mHuBERT-147-br

This model is a fine-tuned version of utter-project/mHuBERT-147 on Mozilla Common Voice 15 Breton dataset and Roadennoù dataset. It achieves the following results on the validation set:

• Loss: 0.7331
• Wer: 50.09
• Cer: 16.45

Model description

This model was trained to assess the performance of mHubert-147 for finetuning a Breton ASR model.

Intended uses & limitations

This model is a research model and shouldn't be used in production.

Training and evaluation data

90% of the Roadennoù dataset was used for training, the remaining 10% was used for validation in addition to MCV15-br validation dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3.8e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_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: 52
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.39.1
• Pytorch 2.0.1+cu117
• Datasets 2.18.0
• Tokenizers 0.15.2