Whisper vs Google Speech-to-Text

Google Speech-to-Text is an API, not an app you open

The first thing to get straight: Google Cloud Speech-to-Text has no window. There's no icon in your dock, no hotkey, no "press to talk". It's a service your software talks to over the network. You send it audio with code; it sends back text. Google's own docs describe it as synchronous, streaming, and asynchronous recognition consumed through an API.

That design is good for what it's for. Streaming recognition returns interim results in real time, which is useful if you're building a live captioning feature or a voice command for your own product. Asynchronous recognition handles long recordings: you upload audio, Google chugs through it in the background, and you poll for the result when it's done. Google documents this batch path as handling audio up to eight hours in one job. That's a real strength. If you've got a warehouse of recorded calls to transcribe overnight, an end-user dictation app is the wrong tool, and an API like Google's is the right one.

It supports a long list of languages and regional locale variants, the BCP-47 codes engineers know, like en-US, en-GB, and es-MX. I'm not going to print an exact language count or a per-minute price here, and I'd be careful of any article that does. Google's pricing and language pages move, and the numbers floating around the web don't all trace back to a primary source I'd stand behind. What I can say without hedging: it's usage-based cloud billing. You pay for what you send, your audio goes to Google's servers, and there's no free local mode.

Two people, two different problems

Here's the cleanest way I've found to tell which side of this line you're on. Picture two people.

The first is a developer. She's building a customer-support tool that turns recorded calls into searchable text. The transcription happens on her server, inside her code, with no human watching it run. She wants an endpoint she can send audio to and a JSON response she can store in a database. She is never going to "open" the transcriber. It lives inside the product she ships to her own customers. That's Google Speech-to-Text's job. The API is the component; her product is the app.

The second is a writer. Or a lawyer drafting on the train, or a student turning a lecture into notes, or a parent answering a teacher's email while stirring dinner. He doesn't have a server. He has a cursor blinking in a document, and he'd rather talk than type. He wants to press one key, say the sentence, and watch it appear in the file he already has open. He is never going to write code, and he shouldn't have to. That's our job.

The confusion in the keyword comes from "Whisper" doing double duty. Google STT is a finished cloud service. Whisper is a model, and a model isn't an app. Someone has to build the app around it: plug in the microphone, wire the hotkey, paste the text at the cursor. That's the part we did.

Whisper-in-our-app is desktop dictation, and it runs on your machine

Whisper is the speech model OpenAI open-sourced. Our app runs it locally: pure Rust, no Python sidecar, no server in the loop for ordinary dictation. You press a hotkey (Ctrl+Space on Windows by default, fully remappable), you talk, you release, and the text lands wherever your cursor already is. No code. No API key for the local path. The audio never leaves the laptop.

That last part is the whole point, and it's the one that doesn't show up in a feature table.

On the local tier you pick from eight Whisper models, from about 140 MB up to 3 GB; you trade download size and CPU time for accuracy. Four are tuned for English; the four multilingual ones cover a wide span of languages and can translate speech to English in the same gesture, which Google's API doesn't fold into one dictation press and most consumer tools skip entirely. There's also Parakeet, a separate NVIDIA engine that's 5 to 10 times faster than Whisper on CPU for English and 24 other European languages, and it runs without a GPU.

The whole local pipeline is free for any signed-in user, with no card at signup: every model, AI cleanup through Ollama, history, custom hotwords, the lot. If you want the cloud surface, that's Whisper Pro: OpenAI cloud transcription (gpt-4o-mini-transcribe or gpt-4o-transcribe), cloud AI cleanup, and web search, all on your own OpenAI key, with Remskill taking no cut. That's optional. The default is local and free.

The boring truth is that for one paragraph of dictated text, your laptop already has a microphone and a CPU. It does not need a data center.

The cost models are not the same shape

This is where the comparison stops being apples-to-apples. A cloud API bills per minute of audio. A local dictation app bills, at most, once.

I watched the per-minute model bite once. A team I worked with had a contractor build an internal "AI dictation" prototype that called a cloud API for every utterance. A "smart retry" routine got too aggressive and re-transcribed the same standup recordings four times over. The team manager opened the cost dashboard at the end of the quarter and found a five-figure bill. The contractor's fix was "we should optimize the prompt". The CFO's fix was "or we should not pay for cloud transcription of meetings that already have notes."

That's not a knock on Google's API. Used as intended, by engineers who watch the meter, it's priced fine for production pipelines. It's a knock on using a metered cloud service for something a local app does for free. Cloud-only transcription is a privacy disaster waiting to be billed for. Your draft contracts, your salary spreadsheet, the email to your kid's school, all leaving your machine because you wanted to talk instead of type. For an individual dictating all day, local-first is the right default, and the meter never starts.

Side by side

Here's the honest layout. Notice the table isn't really "which is better". It's "which category are you in".

No specific Google numbers in that table on purpose. The shape is what matters: server vs machine, code vs hotkey, meter vs free. If those rows point you at the API, good, keep reading the next section. If they point you at the app, the download button is at the bottom.

When Google Speech-to-Text is the right tool

I'd reach for Google's API, not our app, in a few clear cases. This is the section AI articles skip, so here it is plainly.

If any of those is you, close this tab and open Google's docs. We don't do server-side. That's not false modesty; it's a different product.

When Whisper-in-our-app is the right tool

The flip side. You're not building software. You're trying to stop typing.

You want to dictate emails, notes, messages, code comments, and have them appear in whatever app you're already in. You'd rather your audio not go to anyone's servers. You don't want a per-minute meter running while you think. You want it free to start, and you don't want to write a line of code to use it.

Pick Parakeet for speed and English; pick a multilingual Whisper model when you need translation, less common languages, or finer control. The local pipeline costs nothing; the Cloud tier (OpenAI transcription with your own key) is optional and priced on the pricing page.

For the offline, local, free side of this question, I wrote up the broader tradeoffs in local vs cloud transcription. And if you're choosing between the two local engines we ship, Whisper vs Parakeet walks through speed versus language coverage.

If you only remember one thing

Google Speech-to-Text is an API for engineers; Whisper-in-our-app is dictation for people. Asking which is "better" is like asking whether a car engine is better than a car. Depends entirely on whether you're building the thing or driving it.