Wav2Vec2 Indonesian Javanese Sundanese

This model is an automatic speech recognition (ASR) model fine-tuned from facebook/wav2vec2-large-xlsr-53 to support three major Indonesian languages: Indonesian, Javanese, and Sundanese. It was trained on the Indonesian Common Voice dataset, High-quality TTS data for Javanese (SLR41), and High-quality TTS data for Sundanese (SLR44). The model accepts speech input sampled at 16 kHz and outputs transcribed text.

Key strengths include robust multilingual performance across languages that are not natively covered by most single-language ASR systems. Monolingual baselines trained on just one language degrade severely on the other two (e.g., an Indonesian-only model achieves 78.06% WER on Javanese and 64.04% WER on Sundanese), whereas this multilingual model delivers competitive word error rates (WER) on all three.

Notable benchmark results on the Indonesian Common Voice 6.1 test set:

• Without a language model (300-epoch multilingual model): Indonesian WER 11.57%, Javanese WER 16.57%, Sundanese WER 6.72%.
• With a KenLM language model trained on Common Voice text: Indonesian WER drops to 4.27%; with a Wikipedia-based LM it reaches 5.15%.
• For comparison, Google Speech-to-Text reports a WER of 9.22% on the same Indonesian test set—the model with LM outperforms it.

The following figure compares models by WER on the Indonesian test set (without language model):

All models were trained for 200 epochs; the reported 11.57% WER on Indonesian corresponds to the 300-epoch multilingual model. The 4.27% WER with LM was achieved using dataset version v7.