jonatasgrosman commited on
Commit
533a106
1 Parent(s): 26ef1d6

first commit

Browse files
README.md ADDED
@@ -0,0 +1,162 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: it
3
+ datasets:
4
+ - common_voice
5
+ metrics:
6
+ - wer
7
+ - cer
8
+ tags:
9
+ - audio
10
+ - automatic-speech-recognition
11
+ - speech
12
+ - xlsr-fine-tuning-week
13
+ license: apache-2.0
14
+ model-index:
15
+ - name: XLSR Wav2Vec2 Italian by Jonatas Grosman
16
+ results:
17
+ - task:
18
+ name: Speech Recognition
19
+ type: automatic-speech-recognition
20
+ dataset:
21
+ name: Common Voice it
22
+ type: common_voice
23
+ args: it
24
+ metrics:
25
+ - name: Test WER
26
+ type: wer
27
+ value: 11.90
28
+ - name: Test CER
29
+ type: cer
30
+ value: 2.94
31
+ ---
32
+
33
+ # Wav2Vec2-Large-XLSR-53-Italian
34
+
35
+ Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Italian using the [Common Voice](https://huggingface.co/datasets/common_voice).
36
+ When using this model, make sure that your speech input is sampled at 16kHz.
37
+
38
+ The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint
39
+
40
+ ## Usage
41
+
42
+ The model can be used directly (without a language model) as follows:
43
+
44
+ ```python
45
+ import torch
46
+ import librosa
47
+ from datasets import load_dataset
48
+ from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
49
+
50
+ LANG_ID = "it"
51
+ MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-italian"
52
+ SAMPLES = 5
53
+
54
+ test_dataset = load_dataset("common_voice", LANG_ID, split=f"test[:{SAMPLES}]")
55
+
56
+ processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
57
+ model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
58
+
59
+ # Preprocessing the datasets.
60
+ # We need to read the audio files as arrays
61
+ def speech_file_to_array_fn(batch):
62
+ speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
63
+ batch["speech"] = speech_array
64
+ batch["sentence"] = batch["sentence"].upper()
65
+ return batch
66
+
67
+ test_dataset = test_dataset.map(speech_file_to_array_fn)
68
+ inputs = processor(test_dataset["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
69
+
70
+ with torch.no_grad():
71
+ logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
72
+
73
+ predicted_ids = torch.argmax(logits, dim=-1)
74
+ predicted_sentences = processor.batch_decode(predicted_ids)
75
+
76
+ for i, predicted_sentence in enumerate(predicted_sentences):
77
+ print("-" * 100)
78
+ print("Reference:", test_dataset[i]["sentence"])
79
+ print("Prediction:", predicted_sentence)
80
+ ```
81
+
82
+ | Reference | Prediction |
83
+ | ------------- | ------------- |
84
+ | POI LEI MORÌ. | POI LEI MORÌ |
85
+ | IL LIBRO HA SUSCITATO MOLTE POLEMICHE A CAUSA DEI SUOI CONTENUTI. | IL LIBRO HA SUSCITATO MOLTE POLEMICHE A CAUSA DEI SUOI CONTENUTI |
86
+ | "FIN DALL'INIZIO LA SEDE EPISCOPALE È STATA IMMEDIATAMENTE SOGGETTA ALLA SANTA SEDE." | FIN DALL'INIZIO LA SEDE EPISCOPALE È STATA IMMEDIATAMENTE SOGGETTA ALLA SANTA SEDE |
87
+ | IL VUOTO ASSOLUTO? | IL VUOTO ASSOLUTO |
88
+ | DOPO ALCUNI ANNI, EGLI DECISE DI TORNARE IN INDIA PER RACCOGLIERE ALTRI INSEGNAMENTI. | DOPO ALCUNI ANNI EGLI DECISE DI TORNARE IN INDIA PER RACCOGLIERE ALTRI INSEGNAMENTI |
89
+
90
+ ## Evaluation
91
+
92
+ The model can be evaluated as follows on the Italian test data of Common Voice.
93
+
94
+ ```python
95
+ import torch
96
+ import re
97
+ import librosa
98
+ from datasets import load_dataset, load_metric
99
+ from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
100
+
101
+ LANG_ID = "it"
102
+ MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-italian"
103
+ DEVICE = "cuda"
104
+
105
+ CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
106
+ "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
107
+ "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。"]
108
+
109
+ test_dataset = load_dataset("common_voice", LANG_ID, split="test")
110
+
111
+ wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py
112
+ cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py
113
+
114
+ chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
115
+
116
+ processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
117
+ model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
118
+ model.to(DEVICE)
119
+
120
+ # Preprocessing the datasets.
121
+ # We need to read the audio files as arrays
122
+ def speech_file_to_array_fn(batch):
123
+ with warnings.catch_warnings():
124
+ warnings.simplefilter("ignore")
125
+ speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
126
+ batch["speech"] = speech_array
127
+ batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).upper()
128
+ return batch
129
+
130
+ test_dataset = test_dataset.map(speech_file_to_array_fn)
131
+
132
+ # Preprocessing the datasets.
133
+ # We need to read the audio files as arrays
134
+ def evaluate(batch):
135
+ inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
136
+
137
+ with torch.no_grad():
138
+ logits = model(inputs.input_values.to(DEVICE), attention_mask=inputs.attention_mask.to(DEVICE)).logits
139
+
140
+ pred_ids = torch.argmax(logits, dim=-1)
141
+ batch["pred_strings"] = processor.batch_decode(pred_ids)
142
+ return batch
143
+
144
+ result = test_dataset.map(evaluate, batched=True, batch_size=8)
145
+
146
+ predictions = [x.upper() for x in result["pred_strings"]]
147
+ references = [x.upper() for x in result["sentence"]]
148
+
149
+ print(f"WER: {wer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")
150
+ print(f"CER: {cer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")
151
+ ```
152
+
153
+ **Test Result**:
154
+
155
+ My model may report better scores than others because of some specificity of my evaluation script, so I ran the same evaluation script on other models (on 2020-04-21) to make a fairer comparison.
156
+
157
+ | Model | WER | CER |
158
+ | ------------- | ------------- | ------------- |
159
+ | jonatasgrosman/wav2vec2-large-xlsr-53-italian | **11.90%** | **2.94%** |
160
+ | joorock12/wav2vec2-large-xlsr-italian | 12.60% | 3.18% |
161
+ | gchhablani/wav2vec2-large-xlsr-it | 12.99% | 3.11% |
162
+ | facebook/wav2vec2-large-xlsr-53-italian | 22.08% | 6.36% |
config.json ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-large-xlsr-53",
3
+ "activation_dropout": 0.05,
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": 0.05,
46
+ "final_dropout": 0.0,
47
+ "gradient_checkpointing": true,
48
+ "hidden_act": "gelu",
49
+ "hidden_dropout": 0.05,
50
+ "hidden_size": 1024,
51
+ "initializer_range": 0.02,
52
+ "intermediate_size": 4096,
53
+ "layer_norm_eps": 1e-05,
54
+ "layerdrop": 0.05,
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": 75
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 sha256:2d9909cce880814662ccf4b45b2b3dcaf90801c7343acf2884e100ca50555b83
3
+ size 1262241303
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>"}
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<pad>": 0, "<s>": 1, "</s>": 2, "<unk>": 3, "|": 4, "Ð": 5, "Ó": 6, "Ł": 7, "Ô": 8, "O": 9, "Å": 10, "S": 11, "L": 12, "Ě": 13, "Ř": 14, "Ū": 15, "Ş": 16, "Ê": 17, "J": 18, "Q": 19, "Ō": 20, "U": 21, "Ž": 22, "A": 23, "Ć": 24, "B": 25, "Ï": 26, "Č": 27, "Ì": 28, "D": 29, "Y": 30, "Ğ": 31, "$": 32, "C": 33, "Á": 34, "P": 35, "Î": 36, "Ń": 37, "N": 38, "X": 39, "À": 40, "Ë": 41, "'": 42, "Ñ": 43, "Ī": 44, "È": 45, "Š": 46, "F": 47, "Ø": 48, "E": 49, "Ö": 50, "Ź": 51, "K": 52, "G": 53, "Ä": 54, "Ã": 55, "Μ": 56, "T": 57, "Ę": 58, "V": 59, "I": 60, "É": 61, "Z": 62, "R": 63, "-": 64, "W": 65, "Æ": 66, "M": 67, "Í": 68, "Œ": 69, "Ù": 70, "H": 71, "Ḥ": 72, "Ò": 73, "Ú": 74}