Fine-Tune W2V2-Bert for low-resource ASR with 🤗 Transformers

Hugging Face Blog · 2 years ago

Wav2Vec2-BERT, Meta's 580M-parameter audio model pre-trained on 4.5 million hours of multilingual speech data, can be fine-tuned for automatic speech recognition in low-resource languages using the Hugging Face Transformers library. The model achieves competitive word error rates on Mongolian ASR with just 14 hours of labeled training data from Common Voice 16.0. Fine-tuned Wav2Vec2-BERT runs 10 to 30 times faster than Whisper while using 2.5 times fewer computational resources, making it suitable for languages with minimal training data and inference constraints.