anton-jeran/fast-rir

This is the official implementation of our neural-network-based fast diffuse room impulse response generator (FAST-RIR) for generating room impulse responses (RIRs) for a given acoustic environment.

FAST-RIR: FAST NEURAL DIFFUSE ROOM IMPULSE RESPONSE GENERATOR (ICASSP 2022)

This is the official implementation of our neural-network-based fast diffuse room impulse response generator (FAST-RIR) for generating room impulse responses (RIRs) for a given rectangular acoustic environment. Our model is inspired by StackGAN architecture. The audio examples and spectrograms of the generated RIRs are available here.

NEWS:

1) We have generalized our FAST-RIR to generate RIRs for any 3D indoor scenes represented using meshes. Official code of our network MESH2IR is available.

2) We release MULTI-CHANNEL MULTI-SPEAKER MULTI-SPATIAL AUDIO CODEC. The official code of our network M3-AUDIODEC is available.

Requirements

Python3.6
Pytorch
python-dateutil
easydict
pandas
torchfile
gdown
librosa
soundfile
acoustics
wavefile
wavfile
pyyaml==5.4.1
pickle

Embedding

Each normalized embedding is created as follows: If you are using our trained model, you may need to use extra parameter Correction(CRR).

Listener Position = LP
Source Position = SP
Room Dimension = RD
Reverberation Time = T60
Correction = CRR

CRR = 0.1 if 0.5<T60<0.6
CRR = 0.2 if T60>0.6
CRR = 0 otherwise

Embedding = ([LP_X,LP_Y,LP_Z,SP_X,SP_Y,SP_Z,RD_X,RD_Y,RD_Z,(T60+CRR)] /5) - 1

Generete RIRs using trained model

Download the trained model using this command

source download_generate.sh

Create normalized embeddings list in pickle format. You can run following command to generate an example embedding list

 python3 example1.py

Run the following command inside code_new to generate RIRs corresponding to the normalized embeddings list. You can find generated RIRs inside code_new/Generated_RIRs

python3 main.py --cfg cfg/RIR_eval.yml --gpu 0

Range

Our trained NN-DAS is capable of generating RIRs with the following range accurately.

Room Dimension X --> 8m to 11m
Room Dimesnion Y --> 6m to 8m
Room Dimension Z --> 2.5m to 3.5m
Listener Position --> Any position within the room
Speaker Position --> Any position within the room
Reverberation time --> 0.2s to 0.7s

Training the Model

Run the following command to download the training dataset we created using a Diffuse Acoustic Simulator. You also can train the model using your dataset.

source download_data.sh

Run the following command to train the model. You can pass what GPUs to be used for training as an input argument. In this example, I am using 2 GPUs.

python3 main.py --cfg cfg/RIR_s1.yml --gpu 0,1

Related Works

1. IR-GAN: Room Impulse Response Generator for Far-field Speech Recognition (INTERSPEECH2021)
2. TS-RIR: Translated synthetic room impulse responses for speech augmentation (IEEE ASRU 2021)

Citations

If you use our FAST-RIR for your research, please consider citing

@INPROCEEDINGS{9747846, 
author={Ratnarajah, Anton and Zhang, Shi-Xiong and Yu, Meng and Tang, Zhenyu and Manocha, Dinesh and Yu, Dong}, 
booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
title={Fast-Rir: Fast Neural Diffuse Room Impulse Response Generator},
year={2022}, 
volume={},
number={},
pages={571-575},
doi={10.1109/ICASSP43922.2022.9747846}}

Our work is inspired by

@inproceedings{han2017stackgan,
Author = {Han Zhang and Tao Xu and Hongsheng Li and Shaoting Zhang and Xiaogang Wang and Xiaolei Huang and Dimitris Metaxas},
Title = {StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks},
Year = {2017},
booktitle = {{ICCV}},
}

If you use our training dataset generated using Diffuse Acoustic Simulator in your research, please consider citing

@inproceedings{9052932,
  author={Z. {Tang} and L. {Chen} and B. {Wu} and D. {Yu} and D. {Manocha}},  
  booktitle={ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},  
  title={Improving Reverberant Speech Training Using Diffuse Acoustic Simulation},   
  year={2020},  
  volume={},  
  number={},  
  pages={6969-6973},
}

fast-rir's Issues

how to generate multi channel RIRs for micphone array?

Is there any way to generate multi channel RIRs for micphone array, I try to generated 2 sources and 8 micphones mono RIR and stack them together with order of name, but get bad result, some of them look like normal, the rest look like interleave between 2 sources.

by the way, can this rep generate large room, such as 12m*12m?

DAS use question

Hi, how can I use DAS to generate RIR conforming to the given RT60?

Can not hear the RIR voice

i have a question, in the web https://anton-jeran.github.io/FRIR/ i can hear the "RIR generated using DAS", but i use your model to generate theRIR-* .wav, i could not hear the voice, how can i check the generated wav is right.