The TV Speech and Music (TVSM) dataset contains speech and music activity labels across a variety of TV shows and their corresponding audio features extracted from professionally-produced high-quality audio. The dataset aims to facilitate research on speech and music detection tasks.

• The dataset can be downloaded via Zenodo.org.
• The paper can be downloaded via EURASIP open access.
• This repo contains materials and codebase to reproduce the baseline experiment in the paper.

@article{Hung2022,
  title={{A Large TV Dataset for Speech and Music Activity Detection},
  author={Hung, Yun-Ning and Wu, Chih-Wei and Orife, Iroro and Hipple, Aaron and Wolcott, William and Lerch, Alexander},
  journal={EURASIP Journal on Audio, Speech, and Music Processing},
  volume={2022},
  number={1},
  pages={21},
  year={2022},
  publisher={Springer}
}

The TVSM dataset is licensed under a Apache License 2.0 license

The downloaded dataset has the following structure:

└─── READEME.txt
└─── TVSM-cuesheet/
│    └─── labels/
│    └─── mel_features/
│    └─── mfcc/
│    └─── vgg_features/
│    └─── TVSM-xxxx_metadata.csv
└─── TVSM-pseudo/
└─── TVSM-test/

• READEME.txt: basic information about the dataset
• TVSM-cuesheet/: smaller subset used for training. The labels are derived from cuesheet information
• TVSM-pseudo/: larger subset used for training. The labels are labeled from a pre-trained model trained on TVSM-cuesheet
• TVSM-test/: subset for testing. The labels are labeled by human annotators

Each subset folder has the same structure:

• labels/: speech and music activation labels for each sample. Each row in a csv file represents "start time", "end time" and "s(speech)/m(music)"
• mel_features/: the Mel spectrogram feature extracted from the audio of each sample
• mfcc/: the MFCCs feature extracted from the audio of each sample
• vgg_features/: the VGGish feature extracted from the audio of each sample
• TVSM-xxxx_metadata.csv: the metadata of each sample

For more information, please visit our paper

Interested in inferencing existing samples? Please visit predictor.py for usage.

cd training_code
python3 predictor.py --audio_path test.wav

Please install git lfs first then run git-lfs pull to restore the checkpoints

Please replace line 31 in SM_detector.py with self.save_hyperparameters(hparams) if you are using newer pytorch_lightning versions.

└─── Evaluation_Output/
│    └─── AVASpeech/
│    │    └─── T2
│    │    └─── TVSM-cuesheet
│    │    └─── TVSM-pseudo
│    └─── ...
└─── Models/
└─── training_code/

• Evaluation_Output: the output generated by three models across five evaluation sets
  • T2: baseline method
  • TVSM-cuesheet: CRNN-P-Cue method
  • TVSM-pseudo: CRNN-P-Pseu method
• Models: the pre-trained checkpoint from CRNN-P-Cue and CRNN-P-Pseu methods
• training_code: code for training the model

If you encounter error "batch response: This repository is over its data quota. Account responsible for LFS...", can download the model checkpoint from Google Drive

Please feel free to contact [email protected] or open an issue here if you have any questions about the dataset or the support code.