voice_audio_multi/docs/models/micro_wake_word.md

microWakeWord

This runtime executes trained microWakeWord MixedNet TFLite models after conversion to GGUF. It uses the same fixed-point TensorFlow Lite Micro speech frontend as the training project: 16 kHz mono PCM, a 30 ms window, a 10 ms step, and 40 PCAN-normalized filterbank channels.

The default package is MicroWakeWord-GGUF/micro-wake-word-okay-nabu-f32.gguf under audio-cpp/audio.cpp-gguf. It is self-contained, including the model configuration and spec. Only F32 is packaged.

The default Okay Nabu checkpoint is Apache-2.0 licensed. Upstream explicitly includes the trained weights and manifests in its license scope. Retain its LICENSE and NOTICE when redistributing; other converted models may have different licenses.

Convert a model with its published manifest:

conda run -n qwen3-tts python tools/models/convert_micro_wake_word_gguf.py \
  model.tflite model.gguf --manifest model.json

The converter accepts the stateful MixedNet operator set emitted by microWakeWord. It rejects different TFLite graphs explicitly. Converted weights use F32, while the runtime preserves the TFLite activation quantization boundaries and streaming state.

Run an audio file:

build/debug/bin/audiocpp_cli --family micro_wake_word --task wake \
  --model model.gguf --audio input.wav --log --log-file logs/micro_wake_word/run.log

Detections are returned as speech segments. text contains the wake phrase, confidence contains the smoothed probability, and the sample span identifies the feature block that triggered the detection. Request options threshold and sliding_window_size override the values copied from the model manifest.

Use --mode streaming for incremental audio.

Optional NEON network

The default network remains the existing GGML implementation on the session backend. On AArch64, build with -DENGINE_ENABLE_NEON=ON and select the native stateful CNN with:

--session-option 'micro_wake_word.exec_plan={"network":{"impl":"neon"}}'

The NEON network runs on CPU, uses the session thread count, and works in both offline and streaming modes. The fixed-point frontend, input quantization, activation quantization boundaries, recurrent-state updates, probability smoothing and detection thresholds retain their existing behavior. The native network supports batch-one MixedNet manifests with unit spatial width and depthwise multiplier one.

impl: "ggml" or impl: "default" retains the original path. The component accepts no artifacts or params. Existing GGUFs can use the new session option without reconversion, including when loading their embedded spec.