asr-wav2vec2-large-xlsr-53-greek
Audio transcription model for converting Greek audio to Greek text
Notes
huggingface model id: lighteternal/wav2vec2-large-xlsr-53-greek
Greek (el) version of the XLSR-Wav2Vec2 automatic speech recognition (ASR) model
By the Hellenic Army Academy and the Technical University of Crete
- language: el
- licence: apache-2.0
- dataset: CommonVoice (EL), 364MB: https://commonvoice.mozilla.org/el/datasets + CSS10 (EL), 1.22GB: https://github.com/Kyubyong/css10
- model: XLSR-Wav2Vec2, trained for 50 epochs
- metrics: Word Error Rate (WER)
Model description
UPDATE: We repeated the fine-tuning process using an additional 1.22GB dataset from CSS10.
Wav2Vec2 is a pretrained model for Automatic Speech Recognition (ASR) and was released in September 2020 by Alexei Baevski, Michael Auli, and Alex Conneau. Soon after the superior performance of Wav2Vec2 was demonstrated on the English ASR dataset LibriSpeech, Facebook AI presented XLSR-Wav2Vec2. XLSR stands for cross-lingual speech representations and refers to XLSR-Wav2Vec2`s ability to learn speech representations that are useful across multiple languages.
Similar to Wav2Vec2, XLSR-Wav2Vec2 learns powerful speech representations from hundreds of thousands of hours of speech in more than 50 languages of unlabeled speech. Similar, to BERT's masked language modeling, the model learns contextualized speech representations by randomly masking feature vectors before passing them to a transformer network.
This model was trained for 50 epochs on a single NVIDIA RTX 3080, for aprox. 8hrs.
How to use for inference:
For live demo, make sure that speech files are sampled at 16kHz.
Evaluation
The model can be evaluated on the Greek test data of Common Voice.
Test Result: 10.497628 %
How to use for training:
Metrics
| Metric | Value |
|---|---|
| Training Loss | 0.0545 |
| Validation Loss | 0.1661 |
| CER on CommonVoice Test (%) * | 2.8753 |
| WER on CommonVoice Test (%) * | 10.4976 |
* Reference transcripts were lower-cased and striped of punctuation and special characters.
Acknowledgement
The research work was supported by the Hellenic Foundation for Research and Innovation (HFRI) under the HFRI PhD Fellowship grant (Fellowship Number:50, 2nd call) Based on the tutorial of Patrick von Platen: https://huggingface.co/blog/fine-tune-xlsr-wav2vec2 Original colab notebook here: https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Fine_Tune_XLSR_Wav2Vec2_on_Turkish_ASR_with_%F0%9F%A4%97_Transformers.ipynb#scrollTo=V7YOT2mnUiea