wav2vec2-common_voice_13_0-eo-10_1

wav2vec2-common_voice_13_0-eo-10_1, an Esperanto speech recognizer

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the mozilla-foundation/common_voice_13_0 Esperanto dataset. It achieves the following results on the evaluation set:

• Loss: 0.0391
• Cer: 0.0098
• Wer: 0.0534

The first 10 examples in the evaluation set:

The differences in results for the above compared to the previous model (xekri/wav2vec2-common_voice_13_0-eo-10) are:

• eepokon -> epokon
• katedristo -> katedresto
• al la muzeo -> al l muzeo

Model description

See facebook/wav2vec2-large-xlsr-53. This model is a version of xekri/wav2vec2-common_voice_13_0-eo-10 trained for 5 more epochs.

Intended uses & limitations

Speech recognition for Esperanto. The base model was pretrained and finetuned on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16KHz.

The output is all lowercase, no punctuation.

Training and evaluation data

The training split was set to train while the eval split was set to validation. Some files were filtered out of the train and validation dataset due to bad data; see xekri/wav2vec2-common_voice_13_0-eo-3 for a detailed discussion. In summary, I used xekri/wav2vec2-common_voice_13_0-eo-3 as a detector to detect bad files, then hardcoded those files into the trainer code to be filtered out.

Training procedure

I used a modified version of run_speech_recognition_ctc.py for training. See run_speech_recognition_ctc.py in this repo.

The parameters to the trainer are in train.json in this repo.

The key changes between this training run and xekri/wav2vec2-common_voice_13_0-eo-3, aside from the filtering and use of the full training and validation sets are:

• Layer drop probability is 20%
• Train only for 5 epochs

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• layerdrop: 0.2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.29.2
• Pytorch 2.0.1+cu117
• Datasets 2.12.0
• Tokenizers 0.13.3