A kaldi recipe for automatic lyrics transcription and audio-to-lyrics alignment tasks.

If you use this repository, please cite it as follows:

@inproceedings{demirel2020,
  title={Automatic lyrics transcription using dilated convolutional neural networks with self-attention},
  author={Demirel, Emir and Ahlback, Sven and Dixon, Simon},
  booktitle={International Joint Conference on Neural Networks},
  publisher={IEEE},
  year={2020}
}

Link to paper : https://arxiv.org/abs/2007.06486

This framework is built as a Kaldi[1] recipe For instructions on Kaldi installation, please visit https://github.com/kaldi-asr/kaldi

pip install -r requirements.txt

• Modify KALDI_ROOT in s5/path.sh according to where your Kaldi installation is.

• Retrieve Data:

The s5 recipe is based on the DSing!300x30x2 dataset within Smule's DAMP[2] repository. To retrieve the DSing!300x30x2, you need to apply for authorization from https://ccrma.stanford.edu/damp/.

• Set the path to DAMP - Sing!300x30x2 data:

cd s5
damp_data='path-to-your-damp-directory'

We have provided the data files (at data/{train,dev,test}) required in Kaldi pipelines for the ease of using this repository.

• Execute the pipeline:

./run_damp.sh $damp_data

• If you have any problems during the pipeline, look up for the relevant process in run.sh

NOTE: If you use dev and test sets in your experiments, please cite [3]

Run the script for extracting the phoneme posteriorgrams as follows:

audio_path='absolute-path-to-the-input-audio-file'
save_path='path-to-save-the-output
cd s5
./extract_phn_posteriorgram.sh $audio_path $save_path

The output posteriorgrams are saved as numpy arrays (.npy).

Note that we have used 16kHz for the sample rate and 10ms of hop size.

Automatic Lyrics Transcription is the task of translating singing voice into text. Jusy like in hybrid speech recognition, our lyrics transcriber consists of separate acoustic, language and pronunciation models.

Acoustic Model: Sequence discriminative training on MMI criteria[4].

The neural network architecture consists of 2D Convolutions, factorized time-delay and self-attention layers:

Language Model: We use the lyrics of recent (2015-2018) popular songs for training the LM (s5/conf/corpus.txt).

Pronunciation Model: The standard CMU-Sphinx English pronunciation dictionary (http://www.speech.cs.cmu.edu/cgi-bin/cmudict).

(Work in progress : The singing-adapted pronunciation dictionary will be provided soon, as well as grapheme based lexicons for modeling unseen words.)

[1] Povey, Daniel, et al. "The Kaldi speech recognition toolkit." IEEE 2011 workshop on automatic speech recognition and understanding. No. CONF. IEEE Signal Processing Society, 2011.

[2] Digital Archive of Mobile Performances (DAMP) portal hosted by the Stanford Center for Computer Research in Music and Acoustics (CCRMA) (https://ccrma.stanford.edu/damp/)

[3] Dabike, Gerardo Roa, and Jon Barker. "Automatic Lyric Transcription from Karaoke Vocal Tracks: Resources and a Baseline System." INTERSPEECH. 2019.

[4] Povey, Daniel, et al. "Purely sequence-trained neural networks for ASR based on lattice-free MMI." Interspeech. 2016.

This work is licensed under Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International, which means that the reusers can copy, distribute, remix, transform and build upon the material in any media providing the appropriate credits to this repository and to be used for non-commercial purposes.