Skip to content

Latest commit

 

History

History
126 lines (85 loc) · 7.1 KB

README.md

File metadata and controls

126 lines (85 loc) · 7.1 KB
Raw

key_visual

Masked Modeling Duo for Speech (M2D-S)

This repository provides a demo implementation of "Masked Modeling Duo for Speech: Specializing General-Purpose Audio Representation to Speech using Denoising Distillation."

  • Code for pre-training and pre-processing LS-960 features.
  • Pre-trained weights.
  • SUPERB evaluation code and instructions.

1. Getting Started

For installation, follow the instruction in the "1. Getting Started" in the main README.md.

For evaluating on SUPREB, refer to superb/upstream/m2d/README.md.

2. Pre-trained weights

Find pre-trained weight files in releases.

  • M2D-S T=4.0s: m2d_s_vit_base-80x400p80x2-230201
  • M2D-S T=5.12s: m2d_s_vit_base-80x512p80x2-230301
  • M2D-S T=6.08s: m2d_s_vit_base-80x608p80x2-230220
Model Pre-trained dataset PR KS IC SID ER ENV MUS
M2D-S T=4.0s LS-960+AS 5.72 96.47 97.80 81.97 66.36 53.22 41.71
M2D-S T=5.12s LS-960+AS 5.64 96.87 97.65 80.69 65.35 57.34 43.23
M2D-S T=6.08s LS-960+AS 5.33 96.80 97.63 81.74 66.13 54.77 43.75

3. Pre-training from Scratch

3-1. Pre-processing data files

M2D-S learns from the following pre-processed files using LibriSpeech (LS-960) and HuBERT-base pre-trained model.

  • data/ls960_hybrid7s_hubaseL9: Pre-processed data consists of log-mel spectrogram samples converted from LS-960 and HuBERT layer #9 features encoded from LS-960.
  • data/files_ls960_hybrid.csv: List of pre-processed files of the ls960_hybrid7s_hubaseL9 folder.

M2D-S also requires AudioSet as a background noise.

  • data/audioset_lms: Pre-processed log-mel spectrogram samples from AudioSet, as in the original M2D.
  • data/files_audioset.csv: List of pre-processed AudioSet files, as in the original M2D.

3-1-1. LS-960 data files

The following command line will create data/ls960_hybrid7s_hubaseL9 and data/files_ls960_hybrid.csv.

python speech/extract_offline_ls960.py /path/to/LibriSpeech

3-1-2. AudioSet data files

For preparing AudioSet data files (data/audioset_lms and data/files_audioset.csv), please follow the data/README.md.

3-2. Pre-training

The train_speech.py pre-trains for speech.

The following example would run on any affordable GPU, consuming only 7,170MiB. However, please note that it will take very long (It took over 20 minutes for one epoch). You can also change the BG noise dataset by adding --csv_bg_noise data/files_f_s_d_5_0_k.csv, for example.

python -m speech.train_speech --loss_m2d 1. --loss_off 1. --input_size 80x208 --patch_size 80x4 --noise_ratio 0.2 --batch_size 128 --accum_iter 16

The followings are for pre-training high-end models, taking 2.5-3.5 days to complete with 4 A100s.

OMP_NUM_THREADS=1 torchrun --nproc_per_node=4 -m speech.train_speech --loss_m2d 1. --loss_off .5 --input_size 80x400 --patch_size 80x2 --noise_ratio 0.2
OMP_NUM_THREADS=1 torchrun --nproc_per_node=4 -m speech.train_speech --loss_m2d 1. --loss_off .5 --input_size 80x512 --patch_size 80x2 --noise_ratio 0.2 --batch_size 256 --accum_iter 2
OMP_NUM_THREADS=1 torchrun --nproc_per_node=4 -m speech.train_speech --loss_m2d 1. --loss_off .5 --input_size 80x608 --patch_size 80x2 --noise_ratio 0.2 --batch_size 256 --accum_iter 2

3-2-1. Major pre-training options

  • --batch_size: Batch size per GPU, 512 by default.
  • --epochs: Training epochs, 1000 by default.
  • --accum_iter: Iterations to accumulate gradients, 1 by default.
  • --input_size: Input spectrogram size, 80x208 by default.
  • --patch_size: Patch size, 80x4 by default.
  • --mask_ratio: Masking ratio, 0.6 by default.
  • --loss_m2d: Loss ratio for M2D masked prediction, 1.0 by default.
  • --loss_off: Loss ratio for offline target, 0.0 by default.
  • --blr: Base learning rate: absolute_lr = base_lr * total_batch_size / 256.
  • --csv_main: A CSV file to list sample files in the main dataset, 'data/files_ls960_hybrid.csv' by default.
  • --csv_bg_noise: A CSV file to list sample files in the BG noise dataset, 'data/files_audioset.csv' by default.
  • --noise_ratio: Noise mixing ratio, 0.2 by default.

4. SUPERB Evaluation

We provide upstream wrapper implementation, which you can import to your SUPERB environment.

  • Copy the superb/upstream/m2d folder under your s3prl/upstream folder.
  • Make a symbolic link to your copy of M2D repository under your s3prl/upstream/m2d, making s3prl/upstream/m2d/m2d. The wrapper files will find M2D programs under this symbolic link.
  • You will need to run pip install -e . under your s3prl folder, so that you install your local SUPERB in your Python environment.

Please refer to superb/upstream/m2d/README.md for more details.

Acknowledgements

We appreciate these publicly available resources.

References

If you find our M2D-S useful in your research, please consider citing our paper:

@article{niizumi2023m2d4speech,
    title   = {{Masked Modeling Duo for Speech: Specializing General-Purpose Audio Representation to Speech using Denoising Distillation}},
    author  = {Daisuke Niizumi and Daiki Takeuchi and Yasunori Ohishi and Noboru Harada and Kunio Kashino},
    journal = {to appear at Interspeech}, 
    year    = {2023},
    url     = {https://arxiv.org/abs/2305.14079}
}