arampacha commited on
Commit
a69b27b
1 Parent(s): 381c9f4
README.md ADDED
@@ -0,0 +1,133 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+
3
+ language: uk
4
+ dataset: common_voice
5
+ metrics: wer
6
+ tags:
7
+ - audio
8
+ - automatic-speech-recognition
9
+ - speech
10
+ - xlsr-fine-tuning-week
11
+ license: apache-2.0
12
+ model-index:
13
+ - name: Ukrainian XLSR Wav2Vec2 Large 53
14
+ results:
15
+ - task:
16
+ name: Speech Recognition
17
+ type: automatic-speech-recognition
18
+ dataset:
19
+ name: Common Voice uk
20
+ type: common_voice
21
+ args: cs
22
+ metrics:
23
+ - name: Test WER
24
+ type: wer
25
+ value: 37.72
26
+ ---
27
+
28
+ # Wav2Vec2-Large-XLSR-53-Chech
29
+
30
+ Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Czech using the [Common Voice](https://huggingface.co/datasets/common_voice) dataset.
31
+
32
+ When using this model, make sure that your speech input is sampled at 16kHz.
33
+
34
+ ## Usage
35
+
36
+ The model can be used directly (without a language model) as follows:
37
+
38
+ ```python
39
+
40
+ import torch
41
+ import torchaudio
42
+ from datasets import load_dataset
43
+ from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
44
+
45
+ test_dataset = load_dataset("common_voice", "cs", split="test[:2%]")
46
+
47
+ processor = Wav2Vec2Processor.from_pretrained("arampacha/wav2vec2-large-xlsr-ukrainian")
48
+ model = Wav2Vec2ForCTC.from_pretrained("arampacha/wav2vec2-large-xlsr-ukrainian")
49
+ resampler = torchaudio.transforms.Resample(48_000, 16_000)
50
+
51
+ # Preprocessing the datasets.
52
+ # We need to read the aduio files as arrays
53
+
54
+ def speech_file_to_array_fn(batch):
55
+ speech_array, sampling_rate = torchaudio.load(batch["path"])
56
+ batch["speech"] = resampler(speech_array).squeeze().numpy()
57
+ return batch
58
+
59
+ test_dataset = test_dataset.map(speech_file_to_array_fn)
60
+
61
+ inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
62
+
63
+ with torch.no_grad():
64
+ logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
65
+
66
+ predicted_ids = torch.argmax(logits, dim=-1)
67
+
68
+ print("Prediction:", processor.batch_decode(predicted_ids))
69
+ print("Reference:", test_dataset["sentence"][:2])
70
+
71
+ ```
72
+
73
+ ## Evaluation
74
+
75
+ The model can be evaluated as follows on the Ukrainian test data of Common Voice.
76
+
77
+ ```python
78
+
79
+ import torch
80
+ import torchaudio
81
+ from datasets import load_dataset, load_metric
82
+ from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
83
+ import re
84
+
85
+ test_dataset = load_dataset("common_voice", "cs", split="test")
86
+
87
+ wer = load_metric("wer")
88
+ processor = Wav2Vec2Processor.from_pretrained("arampacha/wav2vec2-large-xlsr-czech")
89
+ model = Wav2Vec2ForCTC.from_pretrained("arampacha/wav2vec2-large-xlsr-czech")
90
+ model.to("cuda")
91
+
92
+ chars_to_ignore = [",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�", '«', '»', '—', '…', '(', ')', '*', '”', '“']
93
+ chars_to_ignore_regex = f'[{"".join(chars_to_ignore)}]'
94
+ resampler = torchaudio.transforms.Resample(48_000, 16_000)
95
+
96
+ # Preprocessing the datasets.
97
+ # We need to read the aduio files as arrays
98
+ def speech_file_to_array_fn(batch):
99
+ batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().strip()
100
+ batch["sentence"] = re.sub(re.compile('i'), 'і', batch['sentence'])
101
+ batch['sentence'] = re.sub(' ', ' ', batch['sentence'])
102
+ speech_array, sampling_rate = torchaudio.load(batch["path"])
103
+ batch["speech"] = resampler(speech_array).squeeze().numpy()
104
+ return batch
105
+
106
+ test_dataset = test_dataset.map(speech_file_to_array_fn)
107
+
108
+ # Preprocessing the datasets.
109
+
110
+ # We need to read the aduio files as arrays
111
+
112
+ def evaluate(batch):
113
+ inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
114
+ with torch.no_grad():
115
+ logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
116
+
117
+ pred_ids = torch.argmax(logits, dim=-1)
118
+ batch["pred_strings"] = processor.batch_decode(pred_ids)
119
+ return batch
120
+
121
+ result = test_dataset.map(evaluate, batched=True, batch_size=8)
122
+
123
+ print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
124
+
125
+ ```
126
+
127
+ **Test Result**: 37.72
128
+
129
+ ## Training
130
+
131
+ The Common Voice `train`, `validation`.
132
+
133
+ The script used for training will be available [here](https://github.com/arampacha/hf-sprint-xlsr) soon.
config.json ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-large-xlsr-53",
3
+ "activation_dropout": 0.1,
4
+ "apply_spec_augment": true,
5
+ "architectures": [
6
+ "Wav2Vec2ForCTC"
7
+ ],
8
+ "attention_dropout": 0.1,
9
+ "bos_token_id": 1,
10
+ "conv_bias": true,
11
+ "conv_dim": [
12
+ 512,
13
+ 512,
14
+ 512,
15
+ 512,
16
+ 512,
17
+ 512,
18
+ 512
19
+ ],
20
+ "conv_kernel": [
21
+ 10,
22
+ 3,
23
+ 3,
24
+ 3,
25
+ 3,
26
+ 2,
27
+ 2
28
+ ],
29
+ "conv_stride": [
30
+ 5,
31
+ 2,
32
+ 2,
33
+ 2,
34
+ 2,
35
+ 2,
36
+ 2
37
+ ],
38
+ "ctc_loss_reduction": "mean",
39
+ "ctc_zero_infinity": true,
40
+ "do_stable_layer_norm": true,
41
+ "eos_token_id": 2,
42
+ "feat_extract_activation": "gelu",
43
+ "feat_extract_dropout": 0.0,
44
+ "feat_extract_norm": "layer",
45
+ "feat_proj_dropout": 0.008,
46
+ "final_dropout": 0.0,
47
+ "gradient_checkpointing": true,
48
+ "hidden_act": "gelu",
49
+ "hidden_dropout": 0.1,
50
+ "hidden_size": 1024,
51
+ "initializer_range": 0.02,
52
+ "intermediate_size": 4096,
53
+ "layer_norm_eps": 1e-05,
54
+ "layerdrop": 0.1,
55
+ "mask_channel_length": 10,
56
+ "mask_channel_min_space": 1,
57
+ "mask_channel_other": 0.0,
58
+ "mask_channel_prob": 0.0,
59
+ "mask_channel_selection": "static",
60
+ "mask_feature_length": 10,
61
+ "mask_feature_prob": 0.0,
62
+ "mask_time_length": 10,
63
+ "mask_time_min_space": 1,
64
+ "mask_time_other": 0.0,
65
+ "mask_time_prob": 0.05,
66
+ "mask_time_selection": "static",
67
+ "model_type": "wav2vec2",
68
+ "num_attention_heads": 16,
69
+ "num_conv_pos_embedding_groups": 16,
70
+ "num_conv_pos_embeddings": 128,
71
+ "num_feat_extract_layers": 7,
72
+ "num_hidden_layers": 24,
73
+ "pad_token_id": 0,
74
+ "transformers_version": "4.5.0.dev0",
75
+ "vocab_size": 45
76
+ }
preprocessor_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_normalize": true,
3
+ "feature_size": 1,
4
+ "padding_side": "right",
5
+ "padding_value": 0.0,
6
+ "return_attention_mask": true,
7
+ "sampling_rate": 16000
8
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a8d0ebba6245b55b78fccbc35951b8027c2265d05395bde0f4e7f58113528616
3
+ size 1262118359
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]"}
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|"}
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"[PAD]": 0, "[UNK]": 1, "c": 3, "l": 4, "m": 5, "n": 6, "p": 7, "u": 8, "x": 9, "y": 10, "а": 11, "б": 12, "в": 13, "г": 14, "д": 15, "е": 16, "ж": 17, "з": 18, "и": 19, "й": 20, "к": 21, "л": 22, "м": 23, "н": 24, "о": 25, "п": 26, "р": 27, "с": 28, "т": 29, "у": 30, "ф": 31, "х": 32, "ц": 33, "ч": 34, "ш": 35, "щ": 36, "ь": 37, "ю": 38, "я": 39, "є": 40, "і": 41, "ї": 42, "ґ": 43, "’": 44, "|": 2}