Stop paying to transcribe silence
Cloud speech-to-text bills by the audio minute — and a surprising share of most recordings and calls is not speech at all. Running VoxRT's voice activity detection before audio reaches the meter means you ship, store, and pay for the speech, not the air around it.
Why teams put VAD in front of the meter
Pay for speech, not air
Hold time, thinking pauses, keyboard clatter, dead air — every non-speech minute billed at speech prices. A VAD gate removes it before it becomes a line item.
Same transcripts, smaller pipeline
The words your STT provider returns come from the speech segments anyway. Trimming what isn't speech shrinks transfer, storage, and processing without touching transcript content.
A fix that scales with your growth
The VAD integration is one fixed piece of engineering with no metering of its own — so the savings grow with your traffic instead of the bill.
Where the silence hides
The higher your audio volume, the more of the bill is quiet.
Contact centers & call analyticsHold music off the invoice
Queues, holds, and after-call silence trimmed at ingress — only the conversation reaches the transcription vendor.
Meeting recordingLong rooms, short transcripts
Multi-hour recordings where minutes of quiet between exchanges were being transcribed at full price.
Voice notes & memosUpload only the speech spans
Consumer recorders that clip silence on the device — faster uploads on bad connections, smaller bills behind them.
Voice agentsGate the streaming STT
Real-time agents where VAD already drives turn-taking — the same gate keeps silence out of streaming transcription.
Compliance recordingArchive everything, transcribe speech
Keep the full original audio for the regulator; send only detected speech to the costly transcription stage.
Media & podcast toolingPre-segmentation for batch jobs
Speech spans with timestamps make batch STT jobs shorter and downstream editing cleaner.
Putting a gate in front of the bill
Insert VAD before the metered hop
On the user's device before upload, or at your ingest servers before the STT call — wherever audio last belongs to you.
Segment audio into speech spans
The detector emits per-frame speech probabilities; your code (or the SDK's helpers) turns them into timestamped speech segments, with padding at the edges so words are never clipped.
Send only speech to the provider
Batch or streaming, the provider sees concatenated speech segments — the transcripts come back the same, minus the silence you didn't pay for.
Map results back to the timeline
Keep the segment timestamps and the transcript aligns back to the original recording for playback, search, and audit.
The numbers behind the promise
VAD throughput and footprint are measured and published — see the benchmark methodology and the VAD comparison.
Pipeline. Audio → VAD (per-frame speech probability) → timestamped speech segments → your STT provider. The detector is Silero v5 weights on the VoxRT CPU runtime; it runs on user devices (iOS/Android) or on your Linux servers, whichever side of the upload you want the savings to start.
Platforms. iOS 16+ (Swift Package Manager) and Android 8.0+/API 26 (JitPack/Gradle), plus Linux aarch64 for device-class hardware — validated on Raspberry Pi 3/4/5 and Pi Zero 2 W, NVIDIA Jetson Nano and Orin Nano, AWS Graviton and similar Cortex-A53/A55 boards — shipped as a single ~480 KB shared library with Python, Node.js, Go, C/C++ and Rust wrappers.
Licensing. No usage fees anywhere in the gate: the VAD SDK and model are free for commercial use on devices and servers alike — see licensing.
Cutting STT costs, answered
How much will we actually save?
It equals the share of non-speech in your audio, and that's measurable before you commit: run the VAD over a sample of real traffic and compare speech-span duration to total duration. Contact centers with hold time and meeting recordings tend to have the most to reclaim; dense dictation the least.
Will trimming silence hurt transcript quality?
The standard practice is to pad each speech segment by a fraction of a second on both sides so no word onset is clipped — the STT provider then sees the same speech it would have transcribed anyway, just without the empty stretches between.
Should the gate run on devices or on our servers?
Both work with the same SDK. On-device gating also saves upload bandwidth and keeps silence off your storage; server-side gating is a single deployment with ~54 streams per core and no client changes.
Could we skip the cloud STT bill entirely?
For many products, yes — VoxRT's streaming ASR transcribes on the device at 3.27% WER with no per-minute cost at all. The VAD gate is the incremental fix; on-device transcription is the structural one.