Wav2vec2 Large Xls R 300m Bg

🚀 XLS-R-300M - Bulgarian

This model is a fine - tuned version of [facebook/wav2vec2 - xls - r - 300m](https://huggingface.co/facebook/wav2vec2 - xls - r - 300m) on the MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - BG dataset. It's designed for automatic speech recognition, offering a practical solution for transcribing Bulgarian speech.

🚀 Quick Start

Evaluation Commands

1. To evaluate on mozilla - foundation/common_voice_8_0 with split test

python eval.py --model_id anuragshas/wav2vec2-large-xls-r-300m-bg --dataset mozilla-foundation/common_voice_8_0 --config bg --split test

2. To evaluate on speech - recognition - community - v2/dev_data

python eval.py --model_id anuragshas/wav2vec2-large-xls-r-300m-bg --dataset speech-recognition-community-v2/dev_data --config bg --split validation --chunk_length_s 5.0 --stride_length_s 1.0

Inference With LM

import torch
from datasets import load_dataset
from transformers import AutoModelForCTC, AutoProcessor
import torchaudio.functional as F
model_id = "anuragshas/wav2vec2-large-xls-r-300m-bg"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "bg", split="test", streaming=True, use_auth_token=True))
sample = next(sample_iter)
resampled_audio = F.resample(torch.tensor(sample["audio"]["array"]), 48_000, 16_000).numpy()
model = AutoModelForCTC.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id)
input_values = processor(resampled_audio, return_tensors="pt").input_values
with torch.no_grad():
    logits = model(input_values).logits
transcription = processor.batch_decode(logits.numpy()).text
# => "и надутият му ката блоонкурем взе да се събира"

✨ Features

• Fine - Tuned Model: Based on [facebook/wav2vec2 - xls - r - 300m](https://huggingface.co/facebook/wav2vec2 - xls - r - 300m), fine - tuned on the MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - BG dataset for better performance in Bulgarian speech recognition.
• Multiple Evaluation Metrics: Evaluated on multiple datasets with metrics like WER (Word Error Rate) and CER (Character Error Rate) to measure performance accurately.

📦 Installation

No specific installation steps are provided in the original document, so this section is skipped.

📚 Documentation

Model Performance

This model achieves the following results on the evaluation set:

• Loss: 0.2473
• Wer: 0.3002

Model Index

Training and Evaluation

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 7.5e - 05
• train_batch_size: 32
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training Results

Framework Versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0

Evaluation Results

🔧 Technical Details

The model is a fine - tuned version of [facebook/wav2vec2 - xls - r - 300m](https://huggingface.co/facebook/wav2vec2 - xls - r - 300m). The fine - tuning process involves adjusting the model's parameters on the MOZILLA - FOUNDATION/COMMON_VOICE_8_0 - BG dataset. The use of specific hyperparameters during training, such as the learning rate, batch size, and optimizer, is crucial for achieving good performance.

📄 License

The model is released under the Apache - 2.0 license.