# wav2vec2-large-xlsr-galician
---
language: gl
datasets:
- OpenSLR 77
- mozilla-foundation common_voice_8_0
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: Galician wav2vec2-large-xlsr-galician
  results:
  - task: 
      name: Speech Recognition
      type: automatic-speech-recognition
    dataset_1:
      name: OpenSLR
      type: openslr
      args: gl
    dataset_2:
      name: mozilla-foundation
      type: common voice
      args: gl
    metrics:
       - name: Test WER
         type: wer
         value: 7.12
---

# Model

Fine-tuned model for Galician language

Based on the [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) self-supervised model 
Fine-tune with audio labelled from [OpenSLR](https://openslr.org/77/) and Mozilla [Common_Voice](https://commonvoice.mozilla.org/gl) (both datasets previously refined)

Check training metrics to see results

# Testing

Make sure that the audio speech input is sampled at 16kHz (mono).


```python
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

model = Wav2Vec2ForCTC.from_pretrained("ifrz/wav2vec2-large-xlsr-galician")
processor = Wav2Vec2Processor.from_pretrained("ifrz/wav2vec2-large-xlsr-galician")

# Reading taken audio clip
import librosa, torch
audio, rate = librosa.load("./gl_test_1.wav", sr = 16000)

# Taking an input value
input_values = processor(audio, sampling_rate=16_000, return_tensors = "pt", padding="longest").input_values
# Storing logits (non-normalized prediction values)
logits = model(input_values).logits
# Storing predicted ids
prediction = torch.argmax(logits, dim = -1)

# Passing the prediction to the tokenzer decode to get the transcription
transcription = processor.batch_decode(prediction)[0]
print(transcription)
```