Ukrainian STT model (with Language Model)

?? Join Ukrainian Speech Recognition Community - https://t.me/speech_recognition_uk

⭐ See other Ukrainian models - https://github.com/egorsmkv/speech-recognition-uk

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - UK dataset.

It achieves the following results on the evaluation set without the language model:

• Loss: 0.1875
• Wer: 0.2033
• Cer: 0.0384

Model description

On 100 test example the model shows the following results:

Without LM:

• WER: 0.1862
• CER: 0.0277

With LM:

• WER: 0.1218
• CER: 0.0190

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 20
• total_train_batch_size: 160
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.1.dev0
• Tokenizers 0.11.0

python eval.py --model_id Yehor/wav2vec2-xls-r-1b-uk-with-lm --dataset mozilla-foundation/common_voice_7_0 --config uk --split test

Eval results on Common Voice 7 "test" (WER):