aapot commited on
Commit
de8313e
•
1 Parent(s): 3f73b2c
.gitignore ADDED
@@ -0,0 +1 @@
 
1
+ checkpoint-*/
README.md CHANGED
@@ -1,3 +1,178 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
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: 5.65
30
+ - name: Test CER
31
+ type: cer
32
+ value: 1.2
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 259.57 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](https://huggingface.co/aapot/wav2vec2-xlsr-1b-finnish-lm) model so that model has just been copied/moved to this `Finnish-NLP` Hugging Face organization.
43
+
44
+ **Note**: there is a better V2 version of this model which has been fine-tuned longer with 16 hours of more data: [Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2](https://huggingface.co/Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2)
45
+
46
+ ## Model description
47
+
48
+ 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.
49
+
50
+ 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).
51
+
52
+ This model is fine-tuned version of the pretrained model (1 billion parameter variant) for Finnish ASR.
53
+
54
+ ## Intended uses & limitations
55
+
56
+ You can use this model for Finnish ASR (speech-to-text) task.
57
+
58
+ ### How to use
59
+
60
+ Check the [run-finnish-asr-models.ipynb](https://huggingface.co/Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm/blob/main/run-finnish-asr-models.ipynb) notebook in this repository for an detailed example on how to use this model.
61
+
62
+ ### Limitations and bias
63
+
64
+ 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).
65
+
66
+ 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.
67
+
68
+ The Finnish KenLM language model used in the decoding phase has been trained with text data from the audio transcriptions. Thus, the decoder's language model may not generalize to very different language, for example to spoken daily language with dialects. It may be beneficial to train your own KenLM language model for your domain language and use that in the decoding.
69
+
70
+ ## Training data
71
+
72
+ This model was fine-tuned with 259.57 hours of Finnish transcribed speech data from following datasets:
73
+
74
+ | Dataset | Hours | % of total hours |
75
+ |:----------------------------------------------------------------------------------------------------------------------------------|:--------:|:----------------:|
76
+ | [Common Voice 7.0 Finnish train + evaluation + other splits](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0) | 9.70 h | 3.74 % |
77
+ | [Finnish parliament session 2](https://b2share.eudat.eu/records/4df422d631544ce682d6af1d4714b2d4) | 0.24 h | 0.09 % |
78
+ | [VoxPopuli Finnish](https://github.com/facebookresearch/voxpopuli) | 5.94 h | 2.29 % |
79
+ | [CSS10 Finnish](https://github.com/kyubyong/css10) | 10.32 h | 3.98 % |
80
+ | [Aalto Finnish Parliament ASR Corpus](http://urn.fi/urn:nbn:fi:lb-2021051903) | 228.00 h | 87.84 % |
81
+ | [Finnish Broadcast Corpus](http://urn.fi/urn:nbn:fi:lb-2016042502) | 5.37 h | 2.07 % |
82
+
83
+ Datasets were filtered to include maximum length of 20 seconds long audio samples.
84
+
85
+ ## Training procedure
86
+
87
+ 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.
88
+
89
+ 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.
90
+
91
+ 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.
92
+
93
+ ### Training hyperparameters
94
+
95
+ The following hyperparameters were used during training:
96
+ - learning_rate: 5e-05
97
+ - train_batch_size: 32
98
+ - eval_batch_size: 8
99
+ - seed: 42
100
+ - optimizer: [8-bit Adam](https://github.com/facebookresearch/bitsandbytes) with betas=(0.9,0.999) and epsilon=1e-08
101
+ - lr_scheduler_type: linear
102
+ - lr_scheduler_warmup_steps: 500
103
+ - num_epochs: 5
104
+ - mixed_precision_training: Native AMP
105
+
106
+ The pretrained `facebook/wav2vec2-xls-r-1b` model was initialized with following hyperparameters:
107
+ - attention_dropout: 0.094
108
+ - hidden_dropout: 0.047
109
+ - feat_proj_dropout: 0.04
110
+ - mask_time_prob: 0.082
111
+ - layerdrop: 0.041
112
+ - activation_dropout: 0.055
113
+ - ctc_loss_reduction: "mean"
114
+
115
+ ### Training results
116
+
117
+ | Training Loss | Epoch | Step | Validation Loss | Wer |
118
+ |:-------------:|:-----:|:-----:|:---------------:|:------:|
119
+ | 0.968 | 0.18 | 500 | 0.4870 | 0.4720 |
120
+ | 0.6557 | 0.36 | 1000 | 0.2450 | 0.2931 |
121
+ | 0.647 | 0.54 | 1500 | 0.1818 | 0.2255 |
122
+ | 0.5297 | 0.72 | 2000 | 0.1698 | 0.2354 |
123
+ | 0.5802 | 0.9 | 2500 | 0.1581 | 0.2355 |
124
+ | 0.6351 | 1.07 | 3000 | 0.1689 | 0.2336 |
125
+ | 0.4626 | 1.25 | 3500 | 0.1719 | 0.3099 |
126
+ | 0.4526 | 1.43 | 4000 | 0.1434 | 0.2069 |
127
+ | 0.4692 | 1.61 | 4500 | 0.1645 | 0.2192 |
128
+ | 0.4584 | 1.79 | 5000 | 0.1483 | 0.1987 |
129
+ | 0.4234 | 1.97 | 5500 | 0.1499 | 0.2178 |
130
+ | 0.4243 | 2.15 | 6000 | 0.1345 | 0.2070 |
131
+ | 0.4108 | 2.33 | 6500 | 0.1383 | 0.1850 |
132
+ | 0.4048 | 2.51 | 7000 | 0.1338 | 0.1811 |
133
+ | 0.4085 | 2.69 | 7500 | 0.1290 | 0.1780 |
134
+ | 0.4026 | 2.87 | 8000 | 0.1239 | 0.1650 |
135
+ | 0.4033 | 3.04 | 8500 | 0.1346 | 0.1657 |
136
+ | 0.3986 | 3.22 | 9000 | 0.1310 | 0.1850 |
137
+ | 0.3867 | 3.4 | 9500 | 0.1273 | 0.1741 |
138
+ | 0.3658 | 3.58 | 10000 | 0.1219 | 0.1672 |
139
+ | 0.382 | 3.76 | 10500 | 0.1306 | 0.1698 |
140
+ | 0.3847 | 3.94 | 11000 | 0.1230 | 0.1577 |
141
+ | 0.3691 | 4.12 | 11500 | 0.1310 | 0.1615 |
142
+ | 0.3593 | 4.3 | 12000 | 0.1296 | 0.1622 |
143
+ | 0.3619 | 4.48 | 12500 | 0.1285 | 0.1601 |
144
+ | 0.3361 | 4.66 | 13000 | 0.1261 | 0.1569 |
145
+ | 0.3603 | 4.84 | 13500 | 0.1235 | 0.1533 |
146
+
147
+
148
+ ### Framework versions
149
+
150
+ - Transformers 4.17.0.dev0
151
+ - Pytorch 1.10.2+cu102
152
+ - Datasets 1.18.3
153
+ - Tokenizers 0.11.0
154
+
155
+ ## Evaluation results
156
+
157
+ Evaluation was done with the [Common Voice 7.0 Finnish test split](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0).
158
+
159
+ To evaluate this model, run the `eval.py` script in this repository:
160
+
161
+ ```bash
162
+ python3 eval.py --model_id Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm --dataset mozilla-foundation/common_voice_7_0 --config fi --split test
163
+ ```
164
+
165
+ This model (the second row of the table) achieves the following WER (Word Error Rate) and CER (Character Error Rate) results compared to our other models:
166
+
167
+ | | WER (with LM) | WER (without LM) | CER (with LM) | CER (without LM) |
168
+ |-----------------------------------------------|---------------|------------------|---------------|------------------|
169
+ |Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm-v2 |**4.09** |**9.73** |**0.88** |**1.65** |
170
+ |Finnish-NLP/wav2vec2-xlsr-1b-finnish-lm |5.65 |13.11 |1.20 |2.23 |
171
+ |Finnish-NLP/wav2vec2-xlsr-300m-finnish-lm |8.16 |17.92 |1.97 |3.36 |
172
+
173
+ ## Team Members
174
+
175
+ - Aapo Tanskanen, [Hugging Face profile](https://huggingface.co/aapot), [LinkedIn profile](https://www.linkedin.com/in/aapotanskanen/)
176
+ - Rasmus Toivanen, [Hugging Face profile](https://huggingface.co/RASMUS), [LinkedIn profile](https://www.linkedin.com/in/rasmustoivanen/)
177
+
178
+ Feel free to contact us for more details 🤗
added_tokens.json ADDED
@@ -0,0 +1 @@
 
1
+ {"<s>": 33, "</s>": 34}
alphabet.json ADDED
@@ -0,0 +1 @@
 
1
+ {"labels": [" ", "'", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "\u00e4", "\u00e5", "\u00f6", "\u2047", "", "<s>", "</s>"], "is_bpe": false}
config.json ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-1b",
3
+ "activation_dropout": 0.055,
4
+ "adapter_kernel_size": 3,
5
+ "adapter_stride": 2,
6
+ "add_adapter": false,
7
+ "apply_spec_augment": true,
8
+ "architectures": [
9
+ "Wav2Vec2ForCTC"
10
+ ],
11
+ "attention_dropout": 0.094,
12
+ "bos_token_id": 1,
13
+ "classifier_proj_size": 256,
14
+ "codevector_dim": 1024,
15
+ "contrastive_logits_temperature": 0.1,
16
+ "conv_bias": true,
17
+ "conv_dim": [
18
+ 512,
19
+ 512,
20
+ 512,
21
+ 512,
22
+ 512,
23
+ 512,
24
+ 512
25
+ ],
26
+ "conv_kernel": [
27
+ 10,
28
+ 3,
29
+ 3,
30
+ 3,
31
+ 3,
32
+ 2,
33
+ 2
34
+ ],
35
+ "conv_stride": [
36
+ 5,
37
+ 2,
38
+ 2,
39
+ 2,
40
+ 2,
41
+ 2,
42
+ 2
43
+ ],
44
+ "ctc_loss_reduction": "mean",
45
+ "ctc_zero_infinity": false,
46
+ "diversity_loss_weight": 0.1,
47
+ "do_stable_layer_norm": true,
48
+ "eos_token_id": 2,
49
+ "feat_extract_activation": "gelu",
50
+ "feat_extract_dropout": 0.0,
51
+ "feat_extract_norm": "layer",
52
+ "feat_proj_dropout": 0.04,
53
+ "feat_quantizer_dropout": 0.0,
54
+ "final_dropout": 0.0,
55
+ "gradient_checkpointing": false,
56
+ "hidden_act": "gelu",
57
+ "hidden_dropout": 0.047,
58
+ "hidden_size": 1280,
59
+ "initializer_range": 0.02,
60
+ "intermediate_size": 5120,
61
+ "layer_norm_eps": 1e-05,
62
+ "layerdrop": 0.041,
63
+ "mask_feature_length": 10,
64
+ "mask_feature_min_masks": 0,
65
+ "mask_feature_prob": 0.0,
66
+ "mask_time_length": 10,
67
+ "mask_time_min_masks": 2,
68
+ "mask_time_prob": 0.082,
69
+ "model_type": "wav2vec2",
70
+ "num_adapter_layers": 3,
71
+ "num_attention_heads": 16,
72
+ "num_codevector_groups": 2,
73
+ "num_codevectors_per_group": 320,
74
+ "num_conv_pos_embedding_groups": 16,
75
+ "num_conv_pos_embeddings": 128,
76
+ "num_feat_extract_layers": 7,
77
+ "num_hidden_layers": 48,
78
+ "num_negatives": 100,
79
+ "output_hidden_size": 1280,
80
+ "pad_token_id": 32,
81
+ "proj_codevector_dim": 1024,
82
+ "tdnn_dilation": [
83
+ 1,
84
+ 2,
85
+ 3,
86
+ 1,
87
+ 1
88
+ ],
89
+ "tdnn_dim": [
90
+ 512,
91
+ 512,
92
+ 512,
93
+ 512,
94
+ 1500
95
+ ],
96
+ "tdnn_kernel": [
97
+ 5,
98
+ 3,
99
+ 3,
100
+ 1,
101
+ 1
102
+ ],
103
+ "torch_dtype": "float32",
104
+ "transformers_version": "4.17.0.dev0",
105
+ "use_weighted_layer_sum": false,
106
+ "vocab_size": 35,
107
+ "xvector_output_dim": 512
108
+ }
eval.py ADDED
@@ -0,0 +1,144 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ import argparse
3
+ import re
4
+ from typing import Dict
5
+
6
+ import torch
7
+ from datasets import Audio, Dataset, load_dataset, load_metric
8
+
9
+ from transformers import AutoFeatureExtractor, pipeline
10
+
11
+
12
+ def log_results(result: Dataset, args: Dict[str, str]):
13
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
14
+
15
+ log_outputs = args.log_outputs
16
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
17
+
18
+ # load metric
19
+ wer = load_metric("wer")
20
+ cer = load_metric("cer")
21
+
22
+ # compute metrics
23
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
24
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
25
+
26
+ # print & log results
27
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
28
+ print(result_str)
29
+
30
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
31
+ f.write(result_str)
32
+
33
+ # log all results in text file. Possibly interesting for analysis
34
+ if log_outputs is not None:
35
+ pred_file = f"log_{dataset_id}_predictions.txt"
36
+ target_file = f"log_{dataset_id}_targets.txt"
37
+
38
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
39
+
40
+ # mapping function to write output
41
+ def write_to_file(batch, i):
42
+ p.write(f"{i}" + "\n")
43
+ p.write(batch["prediction"] + "\n")
44
+ t.write(f"{i}" + "\n")
45
+ t.write(batch["target"] + "\n")
46
+
47
+ result.map(write_to_file, with_indices=True)
48
+
49
+
50
+ def normalize_text(text: str) -> str:
51
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
52
+
53
+ CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
54
+ "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
55
+ "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
56
+ "、", "﹂", "﹁", "‧", "~", "﹏", ",", "{", "}", "(", ")", "[", "]", "【", "】", "‥", "〽",
57
+ "『", "』", "〝", "〟", "⟨", "⟩", "〜", ":", "!", "?", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"] # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
58
+
59
+ chars_to_remove_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
60
+
61
+ text = re.sub(chars_to_remove_regex, "", text.lower())
62
+ text = re.sub("[-]", " ", text)
63
+
64
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
65
+ # note that order is important here!
66
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
67
+
68
+ for t in token_sequences_to_ignore:
69
+ text = " ".join(text.split(t))
70
+
71
+ return text
72
+
73
+
74
+ def main(args):
75
+ # load dataset
76
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
77
+
78
+ # for testing: only process the first two examples as a test
79
+ # dataset = dataset.select(range(10))
80
+
81
+ # load processor
82
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
83
+ sampling_rate = feature_extractor.sampling_rate
84
+
85
+ # resample audio
86
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
87
+
88
+ # load eval pipeline
89
+ if args.device is None:
90
+ args.device = 0 if torch.cuda.is_available() else -1
91
+ asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
92
+
93
+ # map function to decode audio
94
+ def map_to_pred(batch):
95
+ prediction = asr(
96
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
97
+ )
98
+
99
+ batch["prediction"] = prediction["text"]
100
+ batch["target"] = normalize_text(batch["sentence"])
101
+ return batch
102
+
103
+ # run inference on all examples
104
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
105
+
106
+ # compute and log_results
107
+ # do not change function below
108
+ log_results(result, args)
109
+
110
+
111
+ if __name__ == "__main__":
112
+ parser = argparse.ArgumentParser()
113
+
114
+ parser.add_argument(
115
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
116
+ )
117
+ parser.add_argument(
118
+ "--dataset",
119
+ type=str,
120
+ required=True,
121
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
122
+ )
123
+ parser.add_argument(
124
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
125
+ )
126
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
127
+ parser.add_argument(
128
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
129
+ )
130
+ parser.add_argument(
131
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
132
+ )
133
+ parser.add_argument(
134
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
135
+ )
136
+ parser.add_argument(
137
+ "--device",
138
+ type=int,
139
+ default=None,
140
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
141
+ )
142
+ args = parser.parse_args()
143
+
144
+ main(args)
language_model/attrs.json ADDED
@@ -0,0 +1 @@
 
1
+ {"alpha": 0.5, "beta": 1.5, "unk_score_offset": -10.0, "score_boundary": true}
language_model/kenlm_finnish.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:45e82be92cfe0ce2e74b0d31ea0e7949b7b185a95730c39ab012f599cf4d8d75
3
+ size 20686116
language_model/unigrams.txt ADDED
The diff for this file is too large to render. See raw diff
log_mozilla-foundation_common_voice_7_0_fi_test_predictions.txt ADDED
@@ -0,0 +1,3198 @@