Every dictation tool we tried had the same problem: it wanted our audio. Not just for transcription — for training. For building voice models. For improving "product quality." The privacy policies were technically accurate but deliberately vague. The data went somewhere. We just couldn't tell where, or for how long, or what it was being used for.
For casual use, maybe that's an acceptable trade-off. But we weren't using these tools for casual purposes. We were dictating technical documents, security analysis, medical notes, legal memos. The kind of material where "your data may be used to improve our services" isn't a minor footnote — it's a dealbreaker.
So we built Verity.
The problem with voice tools today
Voice is intimate in a way that typed text isn't. When you speak, you leave behind more than words. Tone, cadence, vocabulary, the way you phrase things under pressure — these are all captured in an audio stream. Voice data is biometric data. Once someone has a model of how you speak, that model doesn't expire. It doesn't get deleted when you close your account.
The dominant players in dictation handle this in different ways, none fully satisfying. Some offer on-device processing but pair it with cloud products that have the same data concerns as everyone else. Others build detailed context profiles from your dictation patterns — useful for personalization, but requiring you to trust that profile to a third party. Platform-native dictation stays on-device for short phrases but routes longer dictation through the vendor's servers by default.
None of them tell you exactly which third-party systems touch your audio. None of them commit to zero retention by contract, with named processors. None of them publish their audio pipeline in enough detail that a security engineer could evaluate it.
That's the gap we're filling.
What Verity does differently
Verity's architecture is built around one constraint: your audio is not a resource to be mined. It's a signal to be processed and immediately discarded.
Here's what that means concretely:
- Zero retention by design. Audio buffers are released as soon as the transcription comes back. Not written to disk, not queued for later analysis, not batched for training runs. The audio is ephemeral — it exists only for the duration of the request.
- No AI vendor in the audio path. Audio is processed by our own STT pipeline, running open-weight models we host on infrastructure we provision, with redundancy and automatic failover. There is no third-party AI API in the path to trust or audit. The compute is rented from AWS, which supplies servers rather than software and never accesses your data independently of us. See our full privacy breakdown for the complete list.
- No voice model training. We do not build voice profiles. We do not use your audio to improve STT accuracy for other users. Each transcription is independent — there is no persistent model of how you speak.
- Encrypted text, not raw audio. Your transcript is encrypted with AES-256-GCM using a per-user key derived via HKDF-SHA256 before it touches our database. We cannot read your transcripts without your key.
STT processing runs on Verity-operated infrastructure. The models are open-weight releases we host ourselves, served from our own inference servers on compute we provision and control — running a published model is not the same as calling someone else's API. Your audio is never handed to a third-party AI provider; it is transcribed on our own servers and discarded immediately.
Why transparency matters here
Privacy claims are cheap. "We care about your privacy" appears in the marketing copy of products that quietly sell user data. We don't expect you to take our word for it.
So instead of asking for trust, we publish specifics. The privacy page names every processor in the audio path. This blog will document architectural decisions as we make them — including decisions that trade convenience for privacy, and why we made them that way. The technical architecture post goes layer by layer through exactly how audio moves from your microphone to text on your screen.
If you find something in our architecture that concerns you, we want to hear about it. The point of publishing this level of detail is to invite scrutiny, not to deflect it.
Who Verity is for
We built Verity for people who have reasons to care about where their words go:
- Developers and engineers dictating technical documentation or code comments
- Healthcare professionals dictating clinical notes under strict confidentiality requirements
- Lawyers and analysts working with confidential client material
- Anyone who dictates enough to have experienced the specific unease of wondering where their audio ended up
Verity is not trying to be the dictation tool for everyone. There are good options for users who prioritize personalization and convenience over privacy. We're the option for users who need to know exactly what happens to their audio, and for whom "trust us" is not a sufficient answer. Ready to try it? Our beginner's guide shows you how to get started.
What's next
We're exploring specialized domain support for technical vocabulary — method names, library names, jargon that general STT models handle poorly. All speech-to-text runs on Verity's own infrastructure under zero-retention processing. And we're working on per-app profiles — so the dictation context in your IDE is different from the context in your email client.
Privacy-first doesn't mean feature-poor. It means the features are built on a foundation that doesn't require you to give up something you can't get back.
Verity is free during launch — all features included, no credit card required. If you work with sensitive material and you're using a voice tool that can't tell you where your audio goes, it's worth trying something that can.