Wav2vec2-xls-r-phoneme-300m-sv

Note: The tokenizer was created from the official Swedish phoneme vocabulary as defined here: https://github.com/microsoft/UniSpeech/blob/main/UniSpeech/examples/unispeech/data/sv/phonesMatches_reduced.json

One can simply download the file, rename it to vocab.json and load a Wav2Vec2PhonemeCTCTokenizer.from_pretrained("./directory/with/vocab.json/").

This model is a fine-tuned version of wav2vec2-xls-r-300m on the COMMON_VOICE - SV-SE dataset.

It achieves the following results on the evaluation set:

• Loss: 0.9707
• PER: 0.2215

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: 0.0005
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• total_train_batch_size: 32
• total_eval_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: 20.0
• mixed_precision_training: Native AMP

Training results

See Tensorboard traces

Framework versions

• Transformers 4.13.0.dev0
• Pytorch 1.8.1
• Datasets 1.16.2.dev0
• Tokenizers 0.10.3