Almost every microphone on this list is a thin client. Audio leaves the body, a model in someone else’s rack turns it into text, another model turns that into a summary, and a file comes back if you are lucky. That architecture is why Humane could brick a pin, why Plaud can improve templates without a hardware revision, and why “offline” is a rare word in this aisle.
Pebble’s Index 01 is the structural opposite, and it is now a shipping opposite. Speech recognition and the language model both run on the paired phone. Airplane mode works. The companion app’s source is published (Pebble). The cost of that choice is the product: 5–20 cm pickup, five-minute buffer, no room capture. On-device ASR here is not a Plaud feature waiting to be enabled. It is a dictation ring.
Hybrid is a real third path
Omi documents on-phone transcription on the free path, a cloud plan, or a speech provider you bring (Omi subscriptions). The hardware is still always-on and still 10–14 hours. The option to keep Based Hardware off the audio is the privacy score, not the orb’s looks. Most owners will take the cloud. The documentation still matters, because it is the only vendor cloud you are allowed to decline without declining the device.
Everyone else is cloud as the product. Plaud’s templates are the most mature pipeline on the list and they do not run in the card. Bee transcribes in real time and discards audio — there is nothing to process later, locally or not. Vocci’s 100-plus languages and speaker labels are third-party clouds with an export story (Vocci privacy). Limitless’s lifelog was a cloud object with a good API, then a Meta object, then a regional deletion calendar.
Latency is not the axis
People ask about on-device as if it were speed. Sometimes it is. The scored question on methodology is closer to: who has the bits, and can you leave. Intelligence still measures whether the transcript and summary are usable. A local model that cannot summarise a meeting should not win intelligence because it is local. It should win privacy.
| Cost | |
|---|---|
| Price | $75 |
| Plan | None required |
| Capture | |
| Microphones | 1 |
| Capture mode | Press to record |
| Audio out | None |
| Battery (active capture) | 13 h — Not a charge cycle. Pebble quotes 12–15 hours of recording across the ring's entire life, after which the sealed battery is spent and the ring is replaced. |
| On-device storage | 0.083 h of audio |
| Speech recognition | On-device |
| Language model | On-device |
| Body | |
| Form factor | Ring |
| Weight | 4.7 g |
| Materials | Stainless steel |
| Charging case | Not published |
| Recording indicator | No |
| Connectivity | Bluetooth |
| Companion app | iOS, Android, Web |
| Data export | Yes |
| Cost | |
|---|---|
| Price | $179 |
| Plan | On-phone free, or $19/mo |
| Year-one on the encoded plan | $407 if you pay for twelve months of the encoded plan. Not the ranking column. |
| Capture | |
| Microphones | 2 |
| Capture mode | Continuous capture |
| Audio out | None |
| Battery (active capture) | 10 h — Based Hardware quotes 10–14 hours from a 150 mAh cell; the lower bound is used here. Ignore the three-day figure from its CES launch. |
| Speech recognition | Hybrid |
| Language model | Cloud |
| Body | |
| Form factor | Pendant |
| Weight | Not published |
| Materials | Not published |
| Charging case | Not published |
| Recording indicator | Yes |
| Connectivity | Bluetooth, Wi-Fi |
| Companion app | iOS, Android, Web |
| Data export | Yes |
A ranking that treated cloud ASR as a moral failure would put Pebble first and ignore that it cannot hear a table. A ranking that ignored processing location would put Plaud first and ignore that every word takes a trip. The two axes exist so neither sentence has to do both jobs. Buy a recorder and accept the rack, or buy a dictation ring and accept the range. The dishonest pitch is a far-field wearable that also never leaves the device. Nobody here is shipping that.