Wav2vec Tr Lite AG

🚀 wav2vec-tr-lite-AG

This is a model for automatic speech recognition in Turkish, leveraging the XLSR Wav2Vec2 architecture. It offers a practical solution for speech - to - text conversion tasks in the Turkish language.

🚀 Quick Start

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", "tr", split="test[:2%]") 

processor = Wav2Vec2Processor.from_pretrained("emre/wav2vec-tr-lite-AG")
model = Wav2Vec2ForCTC.from_pretrained("emre/wav2vec-tr-lite-AG")

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

💻 Usage Examples

Basic Usage

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

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

processor = Wav2Vec2Processor.from_pretrained("emre/wav2vec-tr-lite-AG")
model = Wav2Vec2ForCTC.from_pretrained("emre/wav2vec-tr-lite-AG")

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

Advanced Usage

# The following hyperparameters were used during training:
# - learning_rate: 0.00005
# - train_batch_size: 2
# - eval_batch_size: 8
# - seed: 42
# - distributed_type: multi - GPU
# - num_devices: 2
# - gradient_accumulation_steps: 8
# - total_train_batch_size: 32
# - total_eval_batch_size: 16
# - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e - 08
# - lr_scheduler_type: linear
# - lr_scheduler_warmup_steps: 500
# - num_epochs: 30.0
# - mixed_precision_training: Native AMP

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

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

processor = Wav2Vec2Processor.from_pretrained("emre/wav2vec-tr-lite-AG")
model = Wav2Vec2ForCTC.from_pretrained("emre/wav2vec-tr-lite-AG")

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

📚 Documentation

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.00005
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• distributed_type: multi - GPU
• num_devices: 2
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e - 08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30.0
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.12.0.dev0
• Pytorch 1.8.1
• Datasets 1.14.1.dev0
• Tokenizers 0.10.3

📄 License

This project is licensed under the Apache 2.0 license.

🔍 Model Information