Voice pipeline layer comparison
Addressee detection vs turn detection
Turn detection finds when speech ends. Addressee detection finds whether it was meant for you.
Architecture comparison
Only addressee detection decides who the speech was for.
Endpointing fires on every completed turn, addressed or not. Addressee detection filters before ASR ever runs.
turn detection
Has the speaker finished this turn?
Selective Auditory Attention (SAA) addressee detection
Was the speech directed at the device?
| Dimension | Turn detection (endpointing) | SAA addressee detection | Why it matters |
|---|---|---|---|
| Core question | Has the speaker finished this turn? | Is this speech addressed to the device? | Different questions, different answers, both required |
| Pipeline position | After ASR begins, pre-response trigger | Pre-ASR, before transcription | Addressee decision must precede transcription to gate it |
| Signal used | Trailing silence, prosodic cues, utterance completeness | Acoustic and visual attention cues, multi-party context | Turn timing and addressee intent are acoustically distinct signals |
| Decides who speech is for? | No | Yes | Only the addressee gate can stop bystander false activations |
| Behavior under uncertainty | Responds after timeout | Fails closed, nothing forwarded | Fail-closed prevents spurious model calls on ambiguous audio |
| Required in multi-party environments? | Yes | Yes, and must run first | Skip this gate and cross-talk still reaches ASR |
Complementary pipeline stages
Addressee detection and turn detection solve different problems.
The pipeline runs VAD, then SAA in real time, then ASR, then endpointing, then the LLM, each stage exactly once.
false triggers and cost
Turn detection fires on every completed turn
Endpointing fires regardless of addressee. ASR transcribes, the LLM reasons, TTS may respond.
Every non-addressed turn that passes the gate is billed as a real inference call.
when the gate is required
Any ambient speech reintroduces the addressee problem
Endpointing alone works only with one speaker and no background voices.
A bystander, a TV, or a nearby conversation reintroduce it; SAA gates that audio out before ASR runs.
Where the gap appears
Turn detection fires on every speaker equally, addressed or not.
Below, non-addressed turns reach ASR without a gate, and the pipeline bills for each one.
drive-thru lanes
Cross-talk between customer and passenger
Customers turn to companions mid-order.
Each exchange is a completed turn endpointing will capture and forward.
SAA holds those at the gate and passes only audio addressed to the agent.
meeting rooms
Participants addressing each other, not the device
A meeting assistant receives turns from every participant.
Without an addressee gate, any completed turn triggers a response attempt.
SAA gates on whether the speaker was addressing the assistant or a colleague.
car cabins
Passenger conversation vs device commands
Passenger chatter and driver commands sound alike; endpointing fires on both.
Addressee detection tells them apart before ASR runs.
Methodology
Measured on held-out multi-party sessions.
Published as "SAA (Selective Auditory Attention): Device-Addressed Speech Detection for Real-Time Voice AI" (arXiv:2604.08412).
Two caveats apply: accuracy drops under distribution shift, and cross-lingual recall is limited.
audio and video fusion
High accuracy on the addressee decision
Held-out multi-party sessions, audio and video combined.
audio-only
Reliable on audio alone
Real multi-speaker audio, no video stream required. Hosted or embedded.
Turn detection decides when. Addressee detection decides whether. The whether comes first.
Common questions
What builders ask about turn detection and addressee detection.
What is turn detection and what question does it answer?
Turn detection, also called endpointing, answers a timing question: has the speaker finished their turn.
It monitors trailing silence, prosodic drop, and end-of-utterance cues.
It carries no information about whether the speech was directed at the device or at another person.
What is addressee detection and how is it different from turn detection?
Addressee detection answers an intent question: was this speech directed at the device, or at another person.
It runs pre-ASR, before transcription begins, routing audio or dropping it.
Turn detection runs after speech is underway and decides when to stop listening.
They are sequential: addressee detection gates what the pipeline sees, turn detection gates when it responds.
Can turn detection alone prevent false activations in multi-party rooms?
No. Turn detection fires on every completed turn regardless of addressee.
A drive-thru customer who says "you want fries?" to a passenger completes a well-formed turn; endpointing fires and the agent responds.
Turn detection performed exactly as designed; it was never designed to decide who the speech was for.
Where does SAA sit in the pipeline relative to turn detection?
SAA runs before ASR and before turn detection.
The order is: VAD (drops silence and noise), SAA (decides addressee in real time, before STT), ASR (transcribes only what SAA forwards), endpointing (detects turn completion), then the LLM.
SAA and turn detection are not alternatives; they occupy adjacent, non-overlapping positions in the pipeline.
If my endpointing is already accurate, do I still need addressee detection?
Yes, in any environment with more than one voice source.
Accurate endpointing means the agent responds exactly when a turn ends, including turns addressed to someone else.
Endpointing alone suffices only when a single speaker has no background voices. Any ambient or bystander speech requires an addressee gate before ASR.
Get in touch
Request an eval or book a call.
Tell us about your voice stack for a free evaluation, licensed per device. Or try the Cloud SDK first to get a feel.
Last updated June 29, 2026
Compare other layers
The addressee-detection comparison cluster.
comparison
What is addressee detection?
The category cornerstone.
What addressee detection is, why it runs before ASR, and how it differs from every prior-art layer in the voice pipeline.
Read the explainercomparison
Addressee detection vs VAD
VAD detects that someone is speaking.
Addressee detection decides if that speech is meant for the device, before ASR runs.
See the comparisoncomparison
Addressee detection vs wake words
Wake words require a trigger phrase and still misfire on background speech.
Addressee detection works on natural speech with no required phrase and no enrollment step.
See the comparisoncomparison
Addressee detection vs speaker diarization
Diarization labels who spoke after transcription.
Addressee detection gates who is addressed before transcription runs.
See the comparisoncomparison
Addressee detection vs noise suppression
Noise suppression cleans audio; it does not decide who the clear voice was talking to.
See the comparison