7.9.31. Google Speech Command Recognition

12-class keyword spotting from 1-second audio clips, using MFCC features and a depthwise separable CNN (DSCNN) on the MSPM0G5187 with NPU.

Example folder: examples/google_speech_command (GitHub)

7.9.31.1. Overview

Task: audio classification (audio_classification, target_module: audio), 12-class keyword spotting
Application: voice command recognition
Dataset: Google Speech Commands v0.02, 12-class variant
Device: MSPM0G5187 (config_MSPM0.yaml, the only config in the folder)
Model: DSCNN_NPU

7.9.31.2. Keyword Classes

Type	Labels
Known keywords (10)	`down`, `go`, `left`, `no`, `off`, `on`, `right`, `stop`, `up`, `yes`
Unknown	`_unknown_`: all other words (for example bird, cat, tree)
Silence	`_silence_`: 1-second clips cut from `_background_noise_`

7.9.31.3. The Dataset

The config downloads the prepared dataset from https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/google_speech_commands_12class.zip, so you can run the example directly. To regenerate it from the original data, run the helper script (needs torch, torchaudio, scipy, pydub, numpy and tqdm) and point dataset.input_data_path at the result:

cd examples/google_speech_command
python generate_dataset.py

The script downloads Google Speech Commands v0.02 through TorchAudio, copies the 10 keyword classes into their own folders, maps all other words into _unknown_ (prefixing the original label to avoid file name collisions), and cuts the background noise files into 1-second windows with 50% overlap for _silence_. The output is SpeechCommands/classes/<class>/.

7.9.31.4. Features

The config uses the feature extraction preset GoogleSpeechCommands_MFCC_Default:

Parameter	Value
Sampling rate	16000 Hz
Audio duration	1000 ms
Frame length / step	30 ms / 20 ms
MFCC coefficients	10
Mel bins	40

The output feature shape is [N, 1, 49, 10]: batch, 1 channel, 49 time frames and 10 MFCC coefficients.

7.9.31.5. Model

DSCNN_NPU is a depthwise separable CNN designed for efficient inference on the NPU: a 10x4 convolution with stride 2, dropout, four blocks of depthwise 3x3 plus pointwise 1x1 convolution, dropout, adaptive average pooling and a fully connected layer to 12 classes (64 filters). Splitting a convolution into depthwise (spatial) and pointwise (channel mixing) parts cuts computation and model size while keeping good keyword spotting accuracy. Other audio models enabled in 1.5.0 are DSCNN_32K_NPU, DSCNN_GB_NPU, TCDS_ResNet_NPU and TCDS_ResNet_FB_NPU (see Wake Word Detection).

7.9.31.6. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/google_speech_command/config_MSPM0.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\google_speech_command\config_MSPM0.yaml

7.9.31.7. Configuration

common:
  target_module: audio
  task_type: audio_classification
  target_device: MSPM0G5187
dataset:
  dataset_name: google_speech_commands_12class
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/google_speech_commands_12class.zip
data_processing_feature_extraction:
  feature_extraction_name: GoogleSpeechCommands_MFCC_Default
  sampling_rate: 16000
  audio_duration_ms: 1000
training:
  model_name: DSCNN_NPU
  batch_size: 64
  training_epochs: 20
  num_gpus: 1
  quantization: 2
  learning_rate: 0.1
  weight_decay: 1e-5
testing: {}
compilation: {}

num_gpus: 1 expects a GPU; use 0 for CPU. Training on this dataset is long compared with the other examples. See Audio Classification for the audio task type.

7.9.31.8. Results

The example readme does not publish accuracy numbers. In the automated 1.5.0 config test run the training reached a best quantized epoch (epoch 7) with a validation accuracy of 91.30%, an F1 score of 0.913 and an AUC ROC of 0.997. That run hit the test harness time limit before it finished, so it is a partial result and not a converged 20-epoch result.

7.9.31.9. Output Location

Results are written to ./data/projects/google_speech_commands_12class/run/<date-time>/DSCNN_NPU/. See Understanding the Run Output for the generated files and Project Compilation Guide: Timeseries Classification Examples for device deployment. The MSPM0 SDK 2.10.04 or later was listed for this example in 1.4.0.