KWS
Keyword spotting, a technique that detects predefined phrases or wake words in an audio stream and outputs a command ID. It is suited to offline voice control with a fixed vocabulary and is lighter and more energy-efficient than full ASR.
Timeline (last 90 days)
The industry typically uses speaker diarization to distinguish the parties in a call, then combines keyword spotting (KWS) and sentiment classification models for downstream analysis.
The ASR processing path is audio → speech-to-text → text matching/intent parsing → action, while the KWS processing path is audio → acoustic features → keyword model → command id → action.
KWS cannot replace free-form voice interaction; if users need to say commands like "set the living room light to 50%" that contain locations and values, the product typically still needs ASR and intent parsing.
For fixed-vocabulary voice control scenarios (such as a wake word and a few fixed commands), keyword spotting (KWS) is more suitable than always running a general ASR model.
A small KWS pipeline is suitable for scenarios with a fixed vocabulary, requiring local or offline operation, battery-powered or resource-constrained, where low latency is important, and where continuous audio is not desired to be sent to cloud services.
Before product release, at minimum it is necessary to measure real trigger recall rate, false activations per hour, distance and speaking volume, different speakers and accents, TV/music/fan and room noise, model size and device memory, and end-to-end latency and power consumption impact.
In always-listening offline voice systems, the keyword model typically outputs scores frame by frame or over a fixed sliding window; a single frame exceeding the threshold does not mean the complete phrase was spoken, and confirmation across consecutive frames (e.g., 3-5 consecutive frames above threshold) and a cooldown period (ignoring retriggering of the same keyword within a few hundred milliseconds) are needed.
All Facts (7)
The industry typically uses speaker diarization to distinguish the parties in a call, then combines keyword spotting (KWS) and sentiment classification models for downstream analysis.
50%UnverifiedThe ASR processing path is audio → speech-to-text → text matching/intent parsing → action, while the KWS processing path is audio → acoustic features → keyword model → command id → action.
50%UnverifiedKWS cannot replace free-form voice interaction; if users need to say commands like "set the living room light to 50%" that contain locations and values, the product typically still needs ASR and intent parsing.
50%UnverifiedFor fixed-vocabulary voice control scenarios (such as a wake word and a few fixed commands), keyword spotting (KWS) is more suitable than always running a general ASR model.
50%UnverifiedA small KWS pipeline is suitable for scenarios with a fixed vocabulary, requiring local or offline operation, battery-powered or resource-constrained, where low latency is important, and where continuous audio is not desired to be sent to cloud services.
50%UnverifiedBefore product release, at minimum it is necessary to measure real trigger recall rate, false activations per hour, distance and speaking volume, different speakers and accents, TV/music/fan and room noise, model size and device memory, and end-to-end latency and power consumption impact.
50%UnverifiedIn always-listening offline voice systems, the keyword model typically outputs scores frame by frame or over a fixed sliding window; a single frame exceeding the threshold does not mean the complete phrase was spoken, and confirmation across consecutive frames (e.g., 3-5 consecutive frames above threshold) and a cooldown period (ignoring retriggering of the same keyword within a few hundred milliseconds) are needed.
50%