KWS, ASR or Speaker Verification: Which Voice Model Do You Need?
These three technologies answer different questions: “Did the user say a known phrase?”, “What did the user say?”, and “Is this the approved speaker?”
Keyword spotting (KWS)
KWS detects a small, predefined set of words or phrases. A wake word such as “Hey device” or a command such as “turn on the light” can be mapped directly to an application callback.
KWS is a strong fit when the vocabulary is fixed and the product needs low latency, local inference and a small model.
Automatic speech recognition (ASR)
ASR converts more open-ended speech into text. It is useful when users can say different sentences, names, numbers or details that cannot be listed in advance.
ASR usually needs more memory and compute, and the product still needs intent parsing after transcription.
Speaker verification
Speaker verification checks whether the voice matches an enrolled person. It answers “who is speaking?” rather than “what command was spoken?” It can be combined with KWS or ASR for actions that require user identity.
Quick decision table
| Product requirement | Best starting point |
|---|---|
| Wake a device with one phrase | Single-keyword KWS |
| Control a product with 2–10 known commands | Multi-keyword KWS |
| Understand arbitrary sentences | ASR + intent parsing |
| Allow only an enrolled user | Speaker verification, often combined with KWS |
The correct choice comes from the product vocabulary and risk model. Test the chosen system with real users and the final device before shipping.
Create a custom KWS model →