Wav2Vec2-Large-XLSR-Persian-ShEMO

<!-- 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. -->

Wav2Vec2 Large XLSR Persian ShEMO

This model is a fine-tuned version of masoudmzb/wav2vec2-xlsr-multilingual-53-fa on the ShEMO dataset for speech recognition in Persian (Farsi). When using this model, make sure that your speech input is sampled at 16 kHz.

It achieves the following results:

• Loss on ShEMO train set: 0.7618
• Loss on ShEMO dev set: 0.6728
• WER on ShEMO train set: 30.47
• WER on ShEMO dev set: 32.85
• WER on Common Voice 13 test set: 19.21

Evaluation results 🧪

As you can see, my model performs better in maximum case :D

Training procedure 🏋️

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 2000
• mixed_precision_training: Native AMP

You may need gradient_accumulation because you need more batch size.

Training log 📉

Hyperparameter tuning 🔧

Several models with differet hyperparameters were trained. The following figures show the training process for three of them. 20_2000_1e-5_hp-mehrdad is the current model (lnxdx/Wav2Vec2-Large-XLSR-Persian-ShEMO) and it's hyperparameters are:

model = Wav2Vec2ForCTC.from_pretrained(
    model_name_or_path if not last_checkpoint else last_checkpoint,
    # hp-mehrdad: Hyperparams of 'm3hrdadfi/wav2vec2-large-xlsr-persian-v3'
    attention_dropout = 0.05316,
    hidden_dropout    = 0.01941,
    feat_proj_dropout = 0.01249,
    mask_time_prob    = 0.04529,
    layerdrop         = 0.01377,
    ctc_loss_reduction = 'mean',
    ctc_zero_infinity = True,
)

The hyperparameters of 19_2000_1e-5_hp-base are:

model = Wav2Vec2ForCTC.from_pretrained(
    model_name_or_path if not last_checkpoint else last_checkpoint,
    # hp-base: Hyperparams simmilar to ('facebook/wav2vec2-large-xlsr-53' or 'facebook/wav2vec2-xls-r-300m')
    attention_dropout = 0.1,
    hidden_dropout    = 0.1,
    feat_proj_dropout = 0.1,
    mask_time_prob    = 0.075,
    layerdrop         = 0.1,
    ctc_loss_reduction = 'mean',
    ctc_zero_infinity = True,
)

And the hyperparameters of 22_2000_1e-5_hp-masoud are:

model = Wav2Vec2ForCTC.from_pretrained(
    model_name_or_path if not last_checkpoint else last_checkpoint,
    # hp-masoud: Hyperparams of 'masoudmzb/wav2vec2-xlsr-multilingual-53-fa'
	attention_dropout = 0.2,
	hidden_dropout    = 0.2,
	feat_proj_dropout = 0.1,
	mask_time_prob    = 0.2,
	layerdrop         = 0.2,
	ctc_loss_reduction = 'mean',
	ctc_zero_infinity = True,
)

Learning rate is 1e-5 for all three models.

As you can see this model performs better with WER metric on validation(evaluation) set.

📒 The script used for training can be found here.

📒 The script used for evaluating on ShEMO and Common Voice can be found here.

Check out this blog for more information.

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.0+cu118
• Datasets 2.15.0
• Tokenizers 0.15.0

Contact us 🛎

If you have any technical question regarding the model, pretraining, code or publication, please create an issue in the repository. This is the best way to reach us.

Citation ↩️

TO DO!

Fine-tuned with ❤️ without ☕︎