metadata

license: apache-2.0
tags:
  - generated_from_trainer
datasets:
  - common_voice_8_0
metrics:
  - wer
model-index:
  - name: wav2vec2-large-xls-r-1b-frisian-cv-8-1h
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: common_voice_8_0
          type: common_voice_8_0
          config: fy-NL
          split: validation
          args: fy-NL
        metrics:
          - name: Wer
            type: wer
            value: 0.23732323953720896
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: common_voice_8_0
          type: common_voice_8_0
          config: fy-NL
          split: test
          args: fy-NL
        metrics:
          - name: Wer
            type: wer
            value: 0.25404682757623936

wav2vec2-large-xls-r-1b-frisian-cv-8-1h

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the common_voice_8_0 dataset. It achieves the following results on the evaluation set:

Loss: 0.4120
Wer: 0.2373

And on the test set:

Wer: 0.2540

Model description

This model has been developed for my Master's thesis in "Voice Technology" at Rijksuniversiteit Groningen - Campus Fryslân. It corresponds to experiment 4 where I use as training set 1 hour of Frisian speech randomly selected from all validated data except the test and evaluation sets.

Intended uses & limitations

The intended use is for recognizing Frisian speech.

Limitations include no LM rescoring and using version 8.0 of Common Voice instead of 13.0.

Training and evaluation data

The evaluation split used is the one available in the Common Voice 8.0 Frisian subset. The train split is 1 hour of Frisian randomly selected from validated data except for the recordings from test and evaluation splits.

Training procedure

The script used for training this model can be found in this GitHub repository: link.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 6e-05
train_batch_size: 32
eval_batch_size: 8
seed: 42
optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
num_epochs: 80
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
6.2987	4.35	100	3.0210	1.0
3.1424	8.7	200	2.9611	1.0
2.6299	13.04	300	0.9929	0.8377
1.3134	17.39	400	0.5679	0.5264
0.9747	21.74	500	0.4516	0.3764
0.8755	26.09	600	0.4515	0.3403
0.7227	30.43	700	0.4169	0.3211
0.6634	34.78	800	0.4159	0.2962
0.5568	39.13	900	0.4081	0.2795
0.7943	43.48	1000	0.4090	0.2709
0.5537	47.83	1100	0.4239	0.2649
0.5596	52.17	1200	0.4029	0.2561
0.5523	56.52	1300	0.4073	0.2524
0.4579	60.87	1400	0.4098	0.2470
0.6477	65.22	1500	0.4099	0.2446
0.4957	69.57	1600	0.4167	0.2475
0.3246	73.91	1700	0.4146	0.2389
0.3937	78.26	1800	0.4120	0.2373

Framework versions

Transformers 4.28.1
Pytorch 2.0.0+cu117
Datasets 2.11.0
Tokenizers 0.13.3