metadata
language: tr
datasets:
- common_voice
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
license: apache-2.0
model-index:
- name: XLSR Wav2Vec2 Turkish by Davut Emre TASAR
results:
- task:
name: Speech Recognition
type: automatic-speech-recognition
dataset:
name: Common Voice tr
type: common_voice
args: tr
metrics:
- name: Test WER
type: wer
wav2vec-tr-lite-AG
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", "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)
### 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
| Training Loss | Epoch | Step | Validation Loss | Wer |
|:-------------:|:-----:|:----:|:---------------:|:------:|
| 0.4388 | 3.7 | 400 | 1.366 | 0.9701 |
| 0.3766 | 7.4 | 800 | 0.4914 | 0.5374 |
| 0.2295 | 11.11 | 1200 | 0.3934 | 0.4125 |
| 0.1121 | 14.81 | 1600 | 0.3264 | 0.2904 |
| 0.1473 | 18.51 | 2000 | 0.3103 | 0.2671 |
| 0.1013 | 22.22 | 2400 | 0.2589 | 0.2324 |
| 0.0704 | 25.92 | 2800 | 0.2826 | 0.2339 |
| 0.0537 | 29.63 | 3200 | 0.2704 | 0.2309 |
### Framework versions
- Transformers 4.12.0.dev0
- Pytorch 1.8.1
- Datasets 1.14.1.dev0
- Tokenizers 0.10.3