Local vs cloud transcription

Local means the model lives on your disk

Local transcription downloads a speech model once, then runs it on your processor. No upload, no server, no account ping during a recording. Pull the network cable and it still types.

Our app does this in pure Rust through a library called transcribe-rs, with no Python runtime bolted on. You pick from eight Whisper models, from Base at about 140 MB up to Large v3 at roughly 3 GB, or NVIDIA's Parakeet at around 600 MB, which is five to ten times faster than Whisper on a CPU. No GPU required. The model loads into RAM, your voice goes in, text comes out, and nothing about that round-trip touches the internet.

The download is the only friction. A 3 GB model is a real download on hotel Wi-Fi, and a 2018 laptop will run a small model fine but choke on the big one. After that first download, though, there is no per-minute cost and no server in the loop. If you want the deeper version of this, I wrote a whole piece on running it fully offline. See offline speech to text on the desktop.

Cloud means your audio takes a trip

Cloud transcription records your audio, sends it to a provider's server, and the server sends back text. You are renting someone else's hardware and their newest model.

In our app, cloud mode is bring-your-own-key OpenAI. Transcription runs on gpt-4o-mini-transcribe or the higher-quality gpt-4o-transcribe, and you can layer on AI cleanup and live web search through the same key. You supply your own OpenAI key and pay OpenAI directly. We take no cut and add no markup. There is no big model to download. It runs the same on a five-year-old netbook as on a new workstation, because the work happens on the server, and it can answer a question by searching the web, which a local model simply cannot do.

The cost is right there in the name. Your audio leaves your machine. You need a live connection. And you pay by the minute, fractions of a cent, but it adds up, and it is metered.

The honest side-by-side

No dollar figures in this table on purpose. See our pricing page for the actual numbers. This is about the shape of each choice.

Read that top to bottom and the pattern is clean. Local trades convenience for privacy, offline use, and a flat cost. Cloud trades privacy and metering for the newest model and a web connection. Neither is better. They are good at different jobs.

When cloud is the better call

I am not going to pretend local wins every time. There are real cases where I would reach for cloud.

If your hardware is old or RAM-starved, cloud is the kinder option. A 2017 laptop with 8 GB of RAM will fight a large local model, while the cloud does the heavy lifting elsewhere and your machine just handles the microphone. If you need the absolute latest transcription quality on hard audio, things like heavy accents, overlapping speakers, or technical jargon, the newest hosted models tend to edge out what you can run at home. And if you want to dictate a question and get a web-sourced answer pasted back at the cursor, that needs the cloud, full stop. A local model has no internet to search.

The thread connecting those: cloud is the escape hatch for weak hardware, top-end quality, and live web access.

When local is the better call

For most people, most of the time, local is what I would start with.

If the words you dictate are private, a salary spreadsheet, an email to your kid's school, a legal draft, they should not end up in a vendor's logs because you wanted to type with your voice. Local keeps that audio on your machine, period. If you work on planes, trains, or in coffee shops with hostile Wi-Fi, local does not care whether you have a signal. And if you dictate a lot, the flat cost matters.

Here is the opinion I will actually commit to: try local first, and treat cloud as the escape hatch, not the default. If your Mac is Apple Silicon or your PC is from the last four years, local handles everyday dictation at 95% to 99% accuracy without a server in the loop. Fall back to cloud when you hit a wall, whether that is weak hardware, hardest-of-the-hard audio, or a need for web search. Most people never hit the wall.

I have a reason for being twitchy about cloud-by-default. A team I worked with once let a contractor build an internal cloud-AI dictation prototype that called the API for every utterance. A smart retry loop transcribed the same standup recordings four times over. The manager opened the cost dashboard at the end of the quarter and found a five-figure bill. The contractor's fix was optimize the prompt. The CFO's fix was stop paying to transcribe meetings that already have notes. Metered cloud is fine until something loops. Local does not have a meter to run away with.

Both modes live in one app

The split above is real, but it is not a fork in the road you commit to once. In our app both modes sit behind the same hotkey and the same recording overlay, and the toggle is one switch. Dictate a private email locally in the morning, flip to cloud to fact-check a claim with web search in the afternoon, flip back. You do not reinstall anything. You do not pick a lane for life.

That is the part the local vs cloud framing tends to miss. It is not a religious war. It is two tools in one drawer, and the right one depends on the sentence you are about to say. If you want the local engines compared against each other, speed versus language coverage, that is its own piece: Whisper vs Parakeet. And if you are weighing us against a specific competitor, the superwhisper comparison walks through one in detail.

If you only remember one thing

Local for privacy, offline, and flat cost. Cloud for the newest model, weak hardware, and web access. Try local first and keep cloud as the escape hatch. The best part is not having to choose forever: one toggle, both modes, whichever fits the sentence you are about to say.