MEMORY NOTCH BLOG

Silero VAD vs WebRTC VAD, and the end of speech

Silero VAD catches far more speech in noise than WebRTC VAD. Neither can say when a sentence is over, and a dictation tool has to solve that on top.

Published October 23, 2026 · The Memory Notch team

Silero VAD catches far more speech in noisy audio; WebRTC VAD is smaller, older and faster. Both are voice activity detector (VAD) models: Silero is a small neural network, and WebRTC's is a statistical detector from Google's browser audio code. Silero shipped v6.2.3 on September 23, 2026. Both answer "is someone speaking right now?" Neither answers "has this person finished?", which dictation has to get right.

	Silero VAD	WebRTC VAD	TEN VAD
Made by	Silero (open source)	Google, part of WebRTC	Agora's TEN project
How it decides	Neural network	Two-Gaussian models over six frequency bands	Neural network
Size	About 2 megabytes	A few C files, no model file	277 to 731 kilobytes (vendor)
Sample rates	8 and 16 kHz	8, 16, 32 and 48 kHz	16 kHz
Frame	32 ms at 16 kHz	10, 20 or 30 ms	10 or 16 ms
Latest release	v6.2.3, September 23, 2026	Part of WebRTC; the common Python binding is py-webrtcvad	v1.0, July 11, 2025
Licence	MIT	BSD-style (WebRTC's licence)	Apache 2.0 with added conditions

Checked September 29, 2026.

What changed this week

Silero v6.2.3, released September 23, makes torchaudio optional. The model itself never used that library; only the helpers that read and write audio files did. Now PyTorch is the one requirement, and if you load audio yourself, you install less.

Six days earlier, v6.2.2 added a separate "sequence" model for whole files at 16 kHz. Neither release changes accuracy. The last quality change was v6.2 in November 2025, which the project says improved child voices, muted speech and "lower quality phone calls".

What does a voice activity detector do?

A voice activity detector labels each short slice of audio as speech or not speech. Apps run one before a speech model, so the model only works on audio with voices in it, and so the app knows when someone starts and stops.

The two detectors get there in different ways. WebRTC's detector takes frames of 10, 20 or 30 milliseconds. It splits each one into six frequency bands, from the deepest voices up to 4 kHz, and measures the energy in each band. Then it compares two statistical models, one for speech and one for noise. The source code calls them Gaussian Mixture Models, and they adapt as the audio changes. A setting from 0 to 3 makes it more or less willing to call something noise.

Silero is a neural network trained on audio in over 6,000 languages. For each 32-millisecond slice, it returns a probability that someone is speaking, and you pick the cut-off. The project says one slice takes under a millisecond on one processor thread.

Band energy against a trained network explains most of what follows. Energy in speech bands is a good sign of a voice in a quiet room. But a dishwasher, a keyboard or a café also puts energy in those bands.

Silero vs WebRTC VAD: which is more accurate in noise?

Silero catches much more speech than WebRTC's detector in noisy audio, in both benchmarks we found. Both come from companies with a stake in the result, so read each as that vendor's own number.

Benchmark	Who ran it	Silero	WebRTC VAD	Their own detector
Area under the curve (AUC), 17 hours of meetings, calls and noisy speech	Silero (its wiki)	0.97 (v6)	0.73	not applicable
Accuracy on 31.25 ms slices of speech sets	Silero (its wiki)	0.92 (v6)	0.74	not applicable
Share of speech caught when 5 % of noise is mislabelled; clean speech mixed with noise at equal loudness	Picovoice (maker of Cobra)	about 88 % (v5.1)	about 50 %	about 99 % (Cobra)

Checked September 29, 2026. Picovoice's figures are read from its published chart, not from a table.

The area under the curve (AUC) scores a detector across every possible cut-off: 0.5 is a coin toss and 1.0 is never wrong. Silero's wiki sums up the gap in its own words. WebRTC's detector is "extremely fast and pretty good at separating noise from silence, but pretty poor at separating speech from noise".

Picovoice's benchmark mixes clean read speech with 18 kinds of recorded noise, from kitchens to traffic, as loud as the voice. Picovoice tested Silero 5.1, not v6.

At a strict setting, where 1 frame of noise in 20 gets called speech, WebRTC catches about half the speech frames and Silero about 88 %. The same chart has a twist. Once you accept far more false alarms, around a quarter of noise frames or more, WebRTC catches slightly more speech than Silero. That makes it a fair pick when a missed word costs more than a false alarm.

WebRTC's strengths are real. It is a few files of C with no model to load, it runs on almost anything, and it takes 48 kHz audio without resampling. In a quiet room with a close microphone, it does the job.

Silero vs TEN VAD: is the newer one better?

TEN VAD, from Agora's open-source TEN framework, says it beats both Silero and WebRTC on precision and reacts faster when speech stops. Every one of those claims comes from its own README.

The project says Silero "suffers from a delay of several hundred milliseconds" at the end of speech. It reports a library of 277 to 731 kilobytes, against about 2.2 megabytes for Silero, at similar speed. It takes 16 kHz audio only. It lists macOS, Linux, Windows, Android, iOS and the web.

Read the licence before you ship it. TEN uses Apache 2.0 "with additional conditions". Among them, you may not deploy it "in a way that competes with Agora's offerings". Silero uses the plain MIT licence, with no such clause. For many products, the licence decides the question before accuracy does.

If you need	Pick
The fewest missed words in noise, with a permissive licence	Silero
The smallest footprint, clean audio, or 32 or 48 kHz input	WebRTC VAD
The fastest end-of-speech reaction, and the licence fits your product	TEN VAD, after testing it on your own audio
The highest published accuracy, and a paid licence is fine	Picovoice Cobra (Picovoice's own numbers)

How do you tell a pause from the end of a sentence?

Wait for a stretch of silence longer than a normal pause. A detector only reports speech and non-speech. How long counts as "done" is a rule you add on top, and the right length depends on the job.

Silero's own helper shows the idea at a small scale. When it cuts a file into speech segments, it closes a segment after 100 milliseconds of silence. That suits cutting files. It is far too short for dictation, where people stop to think mid-sentence.

For dictation, the silence has to be long enough that a thinking pause doesn't paste half a sentence, and short enough that you don't sit waiting. Three other problems sit on top of that:

Words arrive late. A streaming speech model hears audio in chunks, so a word shows up to a chunk after it was spoken. Start the silence clock when the last word appeared, and you add that lag to every phrase. Trust only the text, and a pause looks the same as the end.
The detector can miss quiet speech. A far-off or quiet microphone can make it under-fire while the speech model still hears words.
Someone may never speak, or may talk for a very long time.

Memory Notch's dictation handles all three with one clock. You double-tap ⌘ and speak. Once you have been silent for 2 seconds, the text is pasted into the app you are in.

The clock doesn't start when a word appeared on screen. It starts from the later of two moments: the last time the speech detector heard speech, and the point in the audio where the newest recognised word ended. A word never ends after the speech it came from, so counting it adds no delay when the detector works. When the detector misses quiet speech, the word timings stop the clock from running out while you are still talking.

Two limits close the loop. If no words come within 5 seconds, dictation gives up. A phrase is capped at 120 seconds, and whatever has been heard by then is pasted.

Do you still need a VAD with a streaming speech model?

In most apps, yes. A streaming speech model tells you what was said, a chunk late. A detector tells you, within a few tens of milliseconds, whether someone is still talking. Ending a turn well needs both.

Take the model behind Memory Notch's live transcript. Its words come in 1.12-second chunks, the smallest size its porting library recommends when punctuation matters (why streaming models work in chunks). A silence rule built only on the text would react to a pause about a second late. It could also mistake a slow word for silence.

Some models build end-of-turn detection into the recogniser. NVIDIA's Parakeet Realtime end-of-utterance model emits a token when a speaker finishes, with a median delay of 160 milliseconds in NVIDIA's tests. It supports English only and doesn't punctuate, because it is built for voice agents that must answer quickly. Dictation and meeting notes need many languages and punctuated text. For them, a cheap detector plus word timings is still the practical design.

For the whole path from microphone to transcript, see how AI transcription works.

What we didn't test

We have not benchmarked Silero, WebRTC or TEN ourselves. Every accuracy number above is a vendor's, run on that vendor's choice of audio. None of those tests used laptop microphones in meetings, which is where a Mac dictation tool lives.

The 2-second rule also has a known edge. Say you pause for about a second, then carry on so quietly that the detector misses it. The words for that speech can arrive a chunk late, and the text may be pasted just before they do. Closing that gap would mean waiting up to a chunk on every phrase, which would make every dictation slower.

Where Memory Notch fits

Dictation is free in Memory Notch for Mac from version 0.5.7. It needs the Accessibility permission, which the privacy page explains, and it works in any app that takes text. Speech recognition runs on the Mac. The app makes three network requests, and none carries a recording: a revoked-licence list, a daily update check you can turn off, and an update download when you choose one.

The main job is recording meetings. The app transcribes your microphone and the call's audio as they happen, with numbered speakers, and keeps plain files on your Mac. There are no automatic summaries, and overlapping voices can be difficult to separate.

Memory Notch needs an Apple Silicon Mac on macOS 26 or later and costs $79 once (prices are in US dollars). You can try it free with recordings of up to 5 minutes. Recording other people still needs their consent.

Frequently asked questions

Is Silero VAD better than WebRTC VAD?

In noisy audio, yes, by both Silero's and Picovoice's benchmarks. Silero reports an overall score of 0.97 against 0.73 for WebRTC's detector. Picovoice measured Silero catching about 88 % of speech at a strict setting, against about half for WebRTC. WebRTC's detector is smaller, needs no model file, and is fine for clean, close-miked audio.

Is WebRTC VAD a neural network?

No. It measures energy in six frequency bands and compares Gaussian mixture models of speech and noise, which adapt as the audio changes. That makes it very fast and tiny, but it confuses loud, speech-like noise with speech more often than neural detectors like Silero.

What is the latest Silero VAD version?

Version 6.2.3, released September 23, 2026. It makes torchaudio optional, so only PyTorch is required. v6.2.2, on September 17, added a separate whole-file "sequence" model for 16 kHz audio. The last accuracy update was v6.2 in November 2025.

How long should the silence be before dictation stops listening?

Long enough that a thinking pause doesn't cut you off: seconds, not milliseconds. Memory Notch's dictation pastes after 2 seconds of silence, gives up after 5 seconds with no words, and caps a phrase at 120 seconds. File-cutting tools use far shorter gaps, such as Silero's 100-millisecond default.

Can TEN VAD be used in a commercial app?

Yes, within its licence. TEN uses Apache 2.0 with added conditions from Agora, including that you may not deploy it in a way that competes with Agora's own offerings. Read the licence file before shipping. Silero uses the MIT licence, which has no such condition.

Sources

Download Memory Notch · More from the blog