Keyword Spotting
A technique that continuously monitors an audio stream locally to identify preset keywords or wake words. It typically employs a lightweight model running persistently at low power consumption, avoiding the upload of audio to the cloud.
Timeline (last 90 days)
In offline voice control, wake word detection and command recognition are split into a two-stage architecture: the audio stream is continuously fed to the wake word model first, and once the wake word is hit a brief time window opens, with the audio within that window handed to a multi-keyword model for command recognition.
There are open-source examples that use getUserMedia() to capture the microphone audio stream, convert it to the frame format required by the model, run keyword detection inference locally, and only start ASR or show the command interface after a wake word is detected.
Wake word training is essentially a binary classification problem in keyword spotting (KWS); common lightweight architectures include DS-CNN and LSTM-based variants, which can be compressed to tens of kilobytes for local deployment on MCUs.
Shallow verification methods such as string matching and keyword detection will almost inevitably fail when faced with an AI capable of optimization.
All Facts (4)
In offline voice control, wake word detection and command recognition are split into a two-stage architecture: the audio stream is continuously fed to the wake word model first, and once the wake word is hit a brief time window opens, with the audio within that window handed to a multi-keyword model for command recognition.
50%UnverifiedThere are open-source examples that use getUserMedia() to capture the microphone audio stream, convert it to the frame format required by the model, run keyword detection inference locally, and only start ASR or show the command interface after a wake word is detected.
50%UnverifiedWake word training is essentially a binary classification problem in keyword spotting (KWS); common lightweight architectures include DS-CNN and LSTM-based variants, which can be compressed to tens of kilobytes for local deployment on MCUs.
50%UnverifiedShallow verification methods such as string matching and keyword detection will almost inevitably fail when faced with an AI capable of optimization.
50%