1 ---
2 language: el
3 datasets:
4 - common_voice
5 - CSS10 Greek: Single Speaker Speech Dataset
6 metrics:
7 - wer
8 - cer
9 tags:
10 - audio
11 - automatic-speech-recognition
12 - speech
13 - xlsr-fine-tuning-week
14 license: apache-2.0
15 model-index:
16 - name: V XLSR Wav2Vec2 Large 53 - greek
17 results:
18 - task:
19 name: Speech Recognition
20 type: automatic-speech-recognition
21 dataset:
22 name: Common Voice el
23 type: common_voice
24 args: el
25 metrics:
26 - name: Test WER
27 type: wer
28 value: 18.996669
29 - name: Test CER
30 type: cer
31 value: 5.781874
32 ---
33
34 # Wav2Vec2-Large-XLSR-53-greek
35
36 Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on greek using the [Common Voice](https://huggingface.co/datasets/common_voice) and [CSS10 Greek: Single Speaker Speech Dataset](https://www.kaggle.com/bryanpark/greek-single-speaker-speech-dataset).
37 When using this model, make sure that your speech input is sampled at 16kHz.
38
39 ## Usage
40
41 The model can be used directly (without a language model) as follows:
42
43 ```python
44 import torch
45 import torchaudio
46 from datasets import load_dataset
47 from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
48
49 test_dataset = load_dataset("common_voice", "el", split="test[:2%]") #TODO: replace {lang_id} in your language code here. Make sure the code is one of the *ISO codes* of [this](https://huggingface.co/languages) site.
50
51 processor = Wav2Vec2Processor.from_pretrained("vasilis/wav2vec2-large-xlsr-53-greek") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
52 model = Wav2Vec2ForCTC.from_pretrained("vasilis/wav2vec2-large-xlsr-53-greek") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
53
54 resampler = torchaudio.transforms.Resample(48_000, 16_000)
55
56 # Preprocessing the datasets.
57 # We need to read the aduio files as arrays
58 def speech_file_to_array_fn(batch):
59 speech_array, sampling_rate = torchaudio.load(batch["path"])
60 batch["speech"] = resampler(speech_array).squeeze().numpy()
61 return batch
62
63 test_dataset = test_dataset.map(speech_file_to_array_fn)
64 inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
65
66 with torch.no_grad():
67 logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
68
69 predicted_ids = torch.argmax(logits, dim=-1)
70
71 print("Prediction:", processor.batch_decode(predicted_ids))
72 print("Reference:", test_dataset["sentence"][:2])
73 ```
74
75
76 ## Evaluation
77
78 The model can be evaluated as follows on the greek test data of Common Voice.
79
80
81 ```python
82 import torch
83 import torchaudio
84 from datasets import load_dataset, load_metric
85 from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
86 import re
87
88 test_dataset = load_dataset("common_voice", "el", split="test") #TODO: replace {lang_id} in your language code here. Make sure the code is one of the *ISO codes* of [this](https://huggingface.co/languages) site.
89 wer = load_metric("wer")
90
91 processor = Wav2Vec2Processor.from_pretrained("vasilis/wav2vec2-large-xlsr-53-greek") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
92 model = Wav2Vec2ForCTC.from_pretrained("vasilis/wav2vec2-large-xlsr-53-greek") #TODO: replace {model_id} with your model id. The model id consists of {your_username}/{your_modelname}, *e.g.* `elgeish/wav2vec2-large-xlsr-53-arabic`
93 model.to("cuda")
94
95 chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“]' # TODO: adapt this list to include all special characters you removed from the data
96
97 normalize_greek_letters = {"ς": "σ"}
98 # normalize_greek_letters = {"ά": "α", "έ": "ε", "ί": "ι", 'ϊ': "ι", "ύ": "υ", "ς": "σ", "ΐ": "ι", 'ϋ': "υ", "ή": "η", "ώ": "ω", 'ό': "ο"}
99 remove_chars_greek = {"a": "", "h": "", "n": "", "g": "", "o": "", "v": "", "e": "", "r": "", "t": "", "«": "", "»": "", "m": "", '́': '', "·": "", "’": "", '´': ""}
100 replacements = {**normalize_greek_letters, **remove_chars_greek}
101
102 resampler = {
103 48_000: torchaudio.transforms.Resample(48_000, 16_000),
104 44100: torchaudio.transforms.Resample(44100, 16_000),
105 32000: torchaudio.transforms.Resample(32000, 16_000)
106 }
107
108
109 # Preprocessing the datasets.
110 # We need to read the aduio files as arrays
111 def speech_file_to_array_fn(batch):
112 batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
113 for key, value in replacements.items():
114 batch["sentence"] = batch["sentence"].replace(key, value)
115 speech_array, sampling_rate = torchaudio.load(batch["path"])
116 batch["speech"] = resampler[sampling_rate](speech_array).squeeze().numpy()
117 return batch
118
119
120 test_dataset = test_dataset.map(speech_file_to_array_fn)
121
122 # Preprocessing the datasets.
123 # We need to read the aduio files as arrays
124 def evaluate(batch):
125 inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
126
127 with torch.no_grad():
128 logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
129
130 pred_ids = torch.argmax(logits, dim=-1)
131 batch["pred_strings"] = processor.batch_decode(pred_ids)
132 return batch
133
134 result = test_dataset.map(evaluate, batched=True, batch_size=8)
135
136 print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
137 print("CER: {:2f}".format(100 * wer.compute(predictions=[" ".join(list(entry)) for entry in result["pred_strings"]], references=[" ".join(list(entry)) for entry in result["sentence"]])))
138
139 ```
140
141 **Test Result**: 18.996669 %
142
143
144 ## Training
145
146
147 The Common Voice train dataset was used for training. Also all of `CSS10 Greek` was used using the normalized transcripts.
148 During text preprocessing letter `ς` is normalized to `σ` the reason is that both letters sound the same with `ς` only used as the ending character of words. So, the change can be mapped up to proper dictation easily. I tried removing all accents from letters as well that improved `WER` significantly. The model was reaching `17%` WER easily without having converged. However, the text preprocessing needed to do after to fix transcrtiptions would be more complicated. A language model should fix things easily though. Another thing that could be tried out would be to change all of `ι`, `η` ... etc to a single character since all sound the same. similar for `o` and `ω` these should help the acoustic model part significantly since all these characters map to the same sound. But further text normlization would be needed.
149
150
151
152
153