This repo aims to provide a comparison among the ASR performance of wav2vec models finetuned in monolingual, multilingual and cross-lingual settings. For this purpose, we pretrain the wav2vec model on 6 languages: English, Bulgarian, Mandarin Chinese, Russian, Arabic and Japanese and finetuned on different languages according to the settings.

• Monolingual: Pretraining and finetuning on the same language.
• Cross-lingual: Pretraining on multiple languages and finetuning on one language.
• Multilingual: Pretraining on one language and finetuning on a different language.

We match the total length of the train, valid, test split of the corpus of each language to that of the English: 960 hours of train, 5 hours of valid and 10 hours of test.

• English: LibriSpeech
• Bulgarian:
  • VoxPopuli
  • GlobalPhone
• Mandarin Chinese:
  • Pretraining: United Nations Proceedings Speech
  • Fintuning: GlobalPhone
• Russian:
  • Pretraining: United Nations Proceedings Speech
  • Fintuning: LibriSpeech
• Arabic:
  • Pretraining: United Nations Proceedings Speech
  • Finetuning: GALE Phase 3 Arabic Broadcast News Speech Part 2
• Japanese
  • Pretraining: LaboroTVSpeech
  • Finetuning: Corpus of Spontaneous Japanese

• English
• Bulgarian
• Mandarin Chinese
• Russian
• Arabic
• Japanese
• Multilingual XLSR