You type "what is whisper" into Google and get a dictionary entry, a Wikipedia article on phonation, and a page about an AI transcription tool. All three are right. They just aren't about the same thing.
OpenAI's Whisper model learned from 680,000 hours of audio, and the hosted API version of it is scheduled to shut down on February 26, 2027. That's the half of what is whisper most results skip, so we'll cover it properly (price, deadline, alternatives) after the everyday meaning and the throat mechanics.
The short answer: two meanings
Whisper is either quiet, breathy speech or an AI model that turns speech into text. Which one you want depends on context.
As a verb or noun in everyday English, to whisper is to speak very softly, using breath rather than your voice, so that only someone close can hear you. "She leaned over and whispered something in his ear." It can also mean a rumor or a faint hint: "There's a whisper that the company is being sold."
In software, Whisper is OpenAI's general-purpose speech-recognition model. If you've seen it mentioned next to transcription apps, subtitles, or meeting notes, that's the one. So when someone asks what is whisper without any other clue, check where they saw the word.
What is whisper as a way of speaking?
Whispering is an unvoiced mode of speech. Your vocal cords stay apart instead of vibrating, and air rushes between the arytenoid cartilages, creating turbulence you hear as that hissy, airy sound. Your tongue, lips and jaw still shape words exactly as they do in normal speech. That's why you can understand a whisper at all.
Try it. Say "hello" out loud, then whisper it with your fingers on your throat. The buzz disappears on the second one.
Why people whisper
Mostly it's social. You whisper in libraries, hospitals, and meetings, or when you're sharing something private. Sometimes it's physical: laryngitis, vocal strain, or a doctor telling you to rest your voice.
There's a trade-off most people don't know about. Whispering isn't automatically gentler on your voice, because plenty of people squeeze their throat and push air hard to be heard across a room. Soft, easy, low-volume speech is usually the better bet if your voice is tired.
Can someone without vocal cords whisper?
Usually not in the normal sense. Without a larynx, the airflow that creates whisper turbulence is gone. People who've had a laryngectomy often use an electrolarynx, esophageal speech, or a voice prosthesis instead. Someone with damaged or paralyzed vocal cords, though, can often still whisper, because whispering doesn't need the cords to vibrate.
Describing a whisper in writing
If you're writing fiction, "whispered" gets overused fast. Better to show the situation: breath against an ear, a voice dropping to almost nothing. We've covered alternatives in what you can say instead of whisper, and there's a nuance-focused list in our quick guide to synonyms for whisper. Quick ones: murmur, mutter, breathe, hiss, undertone.
A whisper should earn its place. If the scene already says it's quiet, skip the tag.
What is whisper in software? OpenAI's speech-to-text model
OpenAI's Whisper is an automatic speech recognition model that transcribes audio, translates it into English, and identifies the spoken language. That's the description in OpenAI's own model documentation.
It launched in 2022 as an open-source project. The original model was trained on 680,000 hours of multilingual audio from the web, which is a big reason it copes with accents, background noise and technical vocabulary better than older tools did.
What it actually does
- Transcription: converts speech in dozens of languages into text in that same language.
- Translation: takes non-English speech and outputs English text.
- Language identification: works out which language is being spoken without being told.
- Timestamps: can mark when each phrase occurs, useful for subtitles.
It doesn't separate speakers on its own, and it processes recorded audio in chunks rather than word by word, so it isn't naturally a live-dictation engine. That surprises people. Plenty of apps wrap it to make it feel live, but the model itself was built for files.
Model sizes
There are several versions. Large-v3 dates to November 6, 2023 and has 1.55 billion parameters, according to Metavert's summary of the model family. The faster Large-v3 Turbo arrived September 30, 2024 with 809 million parameters. Smaller ones (tiny, base, small, medium) trade accuracy for speed and can run on a laptop without much fuss.
What Whisper costs in 2026
The code and weights are free to run yourself, but your own hardware and time aren't. Dasha's breakdown of Whisper pricing points out that self-hosting means no model-license fee, yet you still pay for compute and maintenance. If you only transcribe an hour a week, hosting a GPU is silly. If you transcribe thousands of hours, it might make sense.
The hosted route is per-minute billing:
| Option | Price | Best for |
|---|---|---|
| Self-hosted Whisper | No license fee, you pay for compute | High volume, privacy needs |
| whisper-1 API | $0.006/min ($0.36/hour) | Existing integrations |
| gpt-transcribe | $0.0045/min | Recorded audio, newer projects |
| gpt-live-transcribe | $0.017/min | Live transcription |
A quick key to those names. whisper-1 is the hosted API version of the original Whisper model. gpt-transcribe and gpt-live-transcribe are OpenAI's newer transcription models, not Whisper itself: one is meant for recorded audio, the other for live streams.
The whisper-1 rate comes from OpenAI's pricing, and Cartesia's 2026 model-by-model pricing guide lays out the newer tiers side by side. Pay attention to the date on this one: OpenAI deprecated whisper-1 in 2026 and scheduled the API shutdown for February 26, 2027. If you're building something on it, plan the migration now rather than in January 2027. The open-source weights stay downloadable either way.
Whisper versus other things called Whisper
The name is crowded. A few you might have bumped into:
- Whisper app: the anonymous confession social app, once huge. See Whisper app messages: what they were and where they went.
- MacWhisper: a Mac app that runs Whisper models locally on your recordings. Our explainer on what MacWhisper does covers it.
- Horse whisperer: someone with a calm, intuitive way of handling animals. The word's story is in what the word whisperer means.
If you landed here for the plain definition, what does "whisper" mean goes deeper on modern usage.
Should you use Whisper yourself?
If you have audio files to transcribe and some technical patience, yes. If you want to speak and have text appear in whatever app you're using, probably not directly.
Here's my honest take. Whisper is brilliant as an engine and a little awkward as a product. Running it from the command line means installing Python, picking a model size, and waiting while it chews through a file. On an M1 MacBook Air, a 60-minute interview run through the small model takes roughly 15 minutes in my experience, and the large model can take longer than the recording itself. Fine for a podcaster who does it weekly. Useless if you wanted to dictate an email in the next ten seconds.
A practical way to decide:
- Have recordings? Use a Whisper-based app or the API for batch transcription.
- Need live dictation? Pick a tool built for it, not a file-based model with a wrapper.
- Privacy-sensitive? Run a local model so audio never leaves your machine.
- Building a product? Check the 2027 whisper-1 shutdown before committing.
Choose the tool for the moment you'll use it, not the model with the best name.
Dictating on a Mac without the setup
Most people asking what is whisper in the AI sense really want one thing: talk, and get clean text. That's a different job from transcribing a file. Full disclosure, GhostWriter is our app, so weigh this accordingly. It's a macOS app that turns your speech into text in whatever application you have open, adding punctuation and formatting as you go and adjusting tone for contexts like emails or code editors.
A concrete case: you're in Mail, you hold the dictation shortcut, say a three-sentence reply, and it lands in the message ready to send. No model to download, no terminal.
If you're still curious about Whisper itself, grab the small model and run it on a one-minute voice memo. You'll know in five minutes whether you want a transcription engine or just a faster way to write.