Merge branch 'main' of https://huggingface.co/KBLab/wav2vec2-large-xlsr-53-swedish into main

README.md

@@ -93,7 +93,7 @@ processor = Wav2Vec2Processor.from_pretrained("KBLab/wav2vec2-large-xlsr-53-swed
 model = Wav2Vec2ForCTC.from_pretrained("KBLab/wav2vec2-large-xlsr-53-swedish")
 model.to("cuda")
 
-chars_to_ignore_regex = '[,?.!\\\\\\\\\\\\\\\\-;:"“]'
+chars_to_ignore_regex = '[,?.!\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\-;:"“]'
 resampler = torchaudio.transforms.Resample(48_000, 16_000)
 
 # Preprocessing the datasets.
@@ -131,4 +131,4 @@ print("CER: {:2f}".format(100 * wer.compute(predictions=[" ".join(list(entry)) f
 
 ## Training
 
-The [NST Swedish Dictation](https://www.nb.no/sprakbanken/en/resource-catalogue/oai-nb-no-sbr-17/) was used for training. The [Fairseq](https://github.com/fairseq) scripts were used.
+First the XLSR model was further pre-trained for 50 epochs with a corpus consisting of 1000 hours spoken Swedish from various radio stations. Secondly [NST Swedish Dictation](https://www.nb.no/sprakbanken/en/resource-catalogue/oai-nb-no-sbr-17/) was used for fine tuning as well as [Common Voice](https://commonvoice.mozilla.org/en/datasets). Lastly only Common Voice dataset was used for final finetuning. The [Fairseq](https://github.com/fairseq) scripts were used.