--- language: lt datasets: - common_voice tags: - audio - automatic-speech-recognition - speech - xlsr-fine-tuning-week license: apache-2.0 widget: - label: Common Voice sample 11 src: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-lithuanian/resolve/main/sample11.flac - label: Common Voice sample 74 src: https://huggingface.co/m3hrdadfi/wav2vec2-large-xlsr-lithuanian/resolve/main/sample74.flac model-index: - name: XLSR Wav2Vec2 Lithuanian by Mehrdad Farahani results: - task: name: Speech Recognition type: automatic-speech-recognition dataset: name: Common Voice lt type: common_voice args: lt metrics: - name: Test WER type: wer value: 55.13 --- # Wav2Vec2-Large-XLSR-53-Lithuanian Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) in Lithuanian 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 = [ ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�", "#", "!", "?", "«", "»", "(", ")", "؛", ",", "?", ".", "!", "-", ";", ":", '"', "“", "%", "‘", "�", "–", "…", "_", "”", '“', '„' ] 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-lithuanian") model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-lithuanian").to(device) dataset = load_dataset("common_voice", "lt", 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), 10).tolist() for i in max_items: reference, predicted = result["sentence"][i], result["predicted"][i] print("reference:", reference) print("predicted:", predicted) print('---') ``` **Output:** ```text reference: vaistinių augalų rinkimas predicted: vaistiniau augalų rinkimas --- reference: penkias iš jų jie įrašė į kasetę ir pradėjo ieškoti dainininko predicted: penkese iš šių ie ji rašę į kasėtę ir pradėjojos škoti dainininklo --- reference: iki mūsų eros pradžios germanija buvo etniškai mišri predicted: ikimūsų eros pradžios germanija buvo etniškai mišri --- reference: pietrytiniame krante netoli užtvankos įrengtas paplūdimys predicted: pietrytiname klante netoli užtvangos įrengtas paplūdimys --- reference: minta smulkiais bestuburiais predicted: minta smulkiais bestubūriais --- reference: jie gyveno ganykloms tinkamose žemėse tarp miestų visoje vakarų afrikoje predicted: je gyveno gonykloms tinkamase žemėse tarp miestų visojava karų ardykoje --- reference: prefektūra yra kazachstano pasienyje predicted: prefektūrą yra kazahstano pasienyje --- reference: į šiaurę ir pietus nuo kaimo buvusios senovės gyvenvietės predicted: į šiaurė ir pietus nuo kaimo buvusius senovės gyvenvietis --- reference: tai vienintelis lietuvos teritorijoje aptiktas toks vertingas zoologinis radinys predicted: tai vieninteris lietuvos ritorijoje aptiktas toksvirtingas zologinis radinys --- reference: pagrindinis partijos reikalavimas buvo vėl sušaukti steigiamąjį susirinkimą ir įtvirtinti rusijoje demokratiją predicted: pagrindinis partijos reikalavimas buvo vėl sušouktis steigiamajį susirinkimą ir įtvyrtinti ir rusijoje demokratije --- ``` ## Evaluation The model can be evaluated as follows on the Persian (Farsi) 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 = [ ",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�", "#", "!", "?", "«", "»", "(", ")", "؛", ",", "?", ".", "!", "-", ";", ":", '"', "“", "%", "‘", "�", "–", "…", "_", "”", '“', '„' ] 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-lithuanian") model = Wav2Vec2ForCTC.from_pretrained("m3hrdadfi/wav2vec2-large-xlsr-lithuanian").to(device) dataset = load_dataset("common_voice", "lt", 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: 55.13% ## 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_lithuanian/reports/Fine-Tuning-for-Wav2Vec2-Large-XLSR-53-Lithuanian--Vmlldzo1NjEyNTk?accessToken=hflswq0fdfz1yux9wqfxkdodpu41oqtsgg7et8dld0x05epqwv5nq7nvlhpp09cs) The script used for training can be found [here](https://colab.research.google.com/github/m3hrdadfi/notebooks/blob/main/Fine_Tune_XLSR_Wav2Vec2_on_Lithuanian_ASR_with_%F0%9F%A4%97_Transformers_ipynb.ipynb)