Commit
•
f4d2d84
1
Parent(s):
81b36f6
first commit
Browse files- README.md +157 -0
- config.json +69 -0
- preprocessor_config.json +8 -0
- pytorch_model.bin +3 -0
- special_tokens_map.json +1 -0
- vocab.json +1 -0
README.md
ADDED
@@ -0,0 +1,157 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language: fr
|
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 French by Jonatas Grosman
|
16 |
+
results:
|
17 |
+
- task:
|
18 |
+
name: Speech Recognition
|
19 |
+
type: automatic-speech-recognition
|
20 |
+
dataset:
|
21 |
+
name: Common Voice fr
|
22 |
+
type: common_voice
|
23 |
+
args: fr
|
24 |
+
metrics:
|
25 |
+
- name: Test WER
|
26 |
+
type: wer
|
27 |
+
value: 18.01
|
28 |
+
- name: Test CER
|
29 |
+
type: cer
|
30 |
+
value: 6.29
|
31 |
+
|
32 |
+
---
|
33 |
+
|
34 |
+
# Wav2Vec2-Large-XLSR-53-French
|
35 |
+
|
36 |
+
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on French using the [Common Voice](https://huggingface.co/datasets/common_voice) and [CSS10](https://github.com/Kyubyong/css10).
|
37 |
+
When using this model, make sure that your speech input is sampled at 16kHz.
|
38 |
+
|
39 |
+
The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint
|
40 |
+
|
41 |
+
## Usage
|
42 |
+
|
43 |
+
The model can be used directly (without a language model) as follows:
|
44 |
+
|
45 |
+
```python
|
46 |
+
import torch
|
47 |
+
import librosa
|
48 |
+
from datasets import load_dataset
|
49 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
50 |
+
|
51 |
+
LANG_ID = "fr"
|
52 |
+
MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-french"
|
53 |
+
SAMPLES = 5
|
54 |
+
|
55 |
+
test_dataset = load_dataset("common_voice", LANG_ID, split=f"test[:{SAMPLES}]")
|
56 |
+
|
57 |
+
processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
|
58 |
+
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
|
59 |
+
|
60 |
+
# Preprocessing the datasets.
|
61 |
+
# We need to read the audio files as arrays
|
62 |
+
def speech_file_to_array_fn(batch):
|
63 |
+
speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
|
64 |
+
batch["speech"] = speech_array
|
65 |
+
batch["sentence"] = batch["sentence"].upper()
|
66 |
+
return batch
|
67 |
+
|
68 |
+
test_dataset = test_dataset.map(speech_file_to_array_fn)
|
69 |
+
inputs = processor(test_dataset["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
70 |
+
|
71 |
+
with torch.no_grad():
|
72 |
+
logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
|
73 |
+
|
74 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
75 |
+
predicted_sentences = processor.batch_decode(predicted_ids)
|
76 |
+
|
77 |
+
for i, predicted_sentence in enumerate(predicted_sentences):
|
78 |
+
print("-" * 100)
|
79 |
+
print("Reference:", test_dataset[i]["sentence"])
|
80 |
+
print("Prediction:", predicted_sentence)
|
81 |
+
```
|
82 |
+
|
83 |
+
| Reference | Prediction |
|
84 |
+
| ------------- | ------------- |
|
85 |
+
| "CE DERNIER A ÉVOLUÉ TOUT AU LONG DE L'HISTOIRE ROMAINE." | CE DERNIER EST VOLÉ TOUT AU LONG DE L'HISTOIRE ROMAINE |
|
86 |
+
| CE SITE CONTIENT QUATRE TOMBEAUX DE LA DYNASTIE ACHÉMÉNIDE ET SEPT DES SASSANIDES. | CE SITE CONTIENT QUATRE TOMBEAUX DE LA DYNASTIE ASHÉMÉNIDE ET SEPT DES SASANNIDES |
|
87 |
+
| "J'AI DIT QUE LES ACTEURS DE BOIS AVAIENT, SELON MOI, BEAUCOUP D'AVANTAGES SUR LES AUTRES." | JAI DIT QUE LES ACTEURS DE BOIS AVAIENT SELON MOI BEAUCOUP DAVANTAGES SUR LES AUTRES |
|
88 |
+
| LES PAYS-BAS ONT REMPORTÉ TOUTES LES ÉDITIONS. | LE PAYS-BAS AN REMPORTAIT TOUTES LES ÉDITIONS |
|
89 |
+
| IL Y A MAINTENANT UNE GARE ROUTIÈRE. | IL A MA ANDIN GARD DETIRON |
|
90 |
+
|
91 |
+
## Evaluation
|
92 |
+
|
93 |
+
The model can be evaluated as follows on the French test data of Common Voice.
|
94 |
+
|
95 |
+
```python
|
96 |
+
import torch
|
97 |
+
import re
|
98 |
+
import librosa
|
99 |
+
from datasets import load_dataset, load_metric
|
100 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
101 |
+
|
102 |
+
LANG_ID = "fr"
|
103 |
+
MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-french"
|
104 |
+
DEVICE = "cuda"
|
105 |
+
MAX_SAMPLES = 8000
|
106 |
+
|
107 |
+
CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
|
108 |
+
"؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
|
109 |
+
"=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。"]
|
110 |
+
|
111 |
+
test_dataset = load_dataset("common_voice", LANG_ID, split="test")
|
112 |
+
if len(test_dataset) > MAX_SAMPLES:
|
113 |
+
test_dataset = test_dataset.select(range(MAX_SAMPLES))
|
114 |
+
|
115 |
+
wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py
|
116 |
+
cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py
|
117 |
+
|
118 |
+
chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
|
119 |
+
|
120 |
+
processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
|
121 |
+
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
|
122 |
+
model.to(DEVICE)
|
123 |
+
|
124 |
+
# Preprocessing the datasets.
|
125 |
+
# We need to read the audio files as arrays
|
126 |
+
def speech_file_to_array_fn(batch):
|
127 |
+
with warnings.catch_warnings():
|
128 |
+
warnings.simplefilter("ignore")
|
129 |
+
speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
|
130 |
+
batch["speech"] = speech_array
|
131 |
+
batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).upper()
|
132 |
+
return batch
|
133 |
+
|
134 |
+
test_dataset = test_dataset.map(speech_file_to_array_fn)
|
135 |
+
|
136 |
+
# Preprocessing the datasets.
|
137 |
+
# We need to read the audio files as arrays
|
138 |
+
def evaluate(batch):
|
139 |
+
inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
140 |
+
|
141 |
+
with torch.no_grad():
|
142 |
+
logits = model(inputs.input_values.to(DEVICE), attention_mask=inputs.attention_mask.to(DEVICE)).logits
|
143 |
+
|
144 |
+
pred_ids = torch.argmax(logits, dim=-1)
|
145 |
+
batch["pred_strings"] = processor.batch_decode(pred_ids)
|
146 |
+
return batch
|
147 |
+
|
148 |
+
result = test_dataset.map(evaluate, batched=True, batch_size=8)
|
149 |
+
|
150 |
+
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"], chunk_size=1000)))
|
151 |
+
print("CER: {:2f}".format(100 * cer.compute(predictions=result["pred_strings"], references=result["sentence"], chunk_size=1000)))
|
152 |
+
```
|
153 |
+
|
154 |
+
**Test Result**:
|
155 |
+
|
156 |
+
- WER: 18.01%
|
157 |
+
- CER: 6.29%
|
config.json
ADDED
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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": 0.1,
|
47 |
+
"gradient_checkpointing": true,
|
48 |
+
"hidden_act": "gelu",
|
49 |
+
"hidden_dropout": 0.05,
|
50 |
+
"hidden_dropout_prob": 0.1,
|
51 |
+
"hidden_size": 1024,
|
52 |
+
"initializer_range": 0.02,
|
53 |
+
"intermediate_size": 4096,
|
54 |
+
"layer_norm_eps": 1e-05,
|
55 |
+
"layerdrop": 0.05,
|
56 |
+
"mask_feature_length": 10,
|
57 |
+
"mask_feature_prob": 0.0,
|
58 |
+
"mask_time_length": 10,
|
59 |
+
"mask_time_prob": 0.05,
|
60 |
+
"model_type": "wav2vec2",
|
61 |
+
"num_attention_heads": 16,
|
62 |
+
"num_conv_pos_embedding_groups": 16,
|
63 |
+
"num_conv_pos_embeddings": 128,
|
64 |
+
"num_feat_extract_layers": 7,
|
65 |
+
"num_hidden_layers": 24,
|
66 |
+
"pad_token_id": 0,
|
67 |
+
"transformers_version": "4.5.0.dev0",
|
68 |
+
"vocab_size": 49
|
69 |
+
}
|
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:d3d6ee65a78f8aa34f5355f21586fc77a608aa97c60c231948f7fa19354c1ae5
|
3 |
+
size 1262134743
|
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, "E": 5, "S": 6, "A": 7, "I": 8, "T": 9, "N": 10, "R": 11, "U": 12, "L": 13, "O": 14, "D": 15, "M": 16, "C": 17, "P": 18, "É": 19, "V": 20, "'": 21, "Q": 22, "F": 23, "G": 24, "B": 25, "H": 26, "J": 27, "À": 28, "X": 29, "È": 30, "Y": 31, "-": 32, "Ê": 33, "Z": 34, "Â": 35, "Ç": 36, "Î": 37, "Ô": 38, "Û": 39, "Ù": 40, "K": 41, "Œ": 42, "Ï": 43, "W": 44, "Ë": 45, "Ü": 46, "Æ": 47, "Ÿ": 48}
|