7.9.32. Wake Word Detection

Always-on detection of the wake word “OK Kilby” on the MSPM0G5187 with NPU, using a learned filterbank front-end and a TCDS-ResNet classifier. New in 1.5.0.

Example folder: examples/wake_word_detection (GitHub)

7.9.32.1. Overview

Task: audio classification (audio_classification, target_module: audio), 2 classes
Application: low-power, privacy-preserving voice interfaces for smart home, industrial and wearable devices. The wake word is detected on the device so no audio leaves it, and downstream voice-command processing is triggered only after a detection.
Device: MSPM0G5187 (config_MSPM0.yaml, the only config in the folder)
Model: TCDS_ResNet_FB_NPU
Targets in the readme: under 200 ms detection latency from the end of the utterance, a high detection rate with a low false-accept rate, and low average power for always-on, battery-operated listening

7.9.32.2. The Dataset

The wake word is “OK Kilby”. The data is a TI-internal dataset collected for this application and is not derived from a public dataset. It contains multiple speakers saying the wake word with varying tone, distance and speaking rate, background noise (household, office, ambient) and non-wake-word speech to reduce false accepts.

Class	Files in this example
Wake word (“OK Kilby”)	6000
Non-wake-word / background	4600

The subset used by the example is downloaded by the config from https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/wake_word_detection.zip. The readme notes that it is a small part of an internal dataset of more than 50 GB, so for a production-quality model you should collect and add your own wake word and background data (speakers, accents, distances, microphones and noise conditions).

7.9.32.3. Feature Extraction: Learned Filterbank

Instead of a fixed MFCC or FFT front-end, the model learns its filterbank jointly with the classifier. The config uses the preset WakeWordDetection_Filterbank_Default.

Audio is sampled at 8 kHz with a 2-second window per inference (16,000 samples).
A strided Conv1D filterbank (kernel 64, stride 4) slides over the raw audio and projects it into 32 channels (fb_output_channel of the WakeWordDetection_Filterbank_Default preset).
A max-pool over each 20 ms context (160 samples) produces one feature vector per frame.
Filterbank weights are quantized to 2 bits and the input audio to 16 bits.
The features feed directly into the TCDS-ResNet backbone.

The commented lines in the config show the underlying preset keys (audio_feature: "FB", fb_conv_kernel, fb_output_channel, fb_bitwidth, fb_context_ms, fb_conv_stride, input_bit_depth).

7.9.32.4. Model

TCDS_ResNet_FB_NPU combines the learned filterbank with a Temporal Channel-Decoupled Separable ResNet:

Stem: Conv3x1, BatchNorm, ReLU (32 channels)
Backbone: 3 TCDS basic blocks (channels 32, 48, 64, 96). Each block has a main branch of depthwise 9x1, pointwise 1x1, depthwise 9x1, pointwise 1x1, and a skip branch (1x1 projection when the stride or channel count changes). The output is ReLU(main + skip).
Head: Dropout, Flatten, Fully Connected, 2 outputs
Mixed precision: stem and head at 8 bits, residual blocks at 2 bits

The non-filterbank variant TCDS_ResNet_NPU is also available. The audio models enabled in 1.5.0 are DSCNN_NPU, DSCNN_32K_NPU, DSCNN_GB_NPU, TCDS_ResNet_NPU and TCDS_ResNet_FB_NPU.

7.9.32.5. Running the Example

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

./run_tinyml_modelzoo.sh examples/wake_word_detection/config_MSPM0.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\wake_word_detection\config_MSPM0.yaml

7.9.32.6. Configuration

common:
  target_module: audio
  task_type: audio_classification
  target_device: MSPM0G5187
dataset:
  dataset_name: wake_word_detection
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/wake_word_detection.zip
data_processing_feature_extraction:
  feature_extraction_name: WakeWordDetection_Filterbank_Default
  sampling_rate: 8000
  audio_duration_ms: 2000
training:
  model_name: TCDS_ResNet_FB_NPU
  batch_size: 4096
  training_epochs: 100
  num_gpus: 1
  quantization: 2
  learning_rate: 0.01
  weight_decay: 1e-5
  auto_quantization: False
testing: {}
compilation: {}

quantization: 2 with auto_quantization: False runs quantization-aware training with the mixed-precision scheme above. The example uses a batch size of 4096 and a learning rate of 0.01, as in the config shown here. num_gpus: 1 expects a GPU; use 0 for CPU. For the audio task type see Audio Classification.

7.9.32.7. Results

The example readme does not publish accuracy numbers. In the automated 1.5.0 config test run (test set of 1080 samples: 600 wake word, 480 others) the quantized model reached a test accuracy of 96.20% and an AUC ROC of 0.991. Treat these as indicative results from a short test run on the small subset.

7.9.32.8. Output Location

Results are written to ./data/projects/wake_word_detection/run/<date-time>/TCDS_ResNet_FB_NPU/. The compiled model and interface headers are in compilation/artifacts/. See Understanding the Run Output and Project Compilation Guide: Timeseries Classification Examples.

7.9.32.9. Running on Device

Build the deployment with the MSPM0 SDK (2.08.00 or later) and Code Composer Studio, with a microphone as audio input. The compiler flow uses the TI Neural Network Compiler; see Supported Devices for MSPM0G5187.