README.md

EmoBox

This repository holds code, processed meta-data, and benchmark for
EmoBox: Multilingual Multi-corpus Speech Emotion Recognition Toolkit and Benchmark

Guides

EmoBox is an out-of-the-box multilingual multi-corpus speech emotion recognition toolkit, along with a benchmark for both intra-corpus and cross-corpus settings on mainstream pre-trained foundation models. We hope that our toolkit and benchmark can facilitate the research of SER in the community.

Datasets

We include 32 speech emotion datasets spanning 14 distinct languages with download links, some of them require license or registration. We provide data preparation and partitioning of each datasets. Refer to the paper for more details.

Dataset	Source	Lang	Emo	Spk	#Utts	#Hrs
AESDD	Act	Greek	5	5	604	0.7
ASED	Act	Amharic	5	65	2474	2.1
ASVP-ESD	Media	Mix	12	131	13964	18.0
CaFE	Act	French	7	12	936	1.2
CASIA	Act	Mandarin	6	4	1200	0.6
CREMA-D	Act	English	6	91	7442	5.3
EMNS	Act	English	8	1	1181	1.9
EmoDB	Act	German	7	10	535	0.4
EmoV-DB	Act	English	5	4	6887	9.5
EMOVO	Act	Italian	7	6	588	0.5
Emozionalmente	Act	Italian	7	431	6902	6.3
eNTERFACE	Act	English	6	44	1263	1.1
ESD	Act	Mix	5	20	35000	29.1
IEMOCAP	Act	English	5	10	5531	7.0
JL-Corpus	Act	English	5	4	2400	1.4
M3ED	TV	Mandarin	7	626	24437	9.8
MEAD	Act	English	8	48	31729	37.3
MELD	TV	English	7	304	13706	12.1
MER2023	TV	Mandarin	6	/	5030	5.9
MESD	Act	Spanish	6	11	862	0.2
MSP-Podcast	Podcast	English	8	1273	73042	113.6
Oreau	Act	French	7	32	434	0.3
PAVOQUE	Act	German	5	1	7334	12.2
Polish	Act	Polish	3	5	450	0.1
RAVDESS	Act	English	8	24	1440	1.5
RESD	Act	Russian	7	200	1396	2.3
SAVEE	Act	English	7	4	480	0.5
ShEMO	Act	Persian	6	87	2838	3.3
SUBESCO	Act	Bangla	7	20	7000	7.8
TESS	Act	English	7	2	2800	1.6
TurEV-DB	Act	Turkish	4	6	1735	0.5
URDU	Talk show	Urdu	4	29	400	0.3
Total	--	--	--	3510	262020	294.4

Benchmark

Intra-corpus Benchmark

Intra-corpus SER results of 10 pre-trained speech models on 32 emotion datasets spanning 14 distinct languages with EmoBox data partitioning. Refer to the intra-corpus benchmark and the paper for more details.

Cross-corpus Benchmark

Cross-corpus SER results of 10 pre-trained speech models on 4 EmoBox fully balanced test sets. Refer to the cross-corpus benchmark and the paper for more details.

Play with EmoBox

Prepare datasets

You need to download datasets and put them into downloads folder. Make sure paths to your downloaded audio files follow the audio paths in jsonl files in data/.

For example, you need to download the iemocap dataset into relative paths such as downloads/iemocap/Session1/sentences/wav/Ses01F_impro04/Ses01F_impro04_F000.wav, these audio files should follow the audio paths in data/iemocap/iemocap.jsonl

Metadata

We prepare metadata files for each datasets, including several types of formats: json, jsonl, .... For example, the format of metadata in jsonl files is:

[  {
	"key": "Ses01M_impro01_F000",
	"dataset": "iemocap",
	"wav": "downloads/iemocap/Session1/sentences/wav/Ses01M_impro01/Ses01M_impro01_F000.wav",
   "type": "raw" # raw, feature
	"sample_rate": 16000,
	"length": 3.2,
	"task": "category", # category, valence, arousal
	"emo": "hap",
	"channel": 1
	}
	,
	..., 
	{...}
]	

Some datasets (e.g. iemocap) require merging labels, so we provide a label_map.json file for this purpose.

Quick Start

EmoBox provides a torch dataset class EmoDataset and a evaluation class EmoEval. You may train your own models using any recipes or toolkits.

Using EmoDataset and EmoEval, it is easy to compare results from any model trained by any recipie or toolkit. Results can be submitted to our benchmark.

We provide an example pipeline code using EmoDataset and EmoEval:

from EmoBox import EmoDataset, EmoEval

dataset = "iemocap"
folds = 5 # different datasets have different number of folds, which can be find in data/ 
user_data_dir = "./" # path to EmoBox
meta_data_dir = "data/" # path to data folder 
label2idx = {'hap':0, 'sad':1, 'ang':2, 'neu':3} # you may need to define a label to index mapping for your own training, see `data/iemocap/label_map.json`


## take n-flod cross-validation as an example
for fold in range(folds):
	
	train = EmoDataset(dataset, user_data_dir, meta_data_dir, fold=fold, split="train")
	val = EmoDataset(dataset, user_data_dir, meta_data_dir, fold=fold, split="valid")
	
	"""
		Train your model
	"""
	for data in train:
		audio = data["audio"] # raw wav tensor
		label = data["label"] # label, e.g. 'hap'
	
		
	"""
		Evaluate on test data
	"""	
	test = EmoDataset(dataset, user_data_dir, meta_data_dir, fold=fold, split="test")
	test_pred = []
	test_targets = []
	for data in test:
		key = data["key"]
		audio = data["audio"]
		label = data["label"]
		label_index = label2idx[label]
		pred = model(audio) # prediction for label index
		test_pred.append({"key":key, "pred":pred})
		test_targets.append({"key:key", "label":label_index})
	
	emo_eval = EmoEval(test_pred, test_targets)
	scores = emo_eval.compute_metrics()
	"""
	scores is a dict follows the format: 
	{
		"wa":xx, # scikit-learn accuracy
		"ua":xx, # scikit-learn balanced_accuracy
		"micro_f1":xx, # scikit-learn f1_score
		"macro_f1":xx, # scikit-learn f1_score
		"confusion":xx, # scikit-learn confusion matrix
		"num": xx, #num of samples
		"report": xxx # scikit-learn class-wise report 
	}
	"""
	

Example Recipe

We provide a recipe with speechbrain framework:

# setup the environment
cd examples/sb
conda create -n conda-env -f environment.yml
conda activate conda-env

# extract features
python3 speech_feature_extraction.py \
	--model_name wav2vec2-base \
	--model_path pretrained_models/wav2vec2-base \
	--dump_dir dump \
	--device cuda \
	--data data/iemocap/iemocap.json \
	--output_norm

# run training & evaluation
dataset=iemocap
fold=1
n_classes=$(python -c "import json; label_map = json.load(open('${data_dir}/${dataset}/label_map.json')); labels=set([label_map[key] for key in label_map]); print(len(labels)) ")
python3 train.py \
	hparams/wav2vec2-base_freeze.yaml \
	--output_folder exp/iemocap_wav2vec2-base \
    --seed 1234 \
    --batch_size 32 \
    --lr 1e-4 \
    --train_annotation data/${dataset}/fold_${fold}/${dataset}_train_fold_${fold}.json \
    --valid_annotation data/${dataset}/fold_${fold}/${dataset}_valid_fold_${fold}.json \
    --test_annotation data/${dataset}/fold_${fold}/${dataset}_test_fold_${fold}.json  \
    --number_of_epochs 100 \
    --feat_dir dump/${dataset}/wav2vec2-base \
    --label_map data/${dataset}/label_map.json \
    --device cuda \
    --out_n_neurons ${n_classes} \
    --hidden_size 128 \

Submit to benchmark

You may trained your own models following the evaluation setup in our benchmarks. You can send test set results to us through email.

Reference

Please cite our paper below if you use the EmoBox toolkit and benchmark.

@inproceedings{ma2024emobox,
  title={EmoBox: Multilingual Multi-corpus Speech Emotion Recognition Toolkit and Benchmark},
  author={Ziyang Ma and Mingjie Chen and Hezhao Zhang and Zhisheng Zheng and Wenxi Chen and Xiquan Li and Jiaxin Ye and Xie Chen and Thomas Hain},
  booktitle={Proc. INTERSPEECH},
  year={2024}
}