Skip to content

Latest commit

 

History

History
154 lines (100 loc) · 7.46 KB

UsersManual.md

File metadata and controls

154 lines (100 loc) · 7.46 KB
Raw

ToyADMOS2 Dataset Users Manual

For making new dataset with your own configurations, make your recipe file, then run a mixer tool (a python script).

1. Mixer tool (mixer.py)

This tool mixes normal or anomaly samples with environmental noise samples at a specified SNR.

Usage: mix_dataset.py SRC_FOLDER DEST_FOLDER RECIPE_FILE SNR_DB
    SRC_FOLDER   Path to the ToyADMOS2 folder. 
    DEST_FOLDER  Path to a folder you want to create.
    RECIPE_FILE  Recipe file for describing the use of samples.
    SNR_DB       SNR in dB (-6, 6, or any integer), or `clean`.

One more option is written in recipe file:

fig-recipe-settings.png

If True, all the source domain normal samples will be shuffled after mixing all samples. This is for alleviating the data distribution problem that is seen when a dataset is composed of a small number of recordings and the difference of the data distributions between training and test set is too large.

1-1. Example of Runs of Mixer Tool

This will create a clean dataset compatible with the paper benchmark, using source data files from a folder ToyADMOS2, and storing the resulting files under a folder paper_clean. This doesn't mix environmental noises.

python mixer.py ToyADMOS2 paper_clean recipe_benchmark.xlsx clean

This will mix environmental noise at -6dB of the SNR.

python mixer.py ToyADMOS2 paper_clean recipe_benchmark.xlsx -6

2. Making Recipe

Which samples, for how much, with which noise, for what mics? The recipe is for configuring these details, as an Excel spreadsheet.

2-1. Overview of the Recipe File

  • Two machines: ToyCar and ToyTrain.
  • Two types of recordings: Normal sounds and anomalous sounds.
    • Normal sounds are for training and test set, whereas anomalous sounds are for test set only.
    • Normal sounds have 150 recording patterns, whereas anomalous sounds have 300 patterns.
  • A recipe file has five or more sheets: Settings & Notes, NormalToyCar, AnomalyToyCar, NormalToyTrain, and AnomalyToyTrain.
    • Settings & Notes is for writing settings and descriptions/notes of the recipe.
    • NormalToyCar, AnomalyToyCar are for configuring for toy car recordings.
    • NormalToyTrain, AnomalyToyTrain are for toy train recordings.

(You can find sheet tabs at the bottom of Excel app.)

fig-recipe-sheets.png

2-2. NormalToyCar

In the NormalXXX sheets, we have 150 patterns each.

fig-recipe-nor-car

The information on the recording patterns is written in columns A-I.

  • No. -- Recording pattern no.
  • Folder -- Relative path in the ToyADMOS2 dataset.
  • FileID -- Recording pattern ID.
  • Model -- The machine configuration (A to E).
  • CarID -- ID of machine hardware recorded (used in the recording work).
  • Speed -- The operating speed (1 to 5), corresponding to the five voltage levels, 2.8, 3.1, 3.4, 3.7, and 4.0 V.
  • Defect -- Not used for normal samples.
  • D. Level -- The damage level, not used here.
  • # of Rec. -- Number of recording samples per mic. i.e. There are actually 260 x 5 = 1,300 samples if this is 260.

You will write configuration requirements (r# = requirement) in columns named r#_xxx as follows:

fig-recipe-nor-r0

  • r0_pat -- Resulting pathname pattern. ToyCar/train/section_00_source_train_normal_?????.wav will create files starting from section_00_source_train_normal_00001.wav under a folder ToyCar/train.
  • r0_mics -- ID of mics to be used. [3] uses samples from mic 3 only, or [4, 5] uses samples from two mics 4 and 5.
  • r0_nz -- ID of environmental noise to be mixed. A single number is accepted: 1 to 4.
  • r0_qty -- Number of samples to create.

If you need two or more requirements with one recording, here you can write in r1_xxx, r2_xxx, or more requirement columns.

fig-recipe-nor-r1

  • r1_pat -- The same as r0_pat. ToyCar/source_test/section_00_source_test_normal_?????.wav will create files starting from section_00_source_test_normal_00001.wav under a folder ToyCar/source_test.
  • r1_mics -- The same as r0_mics.
  • r1_nz -- The same as r0_nz.
  • r1_qty -- The same as r0_qty. Wait, what's the R25? R means replacement. So R25 will always pick samples from the beginning. If you have 25 in r0_qty and 10 in r1_qty, the first 25 samples will be used for making files with the pattern r0_xxx, and the next 10 samples are for the patten r1_xxx.

2-3. AnomalyToyCar

In the AnomalyXXX sheets, we have 300 patterns each.

fig-recipe-ano-car

  • Defect -- The anomaly conditions defined in the Table 1 on the paper.
  • D. Level -- The damage level: High, mid, or low.
  • # of Rec. -- The same as NormalToyCar, and all the recordings has 27 samples per mic. i.e. There are actually 27 x 5 = 135 samples if we use samples from all mics.

fig-table1-defects.png

The usage of the requirements are the same with NormalXXX sheets.

2-4. NormalToyTrain

fig-recipe-nor-train

The number of mic is 8 for toy train.

  • # of Rec. -- There are actually 65 x 8 = 260 samples if this is 65.

The toy train normal recordings have about 1/4 number of samples compared to toy car.

It is designed to use the samples from four mics as one set; mic [1,2,3,4] is the first set, and [5,6,7,8] is the second.

  • r#_mics -- ID of mics: [1,2,3,4], [5,6,7,8], or any combination for your purpose. [1,2,3,4,5,6,7,8] will use all the mics.

2-5. AnomalyToyTrain

fig-recipe-ano-train

The same design with the NormalToyTrain.

  • # of Rec. -- All the recordings has 7 samples per mic. i.e. There are actually 7 x 8 = 56 samples if we use samples from all mics.

2-6. Other Notes

  • The normal or anomaly samples are used from the beginning of sorted order of list of files, whereas the environmental noise samples are randomly picked with replacement.
  • Logs are created under the destination folder, log-NormalToyCar.txt for example.

(Example log)

2021-05-01 12:03 INFO CN031-carB1-speed1_mic3_00001.wav|0.0406 + N1_mic3_01595.wav|0.0657 -> section_00_source_train_normal_00001_B1s1.wav snr=0dB
2021-05-01 12:03 INFO CN031-carB1-speed1_mic3_00002.wav|0.0406 + N1_mic3_03383.wav|0.0630 -> section_00_source_train_normal_00002_B1s1.wav snr=0dB
2021-05-01 12:03 INFO CN031-carB1-speed1_mic3_00003.wav|0.0406 + N1_mic3_03129.wav|0.0638 -> section_00_source_train_normal_00003_B1s1.wav snr=0dB
2021-05-01 12:03 INFO CN031-carB1-speed1_mic3_00004.wav|0.0406 + N1_mic3_03383.wav|0.0630 -> section_00_source_train_normal_00004_B1s1.wav snr=0dB
2021-05-01 12:03 INFO CN031-carB1-speed1_mic3_00005.wav|0.0406 + N1_mic3_04219.wav|0.0744 -> section_00_source_train_normal_00005_B1s1.wav snr=0dB
2021-05-01 12:04 INFO CN031-carB1-speed1_mic3_00006.wav|0.0406 + N1_mic3_01770.wav|0.0197 -> section_00_source_train_normal_00006_B1s1.wav snr=0dB

This shows that normal samples used are starting from CN031-carB1-speed1_mic3_00001.wav, mixed with randomly picked noise like N1_mic3_01595.wav, and output to the section_00_source_train_normal_00001_B1s1.wav at 0dB of the SNR.

The numbers following filenames (0.0406 for example) are the RMS amplitude;

  • RMS amplitude of normal or anomolous sample is the average value among samples,
  • whereas that of noise sample is calculated from the file.