preprocess_scripts/1.learn_KM_clustering_model.sh

lang=$1
DATA_ROOT=/data/zhangshaolei/datasets/cvss/cvss-c/$lang-en
ROOT=/data/zhangshaolei/StreamSpeech
PRETRAIN_ROOT=$ROOT/pretrain_models

N_CLUSTERS=100 #<number_of_clusters_used_for_kmeans>
TYPE=hubert #<one_of_logmel/cpc/hubert/w2v2>
CKPT_PATH=$PRETRAIN_ROOT/mHuBERT/mhubert_base_vp_en_es_fr_it3.pt #<path_of_pretrained_acoustic_model>
LAYER=11 #<layer_of_acoustic_model_to_extract_features_from>
MANIFEST=$DATA_ROOT #<tab_separated_manifest_of_audio_files_for_training_kmeans>
KM_MODEL_PATH=$ROOT/preprocess_scripts/mhubert.km1000.layer11.pt #<output_path_of_the_kmeans_model>



PYTHONPATH=$ROOT/fairseq python $ROOT/preprocess_scripts/create_manifest.py --data-root $DATA_ROOT

for split in train dev test
do
    PYTHONPATH=$ROOT/fairseq python $ROOT/preprocess_scripts/quantize_with_kmeans.py \
        --feature_type $TYPE \
        --kmeans_model_path $KM_MODEL_PATH \
        --acoustic_model_path $CKPT_PATH \
        --layer $LAYER \
        --manifest_path $DATA_ROOT/$split.txt \
        --out_quantized_file_path $DATA_ROOT/$split.km1000
done