Wav2vec2 Large XLSR 53 Catalan

🚀 Catalan XLSR Wav2Vec Large 53

This is a fine - tuned XLSR Wav2Vec model for Catalan speech recognition, trained on the Common Voice dataset, offering a practical solution for automatic speech - recognition tasks in the Catalan language.

🚀 Quick Start

This model is a fine - tuned version of [facebook/wav2vec2 - large - xlsr - 53](https://huggingface.co/facebook/wav2vec2 - large - xlsr - 53) on Catalan using the Common Voice dataset. When using this model, ensure that your speech input is sampled at 16kHz.

✨ Features

• Trained on Common Voice 6 dataset for Catalan speech recognition.
• Can be used directly without a language model.
• Provides evaluation scripts for calculating Word Error Rate (WER).

📦 Installation

No specific installation steps are provided in the original document.

💻 Usage Examples

Basic 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("PereLluis13/Wav2Vec2-Large-XLSR-53-catalan")
model = Wav2Vec2ForCTC.from_pretrained("PereLluis13/Wav2Vec2-Large-XLSR-53-catalan")

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

# 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"] = 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])

Advanced Usage

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

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

test_dataset = load_dataset("common_voice", "ca", split="test") 
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("PereLluis13/Wav2Vec2-Large-XLSR-53-catalan") 
model = Wav2Vec2ForCTC.from_pretrained("PereLluis13/Wav2Vec2-Large-XLSR-53-catalan") 
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\;\:\"\“]'
resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
    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)

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

# Chunk WER computation due to memory issues, taken from https://huggingface.co/pcuenq/wav2vec2-large-xlsr-53-es
def chunked_wer(targets, predictions, chunk_size=None):                                          
    if chunk_size is None: return jiwer.wer(targets, predictions)                                
    start = 0                                                                                    
    end = chunk_size                                                                             
    H, S, D, I = 0, 0, 0, 0                                                                      
    while start < len(targets):                                                                  
        chunk_metrics = jiwer.compute_measures(targets[start:end], predictions[start:end])       
        H = H + chunk_metrics["hits"]                                                            
        S = S + chunk_metrics["substitutions"]                                                   
        D = D + chunk_metrics["deletions"]                                                       
        I = I + chunk_metrics["insertions"]                                                      
        start += chunk_size                                                                      
        end += chunk_size                                                                        
    return float(S + D + I) / float(H + S + D)

print("WER: {:2f}".format(100 * chunked_wer(result["sentence"], result["pred_strings"], chunk_size=4000)))

Test Result: 8.11 %

📚 Documentation

Disclaimer

This model was trained on Common Voice 6. If you need a Catalan model for ASR, I recommend checking [wav2vec2 - xls - r - 1b - ca - lm](https://huggingface.co/PereLluis13/wav2vec2 - xls - r - 1b - ca - lm) which is a 1b model with a LM on top trained on CV8+ with much better performance or [wav2vec2 - xls - r - 300m - ca - lm](https://huggingface.co/PereLluis13/wav2vec2 - xls - r - 300m - ca - lm) which has the same size (300m) as this model but trained on CV8+ and the same LM.

Training

The Common Voice train, validation datasets were used for training. At the second epoch training was halted due to a memory issue, and was continued with lower batch size, but acc. gradient steps were scaled to keep it at 32 batch size during all training. Then the model was trained for an additional 10 epochs where half the male samples were pitched up.

The script used for training can be found here. Slight modifications were done in order to speed up the ordering by length during training, which can be found [here](https://discuss.huggingface.co/t/spanish-asr-fine - tuning - wav2vec2/4586/6). Another version trained for Catalan can be found [here](https://huggingface.co/ccoreilly/wav2vec2 - large - xlsr - catala), which may be better than this one since it was trained with extra data and for longer time. However, since it used different splits that include part of the Common Voice test set, this version can be used to get a baseline on the Common Voice dataset.

📄 License

This model is licensed under the apache - 2.0 license.

Additional Information