setup_modifications.md

Modifications of the setup script

Contains the issues that lead to the modification of the original setup and training script.

Issue

Dependency librosa is not installed.

ModuleNotFoundError: No module named 'librosa'

Workaround

pip install librosa

Issue

OSError: sndfile library not found

Workaround

sudo apt install libsndfile1

Issue

ValueError: Unknown split "train.100". Should be one of ['train', 'test', 'validation', 'other', 'invalidated'].

Solution

Delete this line (the value of the parameter will default to 'train')

--train_split_name="train.100" \

and add this one to specify the train/validation split:

--validation_split_percentage="5" \

Issue

File "./run_wav2vec2_pretrain_flax.py", line 315, in prepare_dataset
batch["speech"], _ = librosa.load(batch[data_args.speech_file_column], sr=feature_extractor.sampling_rate)
KeyError: 'file'

Solution

Specify the speech file column in the training script by adding:

--speech_file_column="path" \

Issue

RuntimeError: Error opening '/home/mariagrandury/.cache/huggingface/datasets/downloads/extracted/bd58f2e7808a2802cb11d9aae2673fa0a1e54b008404f75a1c63c2751332b806/cv-corpus-6.1-2020-12-11/es/clips/common_voice_es_19819446.mp3': File contains data in an unknown format.

Solution

Specify the format by adding:

--dtype="bfloat16" \

Issue

Warning "PySoundFile failed. Trying audioread instead." is spaming the terminal (librosa/librosa#1015).

Workaround

Ignore warnings with:

python -W ignore