This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - KY dataset. It achieves the following results on the validation set:
- Loss: 0.5497
- Wer: 0.2945
- Cer: 0.0791
For a description of the model architecture, see facebook/wav2vec2-xls-r-300m
The model vocabulary consists of the cyrillic alphabet with punctuation removed.
The kenlm language model is built using the text of the train and invalidated corpus splits.
This model is expected to be of some utility for low-fidelity use cases such as:
- Draft video captions
- Indexing of recorded broadcasts
The model is not reliable enough to use as a substitute for live captions for accessibility purposes, and it should not be used in a manner that would infringe the privacy of any of the contributors to the Common Voice dataset nor any other speakers.
The combination of
other of common voice official splits were used as training data. The half of the official
test split was used as validation data, as and the full
test set was used for final evaluation.
The featurization layers of the XLS-R model are frozen while tuning a final CTC/LM layer on the Kyrgiz CV8 example sentences. A ramped learning rate is used with an initial warmup phase of 500 steps, a max of 0.0001, and cooling back towards 0 for the remainder of the 8100 steps (300 epochs).
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 32
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 300.0
- mixed_precision_training: Native AMP
|Training Loss||Epoch||Step||Validation Loss||Wer||Cer|
- Transformers 4.17.0.dev0
- Pytorch 1.10.2+cu102
- Datasets 1.18.3
- Tokenizers 0.11.0
- Downloads last month