w2v-bert-2.0-nepali-transliterator

w2v-bert-2.0-nepali-transliterator is a speech-to-text transliteration model that converts spoken Nepali audio into Romanized Nepali text. It leverages wav2vec-based embeddings combined with BERT-style processing to enhance accuracy in phonetic transliteration.

This model is a fine-tuned version of facebook/w2v-bert-2.0 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.2366
• Wer: 0.2786

Model Details

• Model Type: Speech-to-Text Transliteration Model
• Language: Nepali (Audio to Romanized Nepali Text)
• Dataset: Labeled Nepali speech dataset with Romanized text pairs
• Base Architecture: wav2vec 2.0 + BERT
• Task: Transliterating spoken Nepali into Romanized Nepali text
• Use Case: Assisting non-Devanagari users in understanding Nepali speech through Romanized output

Direct Use

The model can be used to:

• Convert Nepali speech into Romanized Nepali text
• Assist non-Devanagari users in understanding spoken Nepali
• Enable voice-based transliteration in chat applications

Out-of-Scope Use

• Not for General Nepali Speech-to-Text – This model specifically transliterates into Roman Nepali instead of generating text in Devanagari script.
• Not optimized for noisy environments – Performance may drop in low-quality or multi-speaker recordings.
• May not handle code-switching well – If Nepali is mixed with English or other languages, accuracy might decrease.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 6
• eval_batch_size: 8
• seed: 42
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.4387	1.3636	300	0.3972	0.5025
0.2712	2.7273	600	0.2779	0.3512
0.1335	4.0909	900	0.2366	0.2786

Framework versions

• Transformers 4.49.0
• Pytorch 2.5.1+cu121
• Datasets 3.2.0
• Tokenizers 0.21.0