PerceptiveAgent

Talk With Human-like Agents: Empathetic Dialogue Through Perceptible Acoustic Reception and Reaction (ACL24)

[Arxiv] [Watch Case Study]

To avoid overlooking human communication nuances and misinterpreting speakers' intentions, we propose PerceptiveAgent, an empathetic multi-modal dialogue system designed to discern deeper or more subtle meanings beyond the literal interpretations of words through the integration of speech modality perception. Employing LLMs as a cognitive core, PerceptiveAgent perceives acoustic information from input speech and generates empathetic responses based on speaking styles described in natural language.

Models

• Speech Captioner: download it from GoogleDrive
• Pretrained Synthesizer: download it from GoogleDrive
• MSMA Synthesizer (Finetuned Synthesizer): download it from GoogleDrive

Getting Started

Clone this repository.

git clone https://github.com/Haoqiu-Yan/PerceptiveAgent.git
cd PerceptiveAgent

Configure environment.

Limited by package compatibility, we create two virtual environments. We recommend running on linux using a conda environment. We employ Python 3.8 and Torch 1.13.1 (cuda) in both environments.

1. Environment of Speech Captioner & Chatting with LLM

conda create -n cap38 python=3.8
conda activate cap38

pip install -r cap_requirement.txt

2. Environment of MSMA Synthesizer

conda create -n syn38 python=3.8
conda activate syn38

pip install -r syn_requirement.txt

Speech Captioner

Data Preprocessing

1. For the TextrolSpeeh dataset which provides caption.csv, the following command reads captions one by one and saves them as audio_name.json files to the output directory $MIXDIR. Besides, audio files would be copied to $MIXDIR as well.

   DATA_ROOT=/path/datasets/textrolspeech
   ALLCSV=${DATA_ROOT}/caption/random_train.csv
   MIXDIR=/path/to/save
   python ./captioner/dataset/process.py --dataset textrol --data_dir ${DATA_ROOT} \
       --json_path $ALLCSV \
       --saveto $MIXDIR

2. For the other datasets without caption.csv provided, you can create an empty caption file before processing. Take the EXPRESSO dataset as an example:

   python data_preprocess/create_caption_expresso.py \
       -audio-root /path/expresso/merge_audio_48khz/ \
       --transcript-path /path/expresso/read_transcriptions.txt \
       --saveas /path/expresso/caption/random_read_all.csv

   Then, change the argument --dataset to the responding dataset.

   DATA_ROOT=/path/expresso
   ALLCSV=/path/expresso/caption/random_read_all.csv
   MIXDIR=/path/to/save
   python ./captioner/dataset/process.py --dataset expresso --data_dir ${DATA_ROOT} \
       --json_path $ALLCSV \
       --saveto $MIXDIR

Training

To update llama_model in configs/capsp_train_gpt2.yaml, you can download the pretrained Vicuna weights, according to the instruction in BuboGPT.

To update ABSOLUTE_PATH_OF_bubogpt_7b in configs/capsp_infer.yaml, you can download the pretrained bubogpt_7b from the link.

Then, run the following command to train a speech captioner.

bash scripts/captioner_train.sh ${CUDA_ID} ${CUDA_NUM}

Inference

Infer by the following shell,

bash scripts/captioner_infer.sh ${CUDA_ID} ${CUDA_NUM}

Chat with LLM

To integrate captions into dialogue history, you can chat with ChatGPT by the following command.

python ./chat_llm/conversation_chat.py \
    --id-convs ./input_egs/id_convs_eg.txt \
    --audio-asr-caption ./input_egs/audio_asr_caption_eg.txt \
    --save-root /path/to/savedir > ./logs/MMDD-2200pm.log

Besides, to send dialogue history only to chatGPT, you can chat by:

python ./chat_llm/conversation_chat_text-only.py \
    --id-convs ./input_egs//id_convs_eg.txt \
    --audio-asr-caption ./input_egs/audio_asr_caption_eg.txt \
    --save-root /path/to/savedir > ./logs/MMDD-2200pm.log

MSMA Synthesizer

Data Preprocessing

1. The training phrase takes speech as input. To encode speech into discrete acoustic units, you can run the following command. The HuBERT model, specified by the argument DENSE_NAME in the shell expresso_hubert_gen.sh, would be downloaded automatically.

   bash ./scripts/expresso_hubert_gen.sh

2. The inference phrase takes text as input. To transform text into discrete acoustic units, a text-to-unit (T2U) model is trained by us, which can be downloaded from GoogleDrive.

   The spm model trained by us can be downloaded from GoogleDrive; Otherwise, you can train your own model by:

   spm_train --input=$ALL_TEXT --model_prefix=spm_bpe_1k --vocab_size=1000 --character_coverage=1.0 --model_type=bpe

   Then, run the following command to transform text into units.

   bash ./scripts/t2u_infer.sh

Training

1. Use the EXPRESSO, LJSpeech and VCTK dataset to pretrain a vocoder conditioned on emotion and speaker labels. Note: batch_size in $CONFIG_FILE is the product of batch size and CUDA number.

   CONFIG_FILE=./configs/synthesizer_pretrain_config.json
   OUTPUT_DIR=/path/to/save
   
   python -m torch.distributed.launch --nproc_per_node $GPUS --master_port=29502 \
           synthesizer/examples/pretrain/amp_train.py \
           --checkpoint_path $OUTPUT_DIR \
           --config $CONFIG_FILE \
           --training_epochs 2000 \
           --validation_interval 5000 \
           --checkpoint_interval 25000

2. Use the EXPRESSO dataset to finetune the above vocoder conditioned on pitch, energy and speed labels additionally. To finetune from the lastest checkpoint, you can add --from-latest-ckpt in the following command.

   CONFIG_FILE=./configs/synthesizer_finetune_config.json
   OUTPUT_DIR=/path/to/save
   
   python -m torch.distributed.launch --nproc_per_node $GPUS \
           synthesizer/examples/mcond_expresso/amp_train.py \
           --checkpoint_path $OUTPUT_DIR \
           --config $CONFIG_FILE \
           --training_epochs 2000 \
           --validation_interval 5000 \
           --checkpoint_interval 25000 \
           # --from-latest-ckpt

Inference

1. Infer with the pretrained model.

   CUDA_ID=$1
   GPUS=$2
   
   INPUT_CODE_FILE=./input_egs/syntheizer_pretrain_val.txt
   ckpt=g_00400000
   CHECKPOINT_FILE=/path/of/ckptdir/${ckpt}
   OUTPUT_DIR=/path/to/savedir
   mkdir $OUTPUT_DIR
   
   CUDA_VISIBLE_DEVICES=$CUDA_ID python ./synthesizer/examples/pretrain/inference_example.py \
       --input_code_file $INPUT_CODE_FILE \
       --checkpoint_file $CHECKPOINT_FILE \
       --output_dir $OUTPUT_DIR \
       --num-gpu $GPUS

2. Infer with the finetuned model.

   CUDA_ID=$1
   GPUS=$2
   
   INPUT_CODE_FILE=./input_egs/syntheizer_finetune_dev.txt
   ckpt=g_00200000
   CHECKPOINT_FILE=/path/of/ckptdir/${ckpt}
   OUTPUT_DIR=/path/to/savedir
   mkdir $OUTPUT_DIR
   
   CUDA_VISIBLE_DEVICES=$CUDA_ID python ./synthesizer/examples/mcond_expresso/inference_example.py \
       --input_code_file $INPUT_CODE_FILE \
       --checkpoint_file $CHECKPOINT_FILE \
       --output_dir $OUTPUT_DIR \
       --num-gpu $GPUS \
       --dur-prediction 

Author

Haoqiu Yan*, Yongxin Zhu*, Kai Zheng, Bing Liu, Haoyu Cao, Deqiang Jiang and Linli Xu†. (*Equal Contribution, †Corresponding Author)

How to cite

If you use the code or models from this project in your research, please cite our work as follows:

@article{yan2024talk,
  title={Talk With Human-like Agents: Empathetic Dialogue Through Perceptible Acoustic Reception and Reaction},
  author={Yan, Haoqiu and Zhu, Yongxin and Zheng, Kai and Liu, Bing and Cao, Haoyu and Jiang, Deqiang and Xu, Linli},
  journal={arXiv preprint arXiv:2406.12707},
  year={2024}
}

License

PerceptiveAgent is distributed under the Apache License.

Acknowledgment

This repo is developed based on the following repos:

• bubogpt
• speech-synthesis
• fairseq
• textlesslib
• sentencepiece