
45
- - Cer: 1.65
46
 
47
  ## Model description
48
 
49
- TODO
 
 
 
 
50
 
51
  ## Intended uses & limitations
52
 
53
- TODO
 
 
 
 
 
 
54
 
55
- ## Training and evaluation data
 
 
 
 
56
 
57
  This model was fine-tuned with 275.6 hours of Finnish transcribed speech data from following datasets:
58
 
59
- | Dataset | Hours | % of total hours |
60
- |:------------------------------------------------------------------------------------------------------------------------------|:--------:|:----------------:|
61
- | [Common Voice 7.0 Finnish train+evaluation+other splits](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) | 9.70 h | 3.52 % |
62
- | [Finnish parliament session 2](https://b2share.eudat.eu/records/4df422d631544ce682d6af1d4714b2d4) | 0.24 h | 0.09 % |
63
- | [VoxPopuli Finnish](https://github.com/facebookresearch/voxpopuli) | 21.97 h | 7.97 % |
64
- | [CSS10 Finnish](https://github.com/kyubyong/css10) | 10.32 h | 3.74 % |
65
- | [Aalto Finnish Parliament ASR Corpus](http://urn.fi/urn:nbn:fi:lb-2021051903) | 228.00 h | 82.73 % |
66
- | [Finnish Broadcast Corpus](http://urn.fi/urn:nbn:fi:lb-2016042502) | 5.37 h | 1.95 % |
 
 
67
 
68
  ## Training procedure
69
 
 
 
 
 
70
  ### Training hyperparameters
71
 
72
  The following hyperparameters were used during training:
@@ -80,6 +95,15 @@ The following hyperparameters were used during training:
80
  - num_epochs: 10
81
  - mixed_precision_training: Native AMP
82
 
 
 
 
 
 
 
 
 
 
83
  ### Training results
84
 
85
  | Training Loss | Epoch | Step | Validation Loss | Wer |
@@ -151,3 +175,28 @@ The following hyperparameters were used during training:
151
  - Pytorch 1.10.2+cu102
152
  - Datasets 1.18.3
153
  - Tokenizers 0.11.0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
32
  value: 1.65
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
+ **Note**: there is a version with KenLM language model used in the decoding phase producing better transcriptions: [Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2](https://huggingface.co/Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2)
 
 
 
 
 
41
 
42
  ## Model description
43
 
44
+ 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.
45
+
46
+ 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).
47
+
48
+ This model is fine-tuned version of the pretrained model (1 billion parameter variant) for Finnish ASR.
49
 
50
  ## Intended uses & limitations
51
 
52
+ You can use this model for Finnish ASR (speech-to-text) task.
53
+
54
+ ### How to use
55
+
56
+ Check the [run-finnish-asr-models.ipynb](https://huggingface.co/aapot/wav2vec2-xlsr-1b-finnish-v2/blob/main/run-finnish-asr-models.ipynb) notebook in this repository for an detailed example on how to use this model.
57
+
58
+ ### Limitations and bias
59
 
60
+ 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).
61
+
62
+ 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.
63
+
64
+ ## Training data
65
 
66
  This model was fine-tuned with 275.6 hours of Finnish transcribed speech data from following datasets:
67
 
68
+ | Dataset | Hours | % of total hours |
69
+ |:------------------------------------------------------------------------------------------------------------------------------ |:--------:|:----------------:|
70
+ | [Common Voice 7.0 Finnish train + evaluation + other splits](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) | 9.70 h | 3.52 % |
71
+ | [Finnish parliament session 2](https://b2share.eudat.eu/records/4df422d631544ce682d6af1d4714b2d4) | 0.24 h | 0.09 % |
72
+ | [VoxPopuli Finnish](https://github.com/facebookresearch/voxpopuli) | 21.97 h | 7.97 % |
73
+ | [CSS10 Finnish](https://github.com/kyubyong/css10) | 10.32 h | 3.74 % |
74
+ | [Aalto Finnish Parliament ASR Corpus](http://urn.fi/urn:nbn:fi:lb-2021051903) | 228.00 h | 82.73 % |
75
+ | [Finnish Broadcast Corpus](http://urn.fi/urn:nbn:fi:lb-2016042502) | 5.37 h | 1.95 % |
76
+
77
+ Datasets were filtered to include maximum length of 20 seconds long audio samples.
78
 
79
  ## Training procedure
80
 
81
+ 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.
82
+
83
+ 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.
84
+
85
  ### Training hyperparameters
86
 
87
  The following hyperparameters were used during training:
 
95
  - num_epochs: 10
96
  - mixed_precision_training: Native AMP
97
 
98
+ The pretrained `facebook/wav2vec2-xls-r-1b` model was initialized with following hyperparameters:
99
+ - attention_dropout: 0.094
100
+ - hidden_dropout: 0.047
101
+ - feat_proj_dropout: 0.04
102
+ - mask_time_prob: 0.082
103
+ - layerdrop: 0.041
104
+ - activation_dropout: 0.055
105
+ - ctc_loss_reduction: "mean"
106
+
107
  ### Training results
108
 
109
  | Training Loss | Epoch | Step | Validation Loss | Wer |
 
175
  - Pytorch 1.10.2+cu102
176
  - Datasets 1.18.3
177
  - Tokenizers 0.11.0
178
+
179
+ ## Evaluation results
180
+
181
+ Evaluation was done with the [Common Voice 7.0 Finnish test split](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0).
182
+
183
+ To evaluate this model, run the `eval.py` script in this repository:
184
+
185
+ ```bash
186
+ python3 eval.py --model_id aapot/wav2vec2-xlsr-1b-finnish-v2 --dataset mozilla-foundation/common_voice_7_0 --config fi --split test
187
+ ```
188
+
189
+ 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:
190
+
191
+ | | WER (with LM) | WER (without LM) | CER (with LM) | CER (without LM) |
192
+ |-----------------------------------------|---------------|------------------|---------------|------------------|
193
+ |aapot/wav2vec2-xlsr-1b-finnish-lm-v2 |**4.09** |**9.73** |**0.88** |**1.65** |
194
+ |aapot/wav2vec2-xlsr-1b-finnish-lm |5.65 |13.11 |1.20 |2.23 |
195
+ |aapot/wav2vec2-xlsr-300m-finnish-lm |8.16 |17.92 |1.97 |3.36 |
196
+
197
+ ## Team Members
198
+
199
+ - Aapo Tanskanen, [Hugging Face profile](https://huggingface.co/aapot), [LinkedIn profile](https://www.linkedin.com/in/aapotanskanen/)
200
+ - Rasmus Toivanen, [Hugging Face profile](https://huggingface.co/RASMUS), [LinkedIn profile](https://www.linkedin.com/in/rasmustoivanen/)
201
+
202
+ Feel free to contact us for more details 🤗