README.md

---
language: uk
dataset: common_voice
metrics: wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: Ukrainian XLSR Wav2Vec2 Large 53
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Common Voice uk
      type: common_voice
      args: uk
    metrics:
       - name: Test WER
         type: wer
         value: 29.89
---

# Wav2Vec2-Large-XLSR-53-Ukrainian

Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Ukrainian using the [Common Voice](https://huggingface.co/datasets/common_voice) and sample of [M-AILABS Ukrainian Corpus](https://www.caito.de/2019/01/the-m-ailabs-speech-dataset/) datasets.

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:

```python

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

test_dataset = load_dataset("common_voice", "uk", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("arampacha/wav2vec2-large-xlsr-ukrainian")
model = Wav2Vec2ForCTC.from_pretrained("arampacha/wav2vec2-large-xlsr-ukrainian")

# Preprocessing the datasets.
# We need to read the aduio files as arrays

def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = torchaudio.transforms.Resample(sampling_rate, 16_000)(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)

with torch.no_grad():
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits

predicted_ids = torch.argmax(logits, dim=-1)

print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])

```

## Evaluation

The model can be evaluated as follows on the Ukrainian test data of Common Voice.

```python
import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re

test_dataset = load_dataset("common_voice", "uk", split="test")

wer = load_metric("wer")
processor = Wav2Vec2Processor.from_pretrained("arampacha/wav2vec2-large-xlsr-ukrainian")
model = Wav2Vec2ForCTC.from_pretrained("arampacha/wav2vec2-large-xlsr-ukrainian")
model.to("cuda")

chars_to_ignore = [",", "?", ".", "!", "-", ";", ":", '""', "%", "'", '"', "�", '«', '»', '—', '…', '(', ')', '*', '”', '“']
chars_to_ignore_regex = f'[{"".join(chars_to_ignore)}]'
resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays and normalize charecters
def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(re.compile("['`]"), '’', batch['sentence'])
    batch["sentence"] = re.sub(re.compile(chars_to_ignore_regex), '', batch["sentence"]).lower().strip()
    batch["sentence"] = re.sub(re.compile('i'), 'і', batch['sentence'])
    batch["sentence"] = re.sub(re.compile('o'), 'о', batch['sentence'])
    batch["sentence"] = re.sub(re.compile('a'), 'а', batch['sentence'])
    batch["sentence"] = re.sub(re.compile('ы'), 'и', batch['sentence'])
    batch["sentence"] = re.sub(re.compile("–"), '', batch['sentence'])
    batch['sentence'] = re.sub('  ', ' ', batch['sentence'])
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = torchaudio.transforms.Resample(sampling_rate, 16_000)(speech_array).squeeze().numpy()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

def evaluate(batch):
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
    with torch.no_grad():
        logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits

    pred_ids = torch.argmax(logits, dim=-1)
    batch["pred_strings"] = processor.batch_decode(pred_ids)
    return batch

result = test_dataset.map(evaluate, batched=True, batch_size=8)

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))

```

**Test Result**: 29.89

## Training

The Common Voice `train`, `validation` and the M-AILABS Ukrainian corpus.

The script used for training will be available [here](https://github.com/arampacha/hf-sprint-xlsr) soon.