google_speech_command

Google Speech Commands 12-Class Dataset

A TensorLab-compatible 12-class variant of the Google Speech Commands dataset for keyword spotting and audio classification experiments using TinyML models such as DSCNN. ---

Dataset Source

Downloaded automatically via TorchAudio:

torchaudio.datasets.SPEECHCOMMANDS(
    root=root,
    url="speech_commands_v0.02",
    folder_in_archive="SpeechCommands",
    download=True,
)

Classes

  ┌────────────────┬──────────────────────────────────────────────────────────┐
  │      Type      │                          Labels                          │
  ├────────────────┼──────────────────────────────────────────────────────────┤
  │ Known keywords │ down, go, left, no, off, on, right, stop, up, yes        │
  ├────────────────┼──────────────────────────────────────────────────────────┤
  │ Unknown        │ _unknown_ - all non-keyword words (e.g. bird, cat, tree) │
  ├────────────────┼──────────────────────────────────────────────────────────┤
  │ Silence        │ _silence_ - 1-second clips from _background_noise_/      │
  └────────────────┴──────────────────────────────────────────────────────────┘

Quick Start

• Install dependencies

python -m pip install torch torchaudio scipy pydub numpy tqdm

• Generate the dataset

python generate_dataset.py

Output Structure

  SpeechCommands/
  ├── speech_commands_v0.02/       # Original downloaded dataset
  │   ├── down/
  │   ├── go/
  │   └── _background_noise_/
  └── classes/                     # TensorLab-ready dataset
      ├── down/
      ├── go/
      ├── left/
      ├── no/
      ├── off/
      ├── on/
      ├── right/
      ├── stop/
      ├── up/
      ├── yes/
      ├── _silence_/
      └── _unknown_/

What generate_dataset.py Does:

1. Downloads Google Speech Commands v0.02 via TorchAudio
2. Copies the 10 keyword classes into their own folders
3. Maps all other word classes into unknown (prefixing the original label to avoid filename collisions)
4. Splits background_noise/ audio into 1-second clips for silence
5. Saves everything under SpeechCommands/classes/

How silence Samples Are Generated

                                            Silence samples are not recorded speech - they are synthetic clips cut from the background noise audio files in_background_noise_/.                                                                                                

Source files: all .wav files inside SpeechCommands/speech_commands_v0.02/_background_noise_/

Process (create_silence_samples):

1. Each background noise file is loaded in full via pydub.AudioSegment
2. Raw PCM samples are extracted as a NumPy array
3. The array is sliced into 1-second windows using a sliding loop with 50% overlap.
4. Each segment is written to classes/silence/ as a 16-bit .wav file

Recommended preset

  GoogleSpeechCommands_MFCC_Default = dict(
      data_processing_feature_extraction=dict(
          sampling_rate=16000,
          audio_duration_ms=1000,
          audio_feature="MFCC",
          n_mfcc=10,
          n_mels=40,
          frame_length_ms=30,
          frame_step_ms=20,
          normalize_audio=True,
          mono=True,
          variables=1,
          feat_ext_transform=["MFCC"],
          data_proc_transforms=[],
      ),
      common=dict(
          task_type=TASK_TYPE_AUDIO_CLASSIFICATION,
      ),
  )

MFCC Feature Extraction

MFCCs (Mel Frequency Cepstral Coefficients) compactly represent the frequency characteristics of speech, making them well-suited for keyword spotting.

  ┌───────────────────┬──────────┐
  │     Parameter     │  Value   │
  ├───────────────────┼──────────┤
  │ Sampling rate     │ 16000 Hz │
  ├───────────────────┼──────────┤
  │ Audio duration    │ 1000 ms  │
  ├───────────────────┼──────────┤
  │ Frame length      │ 30 ms    │
  ├───────────────────┼──────────┤
  │ Frame step        │ 20 ms    │
  ├───────────────────┼──────────┤
  │ MFCC coefficients │ 10       │
  ├───────────────────┼──────────┤
  │ Mel bins          │ 40       │
  └───────────────────┴──────────┘

Output feature shape: [N, 1, 49, 10] (batch size × 1 channel × 49 time frames × 10 MFCC coefficients)

DSCNN Model

The recommended model is DSCNN (Depthwise Separable Convolutional Neural Network), designed for efficient TinyML inference on our NPU

Architecture

Conv10x4 / stride 2 Dropout Depthwise3x3 + Pointwise1x1 ×4 Dropout AdaptiveAvgPool Fully Connected (→ 12 classes)

Filters: 64 | Output classes: 12

Why DSCNN

A standard convolution performs spatial filtering and channel mixing in one operation. DSCNN splits this into:

• Depthwise conv - spatial filtering independently per channel
• Pointwise conv - 1×1 convolution for channel mixing

This reduces computation and model size while maintaining strong keyword spotting accuracy, making it suitable for embedded deployment.