Question 1

What is VAD and why does it matter for voice agents?

Accepted Answer

VAD (Voice Activity Detection) is the element of voice agent pipeline that decides whether a chunk of audio contains speech or not. It sits ahead of turn-taking and speech-to-text, so if it misses a word, the agent never gets a chance to hear it, no matter how good the rest of the pipeline is.

Question 2

Why does my voice agent miss what people say?

Accepted Answer

Usually the problem traces back to the VAD, not the ASR or LLM getting it wrong. Conventional VADs are tuned on clean audio, so they can miss speech in engine noise, phone compression, or echo, and once that happens, the words never even reach the transcription step.

Question 3

Why does a VAD perform well in testing but fail on real calls?

Accepted Answer

Most VADs are evaluated on clean speech, where nearly every model performs well. Production conditions, like a drive-thru with the engine running, a bad phone connection, or a crowded room with overlapping speakers, are a different problem, and a VAD that isn't specifically trained against them will start missing speech under exactly those conditions.

Question 4

What is the difference between VAD and speech enhancement?

Accepted Answer

A VAD only classifies whether speech is present in a chunk of audio. Speech enhancement changes the audio itself. Some teams prefer a VAD-only signal because it doesn't alter the audio, while others need enhancement to fix audio quality before it reaches ASR, and many voice-agent pipelines use both.

Question 5

Can I just use Silero VAD instead of a purpose-built VAD?

Accepted Answer

You can, but Silero wasn't built for the conditions voice agents actually run into. ai-coustics' VAD Multi Speaker is trained specifically on messy, real-world audio, so it keeps detecting speech where Silero starts dropping it: on a drive-thru dataset, VAD Multi Speaker caught over 83% of speech against Silero's 48%, a gap that only shows up once the audio stops being clean.

Question 6

What is VAD Multi Speaker and how is it different from Silero VAD?

Accepted Answer

VAD Multi Speaker is ai-coustics' streaming voice-activity-detection model, built as a drop-in replacement for Silero VAD that runs without PyTorch or ONNX runtime. It's trained specifically to keep detecting speech in noisy, degraded audio, where standard VADs like Silero start missing speech, while matching Silero's performance on clean audio.

Question 7

What's the difference between VAD Multi Speaker and VAD Voice Focus?

Accepted Answer

VAD Multi Speaker is the general-purpose model: it triggers on any speech in the audio and is the right choice when every voice in the stream matters. VAD Voice Focus only triggers on the primary speaker and ignores background voices, built for one-to-one interactions like phone bots where secondary speech should be ignored rather than flagged.

Question 8

How much does VAD Multi Speaker improve speech detection?

Accepted Answer

On a real-world drive-thru dataset, VAD Multi Speaker classified 82% of the audio correctly versus 67% for Silero VAD, with most of the gap coming from missed speech: it detected over 83% of all speech in that dataset compared to 48% for Silero. On clean phone-call audio, all three perform comparably, since that's a case general-purpose VADs already handle well.

Question 9

How much latency does ai-coustics VAD add to my pipeline?

Accepted Answer

VAD Multi Speaker and VAD Voice Focus run in real time on CPU with roughly 30ms of latency, through ai-coustics' AirTen runtime rather than PyTorch or ONNX. That keeps it lightweight enough to serve as a direct swap for Silero VAD without adding a separate inference stack.

Question 10

Can I tune how sensitive ai-coustics VAD is?

Accepted Answer

Yes. Our VAD (Multi Speaker and Voice Focus) exposes three tunable parameters (sensitivity, speech hold duration, and minimum speech duration) so you can adjust when it triggers and releases to match your own turn-taking logic. If you want full control, the SDK also exposes the raw per-frame speech probabilities so you can build your own post-processing.

Real-time voice activity detection for Voice AI

Real-time voice activity detection for Voice AI

Fewer false triggers

Reliable turn detection

Real-time on CPU

VAD Multi Speaker

Noise-robust voice activity detection

VAD Voice Focus

Voice activity detection for the primary speaker

One SDK. Integrated in minutes.

Built for your stack

Try for free now in our Developer Platform

Try for free now in our Developer Platform

Try for free now in our Developer Platform

Your questions, answered

Bring real-time audio intelligence into your voice AI stack

Bring real-time audio intelligence into your voice AI stack

Bring real-time audio intelligence into your voice AI stack

Quail

Sparrow