AndrewMcDowell/wav2vec2-xls-r-300m-japanese

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMONVOICE8_0 - JA dataset.

Kanji are converted into Hiragana using the pykakasi library during training and evaluation. The model can output both Hiragana and Katakana characters. Since there is no spacing, WER is not a suitable metric for evaluating performance and CER is more suitable.

On mozilla-foundation/commonvoice8_0 it achieved:

• —cer: 23.64%

On speech-recognition-community-v2/dev_data it achieved:

• —cer: 30.99%

It achieves the following results on the evaluation set:

• —Loss: 0.5212
• —Wer: 1.3068

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 7.5e-05
• —trainbatchsize: 48
• —evalbatchsize: 8
• —seed: 42
• —optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• —lrschedulertype: linear
• —lrschedulerwarmup_steps: 2000
• —num_epochs: 50.0
• —mixedprecisiontraining: Native AMP

Training results

Framework versions

• —Transformers 4.17.0.dev0
• —Pytorch 1.10.2+cu102
• —Datasets 1.18.2.dev0
• —Tokenizers 0.11.0

Evaluation Commands

1. 1.To evaluate on mozilla-foundation/common_voice_8_0 with split test

python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-300m-japanese --dataset mozilla-foundation/common_voice_8_0 --config ja --split test --log_outputs

1. 1.To evaluate on mozilla-foundation/common_voice_8_0 with split test

python ./eval.py --model_id AndrewMcDowell/wav2vec2-xls-r-300m-japanese --dataset speech-recognition-community-v2/dev_data --config de --split validation --chunk_length_s 5.0 --stride_length_s 1.0