sithu31296/audio-tagging

Easy to use Audio Tagging in PyTorch

Audio Classification, Tagging & Sound Event Detection in PyTorch

Progress:

• Fine-tune on audio classification
• Fine-tune on audio tagging
• Fine-tune on sound event detection
• Add tagging metrics
• Add Tutorial
• Add Augmentation Notebook
• Add more schedulers
• Add FSDKaggle2019 dataset
• Add MTT dataset
• Add DESED
• Test in real-time

Model Zoo

AudioSet Pretrained Models

Model	Task	mAP (%)	Sample Rate (kHz)	Window Length	Num Mels	Fmax	Weights
CNN14	Tagging	43.1	32	1024	64	14k	download
CNN14_16k	Tagging	43.8	16	512	64	8k	download
CNN14_DecisionLevelMax	SED	38.5	32	1024	64	14k	download

Note: These models will be used as a pretrained model in the fine-tuning tasks below. Check out audioset-tagging-cnn, if you want to train on AudioSet dataset.

Fine-tuned Classification Models

Model	Dataset	Accuracy (%)	Sample Rate (kHz)	Weights
CNN14	ESC50 (Fold-5)	95.75	32	download
CNN14	FSDKaggle2018 (test)	93.56	32	download
CNN14	SpeechCommandsv1 (val/test)	96.60/96.77	32	download

Fine-tuned Tagging Models

Model	Dataset	mAP(%)	AUC	d-prime	Sample Rate (kHz)	Config	Weights
CNN14	FSDKaggle2019	-	-	-	32	-	-

Fine-tuned SED Models

Model	Dataset	F1	Sample Rate (kHz)	Config	Weights
CNN14_DecisionLevelMax	DESED	-	32	-	-

---

Supported Datasets

Dataset	Task	Classes	Train	Val	Test	Audio Length	Audio Spec	Size
ESC-50	Classification	50	2,000	5 folds	-	5s	44.1kHz, mono	600MB
UrbanSound8k	Classification	10	8,732	10 folds	-	<=4s	Vary	5.6GB
FSDKaggle2018	Classification	41	9,473	-	1,600	300ms~30s	44.1kHz, mono	4.6GB
SpeechCommandsv1	Classification	30	51,088	6,798	6,835	<=1s	16kHz, mono	1.4GB
SpeechCommandsv2	Classification	35	84,843	9,981	11,005	<=1s	16kHz, mono	2.3GB
FSDKaggle2019*	Tagging	80	4,970+19,815	-	4,481	300ms~30s	44.1kHz, mono	24GB
MTT*	Tagging	50	19,000	-	-	-	-	3GB
DESED*	SED	10	-	-	-	10	-	-

Notes: * datasets are not available yet. Classification dataset are treated as multi-class/single-label classification and tagging and sed datasets are treated as multi-label classification.

Dataset Structure (click to expand)

Download the dataset and prepare it into the following structure.

datasets
|__ ESC50
    |__ audio

|__ Urbansound8k
    |__ audio

|__ FSDKaggle2018
    |__ audio_train
    |__ audio_test
    |__ FSDKaggle2018.meta
        |__ train_post_competition.csv
        |__ test_post_competition_scoring_clips.csv

|__ SpeechCommandsv1/v2
    |__ bed
    |__ bird
    |__ ...
    |__ testing_list.txt
    |__ validation_list.txt

Augmentations (click to expand)

Currently, the following augmentations are supported. More will be added in the future. You can test the effects of augmentations with this notebook

WaveForm Augmentations:

• MixUp
• Background Noise
• Gaussian Noise
• Fade In/Out
• Volume
• CutMix

Spectrogram Augmentations:

• Time Masking
• Frequency Masking
• Filter Augmentation

---

Usage

Requirements (click to expand)

• python >= 3.6
• torch >= 1.8.1
• torchaudio >= 0.8.1

Other requirements can be installed with pip install -r requirements.txt.

Configuration (click to expand)

• Create a configuration file in configs. Sample configuration for ESC50 dataset can be found here.
• Copy the contents of this and then edit the fields you think if it is needed.
• This configuration file is needed for all of training, evaluation and prediction scripts.

Training (click to expand)

To train with a single GPU:

$ python tools/train.py --cfg configs/CONFIG_FILE_NAME.yaml

To train with multiple gpus, set DDP field in config file to true and run as follows:

$ python -m torch.distributed.launch --nproc_per_node=2 --use_env tools/train.py --cfg configs/CONFIG_FILE_NAME.yaml

Evaluation (click to expand)

Make sure to set MODEL_PATH of the configuration file to your trained model directory.

$ python tools/val.py --cfg configs/CONFIG_FILE.yaml

Audio Classification/Tagging Inference

• Set MODEL_PATH of the configuration file to your model's trained weights.
• Change the dataset name in DATASET >> NAME as your trained model's dataset.
• Set the testing audio file path in TEST >> FILE.
• Run the following command.

$ python tools/infer.py --cfg configs/CONFIG_FILE.yaml

## for example
$ python tools/infer.py --cfg configs/audioset.yaml

You will get an output similar to this:

Class                     Confidence
----------------------  ------------
Speech                     0.897762
Telephone bell ringing     0.752206
Telephone                  0.219329
Inside, small room         0.20761
Music                      0.0770325

Sound Event Detection Inference

• Set MODEL_PATH of the configuration file to your model's trained weights.
• Change the dataset name in DATASET >> NAME as your trained model's dataset.
• Set the testing audio file path in TEST >> FILE.
• Run the following command.

$ python tools/sed_infer.py --cfg configs/CONFIG_FILE.yaml

## for example
$ python tools/sed_infer.py --cfg configs/audioset_sed.yaml

You will get an output similar to this:

Class                     Start    End
----------------------  -------  -----
Speech                      2.2    7
Telephone bell ringing      0      2.5

The following plot will also be shown, if you set PLOT to true:

References (click to expand)

• https://github.com/qiuqiangkong/audioset_tagging_cnn
• https://github.com/YuanGongND/ast
• https://github.com/frednam93/FilterAugSED
• https://github.com/lRomul/argus-freesound

Citations (click to expand)

@misc{kong2020panns,
      title={PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition}, 
      author={Qiuqiang Kong and Yin Cao and Turab Iqbal and Yuxuan Wang and Wenwu Wang and Mark D. Plumbley},
      year={2020},
      eprint={1912.10211},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
}

@misc{gong2021ast,
      title={AST: Audio Spectrogram Transformer}, 
      author={Yuan Gong and Yu-An Chung and James Glass},
      year={2021},
      eprint={2104.01778},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
}

@misc{nam2021heavily,
      title={Heavily Augmented Sound Event Detection utilizing Weak Predictions}, 
      author={Hyeonuk Nam and Byeong-Yun Ko and Gyeong-Tae Lee and Seong-Hu Kim and Won-Ho Jung and Sang-Min Choi and Yong-Hwa Park},
      year={2021},
      eprint={2107.03649},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}

audio-tagging's Issues

Do you have 8K Sample Rate support

Do you have 8K Sample Rate support