Wake words and keyword spotting.
How a device listens for one short phrase, all the time, on almost no power, and almost never gets it wrong.
Say a wake word and a device that looked asleep is suddenly listening. Behind that small moment sits keyword spotting: the task of detecting one short phrase in a continuous stream of sound, as a quick yes-or-no decision rather than a full transcription. We walk it the way the field is built: what it is, why it is hard, how a modern detector does it, what it unlocks, and where it is heading. The papers behind each idea are linked, so you can read the source yourself.
The three ideas to carry: always-on listening, the false-accept versus false-reject tradeoff, and small streaming models.
Detecting one short phrase in a stream of sound.
The field name is keyword spotting (KWS). The wake word is the special case that turns a system on.
You already meet it everywhere. The wake words on phones and smart speakers, also called hotwords or trigger phrases, all do the same job: a tiny piece of software has been listening the whole time, waiting for one short phrase and ignoring all the rest. When it hears the phrase, it wakes the device and starts the heavier work.
The job sounds narrow, and that narrowness is the whole point. Instead of transcribing everything you say, the device only has to answer one yes-or-no question, over and over. That is the contrast this page builds on: keyword spotting is not speech-to-text. Speech-to-text transcribes everything; keyword spotting answers one cheap question, so it can run all the time. Speech-to-text is what runs after the wake word.
- Audio frames
Short windows reach the keyword detector.
- Keyword score
Estimate whether the trigger phrase occurred.
- Threshold and checks
Decide whether to activate the assistant.
- Larger pipeline
Begin the requested speech interaction.
The words you will meet everywhere.
A short glossary so the rest of the page reads cleanly. These terms show up in every explainer and paper in the field.
| Term | Meaning |
|---|---|
| Wake word | The phrase intended to activate the system. |
| Closed vocabulary | One of a fixed set of trained phrases. |
| User-defined keyword | A phrase configured by the user. |
| Query by example | A phrase specified using spoken examples. |
Two opposite things have to be right at once.
Wake up reliably when you do say the phrase, and stay quiet through hours of conversation, television and noise that must never be mistaken for the trigger.
False accepts versus false rejects.
A detector that wakes too easily fires during TV, chatter and noise. One tuned to never false-fire starts missing the real phrase. The honest way to describe performance is a curve, not a single accuracy number.
Small memory, low power.
Always-on means few parameters, low compute, and ideally a tiny stage that can run while the main processor sleeps. Model size and arithmetic cost are primary constraints, not afterthoughts.
It has to feel instant.
The detector streams. It has to decide within a fraction of a second of the phrase ending, which rules out approaches that need to see a whole utterance before answering.
Far-field, noise, and playback.
Audio across a room, reverberation, background speech, music, traffic, accents, children's voices, and the device's own audio playing back all sit between the microphone and a clean decision.
Words that sound near the trigger.
Phrases that rhyme with or sound close to the trigger drive false accepts, so detectors need explicit hard-negative training on the near-misses.
Few recordings, no shared yardstick.
A new wake word may have very few real recordings, so teams lean on synthetic speech and augmentation. And without a shared dataset and a fixed metric, numbers are not comparable, which is why open benchmarks exist.
- Vertical direction
- More missed intended triggers.
- Horizontal direction
- More false activations per hour.
Small, streaming, and tuned on a curve.
The throughline of the modern field: a compact streaming classifier whose operating point is chosen on the false-accept and false-reject curve, with a stronger second stage confirming before anything heavy runs.
- 01HMM
Model keyword and filler sequences.
- 02DNN
Predict keyword scores.
- 03CNN
Share filters across positions.
- 04Streaming
Reuse state as audio arrives.
The field's throughline: same job, steadily smaller and faster models, accuracy held steady as the budget shrinks.
1The shift to deep classification
From keyword-and-filler to a yes-or-no classifier.
Classic systems used keyword and filler hidden Markov models (HMM): one model for the keyword, one for everything else, with a likelihood-ratio decision. Workable, but heavier and less robust than what replaced it.
The shift was to treat the problem as direct classification. A small neural network looks at a sliding window of audio features and predicts the probability of the keyword, or its sub-word pieces, frame by frame, followed by a simple posterior-smoothing step that makes the firing decision. This is the move that made small-footprint, on-device wake words practical, in Chen, Parada and Heigold (2014).
Horizontal: incoming audio over time. Dashed line: activation threshold.
2Better accuracy per parameter
Convolutional models for a tighter budget.
Swap the fully connected network for a convolutional one (CNN) that shares parameters across time and frequency. It captures local spectral patterns with fewer weights, a better accuracy-per-parameter trade for the tight always-on budget, shown in Sainath and Parada (2015).
Later work goes deeper while staying small. Coucke et al. (2019) adds dilated convolutions, gating and residual connections, WaveNet-style, for a deeper but still small-footprint detector, and releases an open wake-word dataset so a new phrase does not need thousands of real recordings.
Input frames
| Filter application | Weights used |
|---|---|
| Frames 1–3 | Same learned filter |
| Frames 2–4 | Same learned filter |
| Frames 3–5 | Same learned filter |
3Streaming and end-to-end
One network that reuses its own work.
Fold feature handling and the decision into one network, trained end-to-end to emit a keyword score directly. Alvarez and Park (2019) design it to run streaming with memoized activations, so it reuses past computation instead of reprocessing a window each step. The training target is set so the model only needs to detect the end of the keyword, which is what makes it both cheap and instant.
The result is a detector that scores each moment as audio arrives and fires the instant the phrase completes, rather than waiting for a clean utterance boundary.
- Frame t
Compute new features with retained context.
- Frame t + 1
Reuse useful state and add the new frame.
- Phrase complete
Update the keyword score and apply the decision rule.
4A shared benchmark and metrics
How the field measures progress honestly.
None of the above is comparable without a common yardstick. Warden (2018) introduces the Speech Commands dataset, an open set of short spoken words, plus defined accuracy and streaming metrics, so the field can measure progress on small-vocabulary, on-device recognition in a way others can reproduce. It is a widely used starting point for benchmarking keyword spotting.
The honest framing the benchmark encourages: report the false-accept and false-reject behavior as a curve, on a shared dataset, rather than a single headline accuracy that no one else can check.
- Shared evaluation audio
Include intended phrases, speech without the trigger, and noise.
- Fixed metrics
Report false rejects and false accepts at stated operating points.
- Runtime conditions
Include latency, power, or memory when relevant.
A clean front door for the whole speech loop.
At the field level: a hands-free, eyes-free start, a device that can listen for a trigger without running heavy compute the whole time, and a gate that turns on the rest of the pipeline only when you actually want it.
Wake-word detection is the gate that turns on speech-to-text, turn detection and the rest of the pipeline only when you mean to use them. Get that first moment right and everything built on top feels effortless. The hard cases in the section above are exactly the ones that decide whether starting to talk feels instant or fiddly.
Hands-free, eyes-free start.
You begin with your voice, no button, no screen. That is what makes voice help usable while driving or cooking.
Listen without running heavy compute.
A tiny always-on stage means the power-hungry recognizer and any networking only wake on demand. We state the mechanism generally and make no on-device or privacy claim about any one product.
Responsiveness you can feel.
Because the trigger is detected locally in a fraction of a second, the system feels instant before the real request is even processed.
Custom and branded triggers.
Open-vocabulary and query-by-example approaches let products define their own wake phrase from limited data.
For a sales voice co-pilot the bridge is at the outcome level: starting a voice interaction should feel instant and effortless, you speak and it is listening, nothing to press. Everything on this page is the groundwork for that first moment. The honest one-liner: the easiest part of talking to a voice tool is the part you never think about, and that is the point.
The direction the field is heading.
We describe the direction, honestly including the open problems, not a roadmap of our own.
From fixed phrases to your own.
From a phrase baked into the model toward user-defined keywords, including query-by-example that learns a trigger from a few spoken samples.
Learned, streaming, and tiny.
From hand-built features to fully learned end-to-end streaming models, and smaller still via quantization and architecture search for ever-tinier hardware.
Holding up in the real world.
Far-field, multi-speaker and high-noise settings, and confusable phrases, proven in real deployments rather than on clean test sets.
Wake for the right person.
Speaker-conditioned triggers that wake for a specific person, balanced carefully against privacy.
A new keyword from little data.
Synthetic and data-efficient training so a new phrase does not need thousands of real recordings.
The honest unfinished work.
Driving down false accepts without hurting wake rate, fair performance across accents and ages, and evaluation that reflects real always-on conditions, not a single benchmark number.
The papers behind this page.
- Chen, Parada, Heigold (2014). Small-Footprint Keyword Spotting using Deep Neural Networks. ICASSP 2014.
- Sainath, Parada (2015). Convolutional Neural Networks for Small-Footprint Keyword Spotting. Interspeech 2015.
- Warden (2018). Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition. arXiv:1804.03209.
- Alvarez, Park (2019). End-to-End Streaming Keyword Spotting. arXiv:1812.02802.
- Coucke et al. (2019). Efficient Keyword Spotting Using Dilated Convolutions and Gating. arXiv:1811.07684.
Keep reading.
Sibling explainers in the Research section. Keyword spotting sits at the front of the speech loop, so these are what come next.
Talk. Done.
The easiest part of talking to a voice tool is the part you never think about. You talk, and clean, complete notes are waiting. See it on a real call.
No app required. Just your phone number.