Wav2Vec2-Large-XLSR-Català

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Catalan language using the Common Voice and ParlamentParla datasets.

Attention: The split train/dev/test used does not fully map with the CommonVoice 6.1 dataset. A custom split was used combining both the CommonVoice and ParlamentParla dataset and can be found here. Evaluating on the CV test dataset will produce a biased WER as 1144 audio files of that dataset were used in training/evaluation of this model. WER was calculated using this test.csv which was not seen by the model during training/evaluation.

You can find training and evaluation scripts in the github repository ccoreilly/wav2vec2-catala

When using this model, make sure that your speech input is sampled at 16kHz.

Results

Word error rate was evaluated on the following datasets unseen by the model:

Usage

The model can be used directly (without a language model) as follows:

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

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

processor = Wav2Vec2Processor.from_pretrained("ccoreilly/wav2vec2-large-xlsr-catala") 
model = Wav2Vec2ForCTC.from_pretrained("ccoreilly/wav2vec2-large-xlsr-catala")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
	speech_array, sampling_rate = torchaudio.load(batch["path"])
	batch["speech"] = resampler(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])