Frame-level emotional state alignment method for speech emotion recognition

Official implemention for the paper "Frame-level emotion state alignment method for speech emotion recognition", which is submitted to ICASSP 2024.

I am very sorry, I have carefully checked the data division method. The method we used is leave-one-session-out https://github.com/ECNU-Cross-Innovation-Lab/ShiftSER. For each fold, four sessions are used for training while one session is used for testing. The final result is the average of 5 folds. I made a mistake in my paper, and I appreciate everyone's correction. I am extremely sorry for the trouble caused to the readers.

Libraries and dependencies

1. pytorch
2. PyTorch-lighting
3. accelerater
4. sklearn
5. transformers
6. soudfile

Framework of method

Results on IEMOCAP

The format of training, validation, and test sources

You need three scp files containing training, validation and test sources, which are divided according to the methods described in the paper . These scp files are shown as following format (wav_name, wav_path, wav_label):

Phase 1. TPAT and Cluster

TAPT

Task adaptive pretraining stage

First, we need to dump feature and generate pesudo label for pretraining by a pretrained k-means model which is provided by official HuBERT https://github.com/facebookresearch/fairseq/tree/main/examples/hubert. Second, you need copy 5531 samples in IEMOCAP into a folder, for example "./audio". You can find the "hubert-base-ls960" at here https://huggingface.co/facebook/hubert-base-ls960.

#dump feature and generate pesudo label
python preprocess.py --root_dir ./audio 
                     --feat_type hubert 
                     --exp_dir ./exp 
                     --layer_index 9 
                     --num_rank 40 
                     --hubert_base ./hubert-base-ls960 
                     --checkpoint_path ./hubert-base-ls960 
                     --iskmean False
                     --pretrain_kmeans hubert_base_ls960_L9_km500.bin 
                     --percent -1

#pretraining hubert
python train.py --dataset_path ./exp/data/hubert_9/ 
                --hubert_path ./hubert-base-ls960 
                --exp_dir ./exp_iter2_9_500 
                --feature_type hubert 
                --num_class 500 
                --max_updates 20000 
                --learning_rate 0.0005 
                --gpus 4 
                --num_nodes 1

Task adaptive fine-tuning stage

#transfer the format of model into the format of huggingface.
python model_transfer.py --pretrained_model_path ./exp_iter2_9_500/epoch=179-step=17250.ckpt               											    --hubert_base_ls960 ./hubert-base-ls960 
                         --save_path ./transfer_model_9_500

#fine-tune hubert
accelerate fine_tune_hubert.py --CPT_HuBERT_path ./transfer_model_9_500 
                               --train_src_path ./session01_train.scp 
                               --valid_src_path ./session01_valid.scp 
                               --test_src_path ./session01_test.scp 
                               --save_root ./session01_fine_tune_model

Cluster

From here on down, it's all about session 1. You need to repeat the following steps on other sessions.

python preprocess.py --root_dir ./audio 
                     --feat_type hubert 
                     --exp_dir ./exp 
                     --layer_index 9 
                     --num_rank 40 
                     --hubert_base ./hubert-base-ls960 
                     --checkpoint_path ./session01_fine_tune_model 
                     --num_cluster 50 
                     --percent -1

Phase 2. Pretraining HuBERT

python train.py --dataset_path ./exp/data/hubert_9/ 
                --hubert_path ./hubert-base-ls960 
                --exp_dir ./exp_iter2_9_50_session01 
                --feature_type hubert 
                --num_class 50 
                --max_updates 20000 
                --learning_rate 0.0005 
                --gpus 4 
                --num_nodes 1

python model_transfer.py --pretrained_model_path ./exp_iter2_9_50_session01/epoch=168-step=16196.ckpt 
                         --hubert_base_ls960 ./hubert-base-ls960 
                         --save_path ./transfer_model_9_50_session01_CPT_HuBERT

Phase 3. Fine-tuning CPT-HuBERT for SER

accelerate fine_tune_hubert.py --CPT_HuBERT_path ./transfer_model_9_50_session01_CPT_HuBERT           															 
                               --train_src_path ./session01_train.scp 
                               --valid_src_path ./session01_valid.scp 
                               --test_src_path ./session01_test.scp 
                               --save_root 	./session01_final_model_SER