Over the past two years, AI hardware has become a graveyard of ambition. We've seen credit-card-sized gizmos, clip-on pins, smart pendants, and recording earbuds—almost all chasing the exact same premise: record your meetings, generate automated summaries, and manage your to-do lists.
<div class="aice-quick-answer" style="background:#f0f7ff;border-left:4px solid #2563eb;padding:14
At Sandbar, we took a fundamentally different approach with our voice ring, Stream. We recently closed a $23 million Series A led by Adjacent and Kindred Ventures, bringing our total funding to $36 million. But the most important part of our story isn't the capital—it's why we deliberately stripped away the features everyone else thinks are mandatory.
The Real Reason Voice Hardware Keeps Failing
Most voice-first hardware fails because it treats the user as an obstacle in an automated pipeline. The default paradigm forces you to say a wake word, speak, and wait for a response. The entire chain is dictated by the AI, demoting the human to a mere trigger.
Stream flips that script. There is no wake word. It doesn't stream audio to the cloud in real time. Everything is stored locally on the ring until you decide what to do with it. You double-tap the side to capture a thought, and you manually choose which recordings get transcribed and synced to your phone or computer.
There are no auto-generated summaries, no intrusive to-do lists, and no preset assumptions about how you should work. In fact, we deliberately killed three "standard" industry features:
1. No voice activation / wake words
2. No real-time live captioning
3. No on-device AI inference silicon
Instead, Stream uses low-power Bluetooth to offload transcription to your phone. The ring itself does only two things exceptionally well: capture audio and store it. Because we ditched the heavy compute, we achieved a 7-day battery life in a form factor that is just 13.5mm in diameter, 4.2mm thick, and weighs 8.3 grams.
What 172 Beta Testers Taught Us About Real Demand
We gave Stream to 172 users for three consecutive months. The data surprised even us:
- 91% of users actively tapped it more than 3 times a day.
- The average recording length was 47 seconds.
- 0% of users enabled auto-sync—every single person preferred manual export.
- 83% used it exclusively for personal memos and fleeting thoughts, not meeting notes.
- 62% used it in screenless contexts: walking the dog, in the shower, or while driving.
This revealed a clear truth: people don't want an AI agent hijacking their workflow. They want a frictionless safety net for their thoughts—capturing a sudden sentence, a fleeting name, or an unanswered question before it disappears.
Stream doesn't make decisions; it acts as a container. It doesn't replace human memory; it just gives it a physical buffer. By demoting AI from an overbearing orchestrator to an on-demand utility, control stays entirely at your fingertips.
Human creativity doesn't follow an SOP. Breakthrough ideas happen when your hands are full, when you have zero bars of signal, or when you're away from a screen. We aren't trying to make your meetings 10% more productive. We're just making sure you never lose the thought you had 30 seconds ago.
Production Timeline and Yields
While we haven't announced retail pricing yet, mass production for Stream is scheduled for Q4 2026. Micro-microphone array packaging is notoriously difficult, with industry average yields hovering around 72%. Our initial pilot runs have already hit an 87% yield rate.
Building hardware is unforgiving, but the path forward isn't adding more AI bloat—it's knowing exactly what to leave out.
