|
--- |
|
language: ka |
|
datasets: |
|
- common_voice |
|
tags: |
|
- audio |
|
- automatic-speech-recognition |
|
- speech |
|
- xlsr-fine-tuning-week |
|
license: apache-2.0 |
|
widget: |
|
- label: Common Voice sample 566 |
|
src: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-georgian/resolve/main/sample566.flac |
|
- label: Common Voice sample 95 |
|
src: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-georgian/resolve/main/sample95.flac |
|
model-index: |
|
- name: XLSR Wav2Vec2 Georgian by Mehrdad Farahani |
|
results: |
|
- task: |
|
name: Speech Recognition |
|
type: automatic-speech-recognition |
|
dataset: |
|
name: Common Voice ka |
|
type: common_voice |
|
args: ka |
|
metrics: |
|
- name: Test WER |
|
type: wer |
|
value: 54.00 |
|
|
|
--- |
|
|
|
# Wav2Vec2-Large-XLSR-53-Georgian |
|
|
|
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) in Georgian using [Common Voice](https://huggingface.co/datasets/common_voice). When using this model, make sure that your speech input is sampled at 16kHz. |
|
|
|
## Usage |
|
The model can be used directly (without a language model) as follows: |
|
|
|
**Requirements** |
|
```bash |
|
# requirement packages |
|
!pip install git+https://github.com/huggingface/datasets.git |
|
!pip install git+https://github.com/huggingface/transformers.git |
|
!pip install torchaudio |
|
!pip install librosa |
|
!pip install jiwer |
|
``` |
|
|
|
|
|
**Prediction** |
|
```python |
|
import librosa |
|
import torch |
|
import torchaudio |
|
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor |
|
from datasets import load_dataset |
|
|
|
import numpy as np |
|
import re |
|
import string |
|
|
|
import IPython.display as ipd |
|
|
|
chars_to_ignore = [ |
|
",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "๏ฟฝ", |
|
"#", "!", "?", "ยซ", "ยป", "(", ")", "ุ", ",", "?", ".", "!", "-", ";", ":", '"', |
|
"โ", "%", "โ", "๏ฟฝ", "โ", "โฆ", "_", "โ", 'โ', 'โ' |
|
] |
|
chars_to_mapping = { |
|
"\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ", |
|
} |
|
|
|
def multiple_replace(text, chars_to_mapping): |
|
pattern = "|".join(map(re.escape, chars_to_mapping.keys())) |
|
return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text)) |
|
|
|
def remove_special_characters(text, chars_to_ignore_regex): |
|
text = re.sub(chars_to_ignore_regex, '', text).lower() + " " |
|
return text |
|
|
|
def normalizer(batch, chars_to_ignore, chars_to_mapping): |
|
chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]""" |
|
text = batch["sentence"].lower().strip() |
|
|
|
text = multiple_replace(text, chars_to_mapping) |
|
text = remove_special_characters(text, chars_to_ignore_regex) |
|
|
|
batch["sentence"] = text |
|
return batch |
|
|
|
|
|
def speech_file_to_array_fn(batch): |
|
speech_array, sampling_rate = torchaudio.load(batch["path"]) |
|
speech_array = speech_array.squeeze().numpy() |
|
speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000) |
|
|
|
batch["speech"] = speech_array |
|
return batch |
|
|
|
|
|
def predict(batch): |
|
features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True) |
|
|
|
input_values = features.input_values.to(device) |
|
attention_mask = features.attention_mask.to(device) |
|
|
|
with torch.no_grad(): |
|
logits = model(input_values, attention_mask=attention_mask).logits |
|
|
|
pred_ids = torch.argmax(logits, dim=-1) |
|
|
|
batch["predicted"] = processor.batch_decode(pred_ids)[0] |
|
return batch |
|
|
|
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
|
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian") |
|
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian").to(device) |
|
|
|
dataset = load_dataset("common_voice", "ka", split="test[:1%]") |
|
dataset = dataset.map( |
|
normalizer, |
|
fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping}, |
|
remove_columns=list(set(dataset.column_names) - set(['sentence', 'path'])) |
|
) |
|
|
|
dataset = dataset.map(speech_file_to_array_fn) |
|
result = dataset.map(predict) |
|
|
|
max_items = np.random.randint(0, len(result), 20).tolist() |
|
for i in max_items: |
|
reference, predicted = result["sentence"][i], result["predicted"][i] |
|
print("reference:", reference) |
|
print("predicted:", predicted) |
|
print('---') |
|
``` |
|
|
|
**Output:** |
|
```text |
|
reference: แแแแแแแกแขแ แแชแแฃแแ แชแแแขแ แ แฅแแแแฅแ แแแแจแแ |
|
predicted: แแแแแแแกแขแ แแชแแฃแแ แชแแแขแ แ แฅแแแแฅแ แแแแจแแ |
|
--- |
|
reference: แแแแแแแ แแแแแแแขแแก แแฏแแฎแจแ |
|
predicted: แแแแแแ แแแแแแแขแแก แแฏแแฎแจแ |
|
--- |
|
reference: แแฆแกแแแแจแแแแแ แ แแ แกแแแฆแแ แ แฌแแ แแแแแแแแก แแแ แแแแแแ แขแแแกแ แแ แฏแแ แฏ แฐแแ แแกแแแแก แแจแแแแ แแแแแแฃแ แแฃแแขแก |
|
predicted: แแฆแกแแแแจแแแแแแ แ แกแแแฆแ แ แแฌแแ แแแแแแแแก แแแ แแแแแ แแแแก แแ แฏแแ แฉแฎแแ แแกแแแแก แแจแแแแ แแแแแแฃแ แแฃแแแก |
|
--- |
|
reference: แแแ แซแแแแแแแ แฌแแ แแแแแชแแ แฅแแ แแฃแ แแแแแ |
|
predicted: แแแ แซแแแแแแแ แฌแแ แแ แแแชแแ แฅแแ แแฃแ แแแแแ |
|
--- |
|
reference: แแฆแแแ แแฃแแแ แแแแแกแ แแ แแแ แแแก แแแแขแแแแแแก แกแแแฆแแแ แแ |
|
predicted: แแฆแแแ แแฃแแแ แแแแแกแ แแ แแแ แแแก แแแแแแแแแแก แกแแแฆแแแ แแ |
|
--- |
|
reference: แแฅ แแแ แแแแฌแแแแก แกแแแฎแแขแแ แ แแแแแแแแแจแ แกแแแแช แกแแชแแชแฎแแแก แแแแแแแ แแฌแแแแ แแแแแแแแแฃแ แแแฆแแแฌแแแแแก |
|
predicted: แแฅ แแแ แแแแกแฌแ แแแแก แกแแแฎแแขแ แ แแแแแแแ แแจแแกแ แแ แชแแชแแชแฎแแแก แแแแแแแ แแฌแงแแแแ แแ แแแแแแฃแแแแแ แแแงแแแฌแแแแแแก |
|
--- |
|
reference: แแแแ แแกแ แแแแฎแแแแแ แจแแแแแแแแแแแแแ แแ แแแฅแขแแ แแก แแแฎแแแ แแแแ แกแแ แแฃแแ แแแแแแแแก แแแแแก แแแแแแแแ |
|
predicted: แแแแ แแก แแแ แฎแแแแแ แจแแแฃแแแแแแ แแแแ แแ แแแฅแขแแ แแก แแแฎแแแ แแแแช แกแแ แแฃแ แ แแแแแแแแก แแแแแก แแแแแแแแแ |
|
--- |
|
reference: แแแ แซแแแ แขแงแแแแแแ แแแญแ แแแแก แฌแแแแแฆแแแแ |
|
predicted: แแแแ แซแแขแ แขแงแแแแแแ แแแญแ แแแแก แฌแแแแแฆแแแ |
|
--- |
|
reference: แกแแแแแกแก แแฆแแแกแแแแแแแ แแ แแแกแแแแแแแ แแแแ แกแแ แแแแแ แแฅแแก |
|
predicted: แกแแแแแก แแฆแแแกแแแแแแแ แแ แแแกแแแแแ แแแแแ แกแแ แแแแแ แแฅแแก |
|
--- |
|
reference: แแแ แแแแแแ แแแแก แฅแแแแฅแแก แฉแ แแแแแแฆแแแกแแแแแ แแแฌแแแจแ |
|
predicted: แแแ แแแแแแ แแแแก แฅแแแแฅแแก แฉแ แแแแ แแฆแแแกแแแแแ แแแฌแแแจแ |
|
--- |
|
``` |
|
|
|
|
|
## Evaluation |
|
|
|
The model can be evaluated as follows on the Georgian test data of Common Voice. |
|
|
|
```python |
|
import librosa |
|
import torch |
|
import torchaudio |
|
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor |
|
from datasets import load_dataset, load_metric |
|
|
|
import numpy as np |
|
import re |
|
import string |
|
|
|
|
|
chars_to_ignore = [ |
|
",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "๏ฟฝ", |
|
"#", "!", "?", "ยซ", "ยป", "(", ")", "ุ", ",", "?", ".", "!", "-", ";", ":", '"', |
|
"โ", "%", "โ", "๏ฟฝ", "โ", "โฆ", "_", "โ", 'โ', 'โ' |
|
] |
|
chars_to_mapping = { |
|
"\u200c": " ", "\u200d": " ", "\u200e": " ", "\u200f": " ", "\ufeff": " ", |
|
} |
|
|
|
def multiple_replace(text, chars_to_mapping): |
|
pattern = "|".join(map(re.escape, chars_to_mapping.keys())) |
|
return re.sub(pattern, lambda m: chars_to_mapping[m.group()], str(text)) |
|
|
|
def remove_special_characters(text, chars_to_ignore_regex): |
|
text = re.sub(chars_to_ignore_regex, '', text).lower() + " " |
|
return text |
|
|
|
def normalizer(batch, chars_to_ignore, chars_to_mapping): |
|
chars_to_ignore_regex = f"""[{"".join(chars_to_ignore)}]""" |
|
text = batch["sentence"].lower().strip() |
|
|
|
text = multiple_replace(text, chars_to_mapping) |
|
text = remove_special_characters(text, chars_to_ignore_regex) |
|
|
|
batch["sentence"] = text |
|
return batch |
|
|
|
|
|
def speech_file_to_array_fn(batch): |
|
speech_array, sampling_rate = torchaudio.load(batch["path"]) |
|
speech_array = speech_array.squeeze().numpy() |
|
speech_array = librosa.resample(np.asarray(speech_array), sampling_rate, 16_000) |
|
|
|
batch["speech"] = speech_array |
|
return batch |
|
|
|
|
|
def predict(batch): |
|
features = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True) |
|
|
|
input_values = features.input_values.to(device) |
|
attention_mask = features.attention_mask.to(device) |
|
|
|
with torch.no_grad(): |
|
logits = model(input_values, attention_mask=attention_mask).logits |
|
|
|
pred_ids = torch.argmax(logits, dim=-1) |
|
|
|
batch["predicted"] = processor.batch_decode(pred_ids)[0] |
|
return batch |
|
|
|
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
|
processor = Wav2Vec2Processor.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian") |
|
model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-georgian").to(device) |
|
|
|
dataset = load_dataset("common_voice", "ka", split="test") |
|
dataset = dataset.map( |
|
normalizer, |
|
fn_kwargs={"chars_to_ignore": chars_to_ignore, "chars_to_mapping": chars_to_mapping}, |
|
remove_columns=list(set(dataset.column_names) - set(['sentence', 'path'])) |
|
) |
|
|
|
dataset = dataset.map(speech_file_to_array_fn) |
|
result = dataset.map(predict) |
|
|
|
wer = load_metric("wer") |
|
|
|
print("WER: {:.2f}".format(100 * wer.compute(predictions=result["predicted"], references=result["sentence"]))) |
|
``` |
|
|
|
|
|
**Test Result**: |
|
- WER: 54.00% |
|
|
|
|
|
## Training & Report |
|
The Common Voice `train`, `validation` datasets were used for training. |
|
|
|
You can see the training states [here](https://wandb.ai/m3hrdadfi/finetuned_wav2vec_xlsr_georgian/reports/Fine-Tuning-for-Wav2Vec2-Large-XLSR-53-Georgian--Vmlldzo1NTg5MDQ?accessToken=rsmd0p83iln13yq23b9kzj8bim6nco21w8cqn2tb19v51okakqk92c71h6hbxmfj) |
|
|
|
The script used for training can be found [here](https://colab.research.google.com/github/m3hrdadfi/notebooks/blob/main/Fine_Tune_XLSR_Wav2Vec2_on_Georgian_ASR_with_%F0%9F%A4%97_Transformers_ipynb.ipynb) |
