badrex/Ethio-ASR

Automatic Speech Recognition for Languages of Ethiopia # Ethio ASR To run the code, follow the instructions below ### 1. Clone the repo ```shell clone github.com ``` ### 2. Set up environment variables in .env Then create an empty `.env` file in the project root: ```env # Weights & Biases API key WANDB_API_KEY="wandb_api_key_xxx" # Hugging Face API key HF_API_KEY="hf_api_key_xxx" # Set cache NUMBA_CACHE_DIR='/tmp/numba_cache' LIBROSA_CACHE_DIR="/tmp/librosa_cache" # Hugging Face cache directory HF_HOME='/project_dir/huggingface_cache' # MPLCONFIGDIR MPLCONFIGDIR='/tmp/matplotlib_cache' ``` ### 3. Set the YAML config file to run the experiment This config is under config_files. This config below is for an experiment on a small scale dataset. ```yaml # Project settings # this is the WANDB project name project: "Ethio-ASR" # this is the output directory for saving the model and processor output_dir: "inprogress/Ethio-ASR" # set random seed for reproducibility seed: 42 # Model settings pretrained_model: "facebook/w2v-bert-2.0" # or "acebook/mms-300m freeze_feature_encoder: true add_final_layer_adapter: true # should be false for facebook/mms-300m # Training settings batch_size: 8 gradient_accumulation_steps: 4 num_epochs: 25 max_steps: 18400 learning_rate: 0.00003 # or 0.0005 for "facebook/mms-300m" warmup_ratio: 0.1 fp16: true gradient_checkpointing: true save_steps: 800 eval_steps: 800 logging_steps: 5 save_total_limit: 2 # Data settings # if use_custom_dataset is true, then dataset_path is the path to the custom dataset on disk # if use_custom_dataset is false, then dataset_path is the dataset repo name on the HF hub use_custom_dataset: false # if from HF hub, use the repo name for example "badrex/waxalNLP-ethiopic-final" dataset_path: "badrex/waxalNLP-ethiopic-final" train_split: "train" eval_split: "validation" language: "all" # Data sampling settings (for debugging purposes) sample: true sample_size: 197634 # this is the size of the dataset in the HF hub ## Text prep …