Wav2Vec2_XLS-R-300m_Nepali_ASR

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on:

• [Large Nepali ASR training data set from OpenSLR (SLR-54)] (https://www.openslr.org/54/)
• [Common Voice Corpus 17.0] (https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0)

Model description

The model is a fine-tuned version of Wav2Vec2 XLS-R 300 million parameters version fine-tuned for Nepali Automatic Speech Recognition. The reported results are on the OpenSLR test split.

• WER on OpenSLR: 16.82%
• CER on OpenSLR: 2.72%

Intended uses & limitations

• Research on Nepali ASR
• Transcriptions on Nepali audio
• Further Fine-tuning

• Limitations:

• The model is trained on the OpenSLR Nepali ASR dataset which upon inspection was found to be quite noisy and inconsistent.
• Due to resources limitations, utterances longer than 5 sec have been filtered out from the dataset during training and evaluation.
• Numerals have been filtered out as well.
• The vocabulary doesn't contain all the Nepali alphabets.
• Might perform poorly on audio segments longer than 5 seconds. Or, needs some mechanism to segment audio into 5 seconds chunks which might increase processing time.
• May struggle with background noises and overlapping speech.

Training and evaluation data

Common Voice v17.0

• This model has been fine-tuned on OpenSLR-54 (Nepali ASR training dataset) and CommonVoice Corpus v17.0
• Initially, the model was trained on CommonVoice v17.0 ne-NP which consists of about 2 hours of voice data of which 1 hours have been manually validated.
• We combined the validated and other split first since the dataset is very small. So, we had a total of 1337 utterances.
• We have preprocessed the data by removing all punctuations and symbols.
• Then, we used 80% of the total utterances for training and 10% for evaluation.
• And, we used the test split consisting of 217 utterances for testing. (It might have been present in the train split as well.)
• It was trained for 30 epochs. The WER started fluctuating around 37% to 39%.

OpenSLR Nepali ASR training data

• Then, it was further trained on the larger OpenSLR Nepali ASR training dataset which has 157,000 utterances.
• Firstly, the numerals were removed as the utterances were inconsistent with transcriptions.
• And, segments longer than 5 seconds were removed because of resource limitations.
• Less frequently used 'alphabets' were removed to reduce the vocabulary size.
• Finally, we ended up with 136083 utterances for whole dataset. The dataset has been uploaded here.
• 80% was used for training, 10% for evaluation and 10% for testing.

Training procedure

Training on CommonVoice 17.0

The following hyperparameters were used during training:

• learning_rate: 3e-04
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 400
• num_epochs: 30
• mixed_precision_training: Native AMP

Initial Training on OpenSLR-54 for 16 epochs

The following hyperparameters were used:

• learning_rate: 3e-04
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• warmup_steps: 500
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 16
• mixed_precision_training: Native AMP

Further Training on OpenSLR-54 for further 3 epochs

We used the following:

• learning_rate: 2e-5
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 700
• num_epochs: 3
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.1+cu121
• Datasets 3.0.0
• Tokenizers 0.19.1