Local AI transcription: on-device speech to text, explained

Local vs cloud transcription

Most transcription services, including popular meeting note takers, are cloud services. Your recording is uploaded, a large model on the company's servers transcribes it, and the text comes back. That's convenient and works on any device, but your audio sits on someone else's computer, often with policies about how it can be used.

With on-device transcription, the model file lives on your computer. Audio goes in, text comes out, and nothing leaves the machine. It also means no per-minute fees and no internet needed once the model is downloaded, which is why people search for "offline meeting transcription."

How local speech recognition works

1. Prepare the audio. The recording is converted to a simple format, typically 16,000 samples per second in mono, which is what most speech models expect.
2. Find the speech. A small voice activity detection model marks where people are talking and skips silence, so the main model doesn't waste time or invent words in quiet stretches.
3. Cut it into pieces. Long recordings are split into chunks, ideally at pauses so words aren't cut in half.
4. Recognize the words. The speech model turns each chunk into text, usually with timing for each word or phrase.
5. Work out who spoke (optional). A separate step, called speaker diarization, groups the audio by voice. Our guide to speaker labels in transcription explains it.

The models: Whisper and Parakeet

OpenAI Whisper

Whisper is the best-known open speech model. Its code and weights are released under the MIT license and it handles many languages. It comes in several sizes, and you trade speed for accuracy:

Figures from OpenAI's Whisper repository; speeds were measured on a data-center GPU, so expect slower on a laptop.

Community projects such as whisper.cpp make Whisper run efficiently on ordinary CPUs and Apple Silicon, which is what many desktop apps use under the hood.

NVIDIA Parakeet

Parakeet is a family of speech models from NVIDIA. The English model Parakeet TDT 0.6B v2 has 600 million parameters, is licensed CC-BY-4.0, and reports an average word error rate of 6.05% across the Hugging Face Open ASR Leaderboard test sets, with word-level timestamps. It's designed to be very fast, which matters when you want a transcript while the meeting is still going. The same model card shows the honest catch: in heavy background noise, its error rate roughly doubles.

What hardware do you need?

• Processor: any recent laptop can run small and mid-sized models on the CPU. Transcribing an hour of audio takes from a few minutes to longer than the recording, depending on the model and machine.
• Graphics card: optional for most apps, but it speeds up the largest models a lot. Apple Silicon Macs are fast at this.
• Memory: the model has to fit in memory alongside your meeting app, so more RAM helps, and the largest models need the most.
• Disk: model files range from tens of megabytes to a few gigabytes. You download them once.
• Battery: transcription is heavy work, so expect your fan to spin and battery to drain faster while it runs.

Accuracy and privacy trade-offs

On accuracy, the gap has narrowed. Clean recordings come out well with good local models. The things that hurt every system are the same: people talking over each other, distant microphones, heavy accents, and unusual names. A good tool lets you fix mistakes quickly and keeps your corrections.

One more nuance: "local transcription" doesn't always mean "everything local." Many apps transcribe on your device and then send the text to a cloud AI for summaries. That's often a reasonable trade, since text is smaller and less identifying than your voice, but it's worth knowing which parts leave your computer.

Local transcription options

• Windows Live Captions (built in). On Windows 11 22H2 or later, press Win + Ctrl + L to show captions for any audio. Microsoft says the captions are processed on the device and not stored, so it helps you follow a call but doesn't give you a transcript to keep.
• MacWhisper (Mac). A popular Mac app that runs Whisper locally to transcribe audio and video files.
• Buzz (Windows, Mac, Linux). A free, open-source app that runs Whisper on your computer to transcribe files.
• whisper.cpp and other command-line tools. Free and flexible if you're comfortable with a terminal.
• Krisp. A meeting assistant that says it transcribes English on-device and uses servers for other languages.
• Hark (Windows, Mac, Linux). Built for meetings: records, transcribes locally, and labels speakers. Details below.

File transcribers like MacWhisper and Buzz are great for recordings you already have. For live meetings, you also need something that records both sides of the call; see our list of transcription software for Mac for more.

How Hark does local transcription

Hark is a desktop meeting recorder that does the speech work on your computer:

• Models: NVIDIA Parakeet for speech recognition, Silero for finding speech and NVIDIA TitaNet voiceprints for telling speakers apart, all run through the open-source sherpa-onnx engine.
• One-time download: about 730 MB of models on first use. After that, transcription works offline.
• Two passes: a live transcript while the meeting runs, then a more accurate full transcript after it ends.
• Word timing: the word being said is highlighted as you play back, and you can click a word to jump there.
• Speakers you name are remembered and suggested automatically in future meetings.
• No limits: because it runs on your computer, recording and transcription are free with no minute caps.
• Imports too: drop in an mp3, m4a, wav, mp4 or other file and it's transcribed the same way.

What to know before you choose it: Hark transcribes English only. On a Mac it records the microphone only (not the computer's audio). And while transcription is local, Hark's optional AI features (meeting notes, titles, proofreading) send transcript text, not audio, to Claude, so they're not offline. If you only use the free features, nothing leaves your computer. For more on what that means in practice, read AI note takers without a bot or how one cloud note taker handles data.

If you record calls with other people, remember that recording laws vary, and some places require everyone's consent. Let people know before you start.