“Local voice AI” is a placement decision for each stage of the pipeline, not a single product feature. Voice activity detection can run on a recorder, ASR on a computer, search in a local index, and reasoning in a cloud model—or every stage can use a different boundary.

Architecture scope. This guide explains execution placement and trade-offs. It does not claim that local models are always more accurate, faster, cheaper, or safer than managed cloud systems.

Which voice-AI tasks can run in different places?

Task Recorder/device Computer Cloud
Voice activity detection Low-latency slicing, low power target More compute and tunability Requires streaming/upload
Audio enhancement Immediate but resource constrained Can use larger models Managed model and scaling
ASR Maximum autonomy, tight compute limit Offline-capable with local CPU/GPU/NPU Central models and elastic compute
Speaker labeling Usually constrained Local batch processing possible Managed pipeline and collaboration
Search/index Limited UI and storage model Open local files/database Cross-device workspace
LLM reasoning Rare on a small recorder Local model if hardware permits Large managed models and tools

Ask where each row runs and what bytes cross each arrow. A vendor can truthfully advertise “on-device VAD” while sending audio to the cloud for transcription. Another can run ASR locally but send the resulting text to a cloud LLM for summarization. Neither label alone describes the full data path.

What runs where in the documented C1 design?

C1 performs capture and on-device VAD, then stores encrypted audio slices temporarily. It connects to a paired computer over Bluetooth 5.4 or USB-C and has no Wi-Fi. The computer runs OneMira speech recognition, transcription, and speaker labeling, then writes local OKF Markdown/plain-text notes. Each slice is deleted from C1 after the computer consumes it for transcription.

Search, organization, CLI access, and MCP retrieval operate against local files and indexes. If the user asks a third-party AI agent to reason over selected text, the text follows that provider’s data path. This architecture is best described as device capture + computer-local ASR + user-selected reasoning.

Close model render of the OneMira C1 microphone-side LED and enclosure
The recorder handles capture and status; the heavier ASR workload belongs to the paired computer in the documented architecture.

Why put VAD on the recorder?

VAD can divide long capture into speech-relevant slices and reduce transfer/storage work. It also creates risk: a false negative can remove speech before a larger model sees it. Device VAD should therefore be conservative enough for the supported scenes and validated against quiet starts, pauses, keyboard noise, distant speech, multiple speakers, and non-speech transients.

Keep a diagnostic mode or labeled corpus that reveals boundary behavior. A high aggregate accuracy score can hide systematic clipping of the first word after silence.

Why put ASR on the computer instead of the recorder?

A laptop or desktop has more memory, sustained power, storage, and acceleration options than a 70 g recorder. It can host larger acoustic models and update them without replacing the capture device. The trade-off is hardware variability: macOS and Windows systems span different CPU generations, Apple silicon, integrated graphics, NVIDIA/AMD discrete GPUs, memory sizes, drivers, and thermal limits.

A correct runtime should detect supported acceleration and publish the selected execution provider. It also needs a tested fallback. GPU availability is not enough: the model runtime, operator set, driver, precision, and memory must all be compatible. Objective validation records real-time factor, peak memory, model version, execution provider, and output quality on a representative device matrix.

How much data does compressed speech require?

The documented OneMira audio design uses Opus, 16 kHz mono, at 16 kbps. The arithmetic is straightforward: 

16,000 bits/second × 3,600 seconds ÷ 8
= 7,200,000 bytes per hour
≈ 7.2 MB/hour (decimal)

RFC 6716 defines Opus for interactive speech and audio across a broad bitrate range; 16 kbps is the OneMira design point, not a universal requirement. Actual file size can differ with container overhead and bitrate mode. Storage-capacity claims should specify whether they use decimal or binary units and how much space is allocated to audio versus other data.

Why does Bluetooth’s 2 Mb/s figure not equal file throughput?

The Bluetooth Core Specification defines LE 2M at 2 Mb/s on the physical layer. Bluetooth SIG’s LE primer estimates a lower maximum application rate because packets also carry protocol overhead and must share timed radio events. Real throughput falls further with operating-system scheduling, connection parameters, interference, retransmission, and implementation.

That is why C1 also supports USB-C for larger transfers. Do not calculate transfer time by dividing file size by the PHY headline. Measure application bytes committed per second on supported macOS and Windows machines, including weak-signal and reconnect cases.

What do local files add beyond local ASR?

Local ASR can still write into a proprietary database. Open Markdown changes the post-transcription boundary: users can inspect, search, version, back up, and migrate notes with ordinary tools. A local SQLite index can accelerate search while Markdown remains the durable readable representation.

This also makes failures observable. If an index is deleted, it can be rebuilt from files; if a parser changes, the source text remains inspectable. Open files do not remove the need for backups, encryption, permissions, or schema evolution.

How should you evaluate a local voice-AI system?

  • Offline test: after model installation, capture, transfer, transcribe, search, and export with network disconnected.
  • Execution test: log the CPU/GPU/NPU provider actually selected and verify fallback.
  • Quality test: use a labeled corpus and critical-term recall, not a demo sentence.
  • Recovery test: interrupt at every destructive state transition.
  • Portability test: bulk-open notes with an independent editor and search tool.
  • Disclosure test: identify which text, if any, is sent to a third-party model.

For the data-boundary comparison, read Cloud vs. local AI voice recorders. For recovery testing, read the DVT guide.

Sources and further reading