5.5. Audio and Radar Dataset Format

Tiny ML ModelZoo 1.5.0 adds two new task types, audio_classification and radar_classification. Both read datasets with the same classes/ layout used for time series classification, so the folder name of each class becomes its label. Read this page together with Data Splitting, which applies unchanged.

5.5.1. Audio Classification

Task overview: Audio Classification.

5.5.1.1. Directory structure

my_audio_dataset/
├── classes/                      # default data_dir for audio
│   ├── yes/
│   │   ├── 0001.wav
│   │   └── 0002.wav
│   ├── no/
│   │   └── 0001.wav
│   └── _silence_/
│       └── 0001.wav
└── annotations/                  # optional, generated if missing
    ├── instances_train_list.txt
    ├── instances_val_list.txt
    └── instances_test_list.txt

  • Audio files are read from classes/<class_name>/*.wav or *.flac (upper-case extensions are also accepted). Files nested deeper than one folder below classes/ are not found.

  • Every subfolder of classes/ is a class. Files whose path contains _background_noise_ are ignored when the split lists are generated.

  • Set data_processing_feature_extraction.sampling_rate and audio_duration_ms to match your recordings. The shipped examples use 16000 Hz / 1000 ms (Google Speech Commands), 8000 Hz / 2000 ms (wake word), 44100 Hz / 1000 ms (glass break) and the 8820 Hz LPC preset (cough detection).

  • The audio feature-extraction presets set mono: True; normalize_audio is a preset option.

  • Annotation lists use the same names and search order as other tasks (see Classification Dataset Format); each line is a path such as yes/0001.wav.

5.5.1.2. Preparing public datasets

1.5.0 includes helper scripts in the examples folder to turn raw or public data into this layout:

Script

Purpose

examples/google_speech_command/generate_dataset.py

Arranges the Google Speech Commands data into class folders: ten command words (down, go, left, no, off, on, right, stop, up, yes) plus _unknown_ and _silence_ (silence clips are cut from _background_noise_).

examples/cough_detection/prepare_dataset.py

Converts raw Edge AI Studio CSV recordings to segmented 8820 Hz WAV clips in class folders. Use --data-dir and --output-dir.

The prepared datasets used by the examples are also available as ZIP files (see Google Speech Command Recognition, Wake Word Detection, Cough Detection and Glass Break Detection).

5.5.1.3. Config

common:
  task_type: audio_classification
  target_device: MSPM0G5187

dataset:
  dataset_name: my_audio_dataset
  input_data_path: /path/to/my_audio_dataset   # folder, ZIP or URL

data_processing_feature_extraction:
  feature_extraction_name: GoogleSpeechCommands_MFCC_Default
  sampling_rate: 16000
  audio_duration_ms: 1000

5.5.2. Radar Classification

Task overview: Radar Classification.

5.5.2.1. Directory structure

my_radar_dataset/
├── classes/                      # default data_dir for radar
│   ├── standing/
│   │   ├── capture1.csv
│   │   └── capture2.csv
│   ├── falling/
│   │   └── capture1.csv
│   └── lying/
│       └── capture1.csv
└── annotations/                  # optional, generated if missing
    ├── instances_train_list.txt
    ├── instances_val_list.txt
    └── instances_test_list.txt

  • Files are .csv or .npy placed directly in a class folder.

  • Each row is one frame of radar point-cloud features and the data is windowed with SimpleWindow (frame_size rows per window).

  • A CSV may have a header row. When the first row is non-numeric it is treated as the header and any column whose name contains session, recording, time, track, tid or classification is dropped. Remaining columns must be numeric.

  • If different classes have different numbers of columns, all classes are truncated to the narrowest one (a warning is logged).

  • Class names are the folder names, sorted alphabetically.

5.5.2.2. Preparing the public dataset

examples/radar_pose_and_fall_detection/pose_and_fall_dataset.py is an optional preprocessing step (point-count and height/distance filtering, train split). You only need it if you do not use the public radar_human_pose_detection dataset referenced in the example config. See Radar Pose and Fall Detection.

5.5.2.3. Config

This is the configuration of examples/radar_pose_and_fall_detection/config.yaml:

common:
  task_type: radar_classification
  target_device: IWRL6432

dataset:
  dataset_name: radar_human_pose_detection
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/radar_human_pose_detection.zip
  split_type: within_files
  split_factor: [0.8, 0.1, 0.1]

data_processing_feature_extraction:
  data_proc_transforms: [SimpleWindow]
  frame_size: 8
  stride_size: 0.5

5.5.3. See Also