When a real-time voice app feels slow or keeps misunderstanding users, it is tempting to start debugging the speech recognizer, the language model, or the network.
Sometimes the problem is earlier.
Before ASR processes a word, voice activity detection (VAD) has already decided whether the incoming audio is worth sending downstream.
A good VAD configuration makes the pipeline feel responsive. A bad one can send keyboard noise into transcription, clip the first word of an utterance, delay turn-taking, or make an agent respond to its own audio.
That makes VAD less of a preprocessing checkbox and more of a load-bearing part of real-time voice architecture.
The broader Smallest AI stack covers real-time speech models and voice applications, but this article focuses specifically on the boundary before transcription: how to decide when speech begins, when it stops, and how aggressively the pipeline should react.
What VAD actually controls
At its simplest, VAD repeatedly answers one question:
Is this frame speech or non-speech?
Most implementations process audio in short windows, commonly around 10 to 30 milliseconds.
Those frame-level decisions then control the rest of the system:
- Start sending audio to ASR.
- Continue an active speech segment.
- Hold during a brief pause.
- Flush buffered audio.
- Stop the stream or begin endpointing logic.
In VoIP systems, VAD can avoid sending long stretches of silence.
In transcription systems, it determines which audio reaches ASR.
In conversational voice agents, VAD becomes one of the signals used to decide whether the user has started or stopped talking.
The difficult part is that production audio rarely looks like a clean speech dataset.
Your microphone may also capture HVAC noise, typing, traffic, music, another speaker, breathing, lip noise, or audio from the agent itself.
VAD has to make its decision immediately, without knowing what the next few hundred milliseconds will contain.
Frame size is part of your latency budget
Frame size is one of the first VAD parameters worth examining.
Short frames, such as 10 ms windows, let the system detect speech onset quickly. That can reduce the delay before ASR begins receiving useful audio.
The trade-off is context.
With less audio inside each decision window, short frames may be easier to misclassify in difficult acoustic environments.
Longer frames, such as 30 ms windows, contain more information and can make classification more stable. But they also delay the first speech decision.
Thirty milliseconds does not sound significant in isolation.
It becomes significant when you add it to:
- Audio capture and buffering
- Network transport
- Speech recognition
- Endpoint detection
- LLM inference
- Text-to-speech generation
- Playback buffering
Voice engineering guidance has traditionally treated roughly 150 ms of one-way delay as an important quality boundary for highly interactive communication, while normal human turn-taking can occur on the order of a few hundred milliseconds.
In that environment, several small delays can consume a meaningful percentage of the entire interaction budget.
That does not mean every system should use 10 ms frames.
A controlled call-center deployment with standardized headsets has a very different noise profile from a mobile application being used in kitchens, cars, cafés, and sidewalks.
Choose frame size against the audio your application actually receives.
Classical VAD vs. neural VAD
The right VAD architecture depends heavily on compute constraints and acoustic conditions.
Classical VAD
Classical detectors remain useful because they are inexpensive and predictable.
The WebRTC VAD implementation, for example, uses Gaussian Mixture Models to compare speech and background-noise probabilities from audio features. Its implementation supports frame durations such as 10, 20, and 30 ms.
You can inspect the WebRTC VAD implementation directly if you want to understand the statistical decision path.
This kind of detector makes sense when you need:
- Low compute overhead
- On-device execution
- Predictable processing time
- Controlled acoustic environments
- No dependency on GPU inference
Neural VAD
Neural VAD usually uses a compact learned model rather than relying only on hand-designed spectral rules.
That can make it more robust when background noise changes over time.
Music, television, overlapping speakers, traffic, and other non-stationary noise can be particularly difficult for simpler statistical detectors.
The trade-off is inference cost.
For server-side pipelines that already have sufficient compute, that cost may be acceptable. On constrained devices or latency-sensitive edge deployments, it may not be.
Hybrid VAD
You do not necessarily have to choose one detector for every frame.
A production system can use a lightweight detector as the first gate and invoke a more expensive model only for uncertain or difficult segments.
The objective is not architectural purity.
The objective is to avoid spending expensive inference on obvious silence while still handling noisy edge cases reliably.
VAD and endpointing solve different problems
A common implementation mistake is treating VAD and endpointing as interchangeable.
They are not.
VAD answers:
Is the user producing speech right now?
Endpointing answers:
Has the user finished their turn?
Those are different questions.
Consider this sentence:
"Can you book a meeting with... Sarah tomorrow?"
A speaker may naturally pause after "with" while remembering the name.
During that pause, VAD may correctly classify several frames as non-speech.
That does not mean the conversational turn has ended.
A basic endpointing system might use a rule such as:
800 ms of silence → end of turn
That is easy to implement, but it can cause interruptions when users pause to think.
More sophisticated endpointing systems can combine VAD state with additional information such as partial transcription, semantic completeness, or dedicated turn-taking models.
VAD remains a low-level signal.
Endpointing turns that signal into a conversational decision.
If you are budgeting latency across the entire conversational pipeline, the distinction matters. Smallest AI's guide to STT, LLM, TTS, tools, and latency budgeting looks at how those delays accumulate across the broader voice-agent stack.
False triggers are usually where production VAD hurts
VAD errors broadly fall into two categories.
False positives
A false positive happens when non-speech audio is classified as speech.
That can push unnecessary audio into ASR.
Sometimes the result is harmless gibberish.
A more dangerous failure happens when the recognizer produces plausible text from noise. Downstream components may then treat something that never happened as a real user utterance.
For a conversational agent, a cough, television voice, keyboard sound, or door slam should not become an actionable request.
False negatives
False negatives happen when actual speech is classified as non-speech.
The most noticeable version is a clipped speech onset.
If VAD opens the gate too late, the downstream recognizer may receive:
...eed to change my booking
instead of:
I need to change my booking
Users notice this immediately.
They start repeating words, speaking unnaturally, or assuming the application is not listening.
In production traffic, several inputs frequently cause trouble:
- Non-stationary noise: music, television, traffic, and overlapping speakers.
- Breath and lip sounds: these can resemble speech-like acoustic events.
- Hesitation sounds: "um," "uh," and other short, quiet speech can disappear when onset thresholds are too conservative.
- Playback leakage: an agent's own TTS may reach the microphone and trigger the detector.
- Threshold-edge audio: frames hovering near the activation threshold can cause rapid speech/non-speech switching.
The solution is not simply "increase the threshold."
Every adjustment changes which class of errors you are accepting.
Practical VAD tuning in a real pipeline
There is no universal production threshold.
Tune against the failure distribution of your application.
| Problem | Likely cause | What to adjust |
|---|---|---|
| First word gets clipped | Activation threshold too high or no leading buffer | Lower the threshold and add pre-roll |
| Agent responds too late | Trailing padding or endpoint window is too long | Reduce the trailing-silence window |
| Background noise triggers ASR | Threshold too low or weak upstream cleanup | Increase the threshold or improve noise suppression |
| VAD rapidly flips states | Signal is hovering near the boundary | Add hysteresis smoothing |
| Agent detects its own TTS | Playback is leaking into the microphone | Add echo cancellation and playback-specific VAD behavior |
Log the decisions, not just the transcripts
If possible, capture VAD state alongside a representative sample of audio.
You want to know:
- Which frames triggered speech?
- What did the audio actually contain?
- How often did speech onset get clipped?
- Which noise types produced false positives?
- How long did the system wait before closing an utterance?
- Were failures concentrated on specific devices or environments?
Synthetic noise tests are useful for regression.
They are not a substitute for observing the acoustic environments your users actually create.
Tune activation threshold and padding together
Two settings tend to dominate day-to-day tuning:
Activation threshold controls how confident the detector must be before a frame becomes speech.
Lowering it improves sensitivity but usually increases false positives.
Padding controls how much audio remains part of the active segment around speech boundaries.
On the trailing edge, a practical conversational starting point is often around 200–300 ms.
That is long enough to absorb many natural micro-pauses without leaving the pipeline open indefinitely.
It is a baseline, not a universal default.
Your application still needs measurement.
On the leading edge, sensitivity often deserves extra weight because a small pre-roll buffer is cheaper than losing the beginning of a user's sentence.
Why hysteresis matters
Suppose your detector produces confidence values close to a threshold:
0.48, 0.52, 0.49, 0.53, 0.47
With a hard threshold at 0.50, the VAD state may repeatedly switch:
off → on → off → on → off
That state chatter is difficult for downstream components.
Hysteresis introduces stability.
Instead of using exactly the same transition rule in both directions, require sustained evidence before changing states.
For example, entering speech may require several qualifying frames, while leaving speech may require a longer sequence below the deactivation boundary.
The exact values depend on your detector, but the principle is broadly useful:
A frame-level classifier does not have to become a frame-level state transition.
Barge-in changes the operating conditions
Voice agents introduce a problem that pure transcription systems often avoid.
The system may be speaking while the user starts talking.
Now the microphone contains both:
- The user's interruption
- The agent's synthesized voice
This is the barge-in problem.
Acoustic echo cancellation is one of the first defenses because it reduces playback leakage before VAD evaluates the microphone signal.
But echo cancellation alone does not eliminate the tuning problem.
Your acoustic conditions during playback are different from your acoustic conditions during silence.
That means the same VAD sensitivity may not be optimal in both states.
If the playback-mode threshold is too high, real interruptions get missed.
If it is too low, the system detects its own TTS and interrupts itself.
Treat playback and non-playback as distinct operating modes.
Multi-microphone systems can improve the input before VAD
When multiple microphone channels are available, spatial processing can improve what the detector sees.
Beamforming attempts to emphasize sound arriving from a target direction while suppressing other sources.
That improves signal-to-noise ratio before classification.
A cleaner input can reduce both false positives and false negatives without changing the VAD model itself.
This illustrates a broader point:
Not every VAD problem should be solved inside VAD.
Sometimes the correct fix is upstream.
Noise suppression, acoustic echo cancellation, beamforming, microphone placement, gain control, and device-specific audio processing can all change the detector's error rate.
Where VAD belongs in the voice stack
A simplified real-time speech pipeline might look like this:
microphone → echo cancellation → noise suppression → VAD → ASR → endpointing/application logic
The exact order depends on the system, but placement matters.
Running noise suppression before VAD gives the detector a cleaner signal.
Running VAD directly on raw audio means its threshold must tolerate every acoustic artifact that reaches the microphone.
The same rule applies downstream.
If VAD discards a frame containing speech, ASR cannot reconstruct audio it never received.
If VAD sends noise into ASR, the recognizer has to decide what to do with audio that should have been filtered earlier.
For a hosted downstream STT layer, Pulse speech-to-text is Smallest AI's product for real-time and recorded transcription, including live-audio and voice-agent workloads.
That does not make the STT model a replacement for your VAD design.
It makes the VAD boundary easier to reason about: your recognizer can only process the audio your preprocessing layer decides to pass downstream.
Testing VAD against a real speech pipeline
A useful VAD evaluation should not stop at frame-level accuracy.
Measure what happens to the rest of the application.
For example:
- Run representative audio through your VAD configuration.
- Preserve the accepted audio exactly as the recognizer would receive it.
- Send that audio through your STT layer.
- Compare transcripts across threshold and padding configurations.
- Measure speech-onset clipping.
- Measure false ASR activations caused by noise.
- Measure how VAD and trailing padding affect end-to-end response time.
This tells you something a standalone VAD score cannot:
whether the detector's mistakes actually damage the product.
Developers who want to test the downstream transcription side can use the Smallest AI API as the speech layer in this type of evaluation.
The VAD itself can remain in your client, media server, or preprocessing service. The API then gives you a consistent downstream component against which you can evaluate how different gating decisions affect transcription.
Create and store the API key
Keep the API key in an environment variable rather than hard-coding it into the application.
Before running the snippet, create a Smallest.ai API key in the dashboard and store it in the SMALLEST_API_KEY environment variable.
export SMALLEST_API_KEY="your-api-key-here"
Every authenticated request sends the value through the Authorization header:
Authorization: Bearer <SMALLEST_API_KEY value>
Keep the key on your server.
Do not expose it in browser JavaScript, client-side React code, mobile application code, public repositories, screenshots, query parameters, or client-side logs.
For production deployments, store it in your hosting environment's server-side secrets manager rather than committing credentials to configuration files.
This article intentionally does not invent an API request specifically for VAD because the VAD architecture described here is an upstream pipeline concern rather than a Smallest AI-specific VAD endpoint.
What to measure before shipping
If VAD is going into production, evaluate it as a system component rather than a classifier in isolation.
Useful measurements include:
- Speech-start detection delay
- Speech-end detection delay
- False-positive rate by noise category
- False-negative rate at utterance onset
- Percentage of utterances with clipped first words
- Number of unnecessary ASR activations
- Endpointing delay after actual user completion
- Barge-in success during TTS playback
- Behavior across microphones, devices, and environments
The correct configuration depends on the cost of each failure.
A dictation application may tolerate a slightly slower onset if it reduces false activations.
A voice agent may prefer a more sensitive leading edge because clipped words damage conversational flow immediately.
Production tuning is about choosing those trade-offs deliberately.
Key takeaways
- VAD is an upstream latency and quality control point, not just a silence detector.
- Frame size affects how quickly speech can be detected and how much acoustic context the classifier receives.
- False positives waste downstream work; false negatives can remove speech permanently.
- VAD and endpointing solve different problems.
- Trailing padding directly affects how quickly a conversational system can respond.
- Around 200–300 ms of trailing padding is a reasonable starting point for many conversational systems, but it should be validated against real traffic.
- Hysteresis can stabilize frame-level decisions without replacing the underlying detector.
- Barge-in requires echo handling and often different VAD behavior during playback.
- Beamforming and noise suppression can improve VAD performance before you touch the detector itself.
- The right threshold is the one that minimizes the errors that matter most to your application.
Conclusion
Voice activity detection is easy to describe because the output looks binary.
Production behavior is not.
Every threshold, frame size, pre-roll buffer, silence window, and state-transition rule changes what the rest of your voice pipeline receives and when it receives it.
That is why VAD tuning should be evaluated alongside ASR and endpointing rather than as an isolated preprocessing benchmark.
Measure with real audio. Log the boundaries. Inspect false triggers. Watch for clipped onsets. Test again while TTS is playing.
Then optimize the failure mode that actually damages your application.
If you want to evaluate how those VAD decisions affect downstream transcription, create an API key and test the pipeline with your own audio using Smallest AI.
Top comments (0)