Is on-device speech recognition as accurate as cloud?

For short dictation in major languages, on-device speech recognition is now about as accurate as cloud recognition. Some reports say it matches cloud accuracy, and one peer-reviewed study found an on-device model that beat a conventional model at the same latency. That is not the same as saying every on-device engine equals every cloud service. Accuracy depends more on the engine, the language and the recording than on whether the model runs on your device or on a server.

The guide to on-device and cloud speech recognition covers the wider trade-offs, including privacy, cost and offline use. This page answers the accuracy question on its own.

Accuracy depends on the engine more than on where it runs

The clearest way to see this is to compare two engines on the same recording. In a 2025 test by 9to5Mac, Apple's speech API had a word error rate of 10.3 percent on one 7-minute-31-second episode ¹. Whisper Large V3 Turbo had 1.5 percent on the same episode ². On that recording, Apple's error rate was about 6.9 times the Whisper rate ³.

The gap is a reminder that a single test is a single test. The author said the models were evaluated the same way, but that the exact numbers would differ under slightly different methods. The lesson is not that one engine is always better. It is that you should not assume an engine is accurate because of where it runs. Measure the engine you plan to use.

Studies and vendors that favor on-device recognition

Google said in 2019 that its new on-device recognizer offered the same accuracy as its traditional server-based models in a much smaller model ⁴. That is a vendor claim, made at launch under Google's own tests, so treat it as a claim to check rather than as an independent result.

A peer-reviewed study is stronger evidence. Sainath and colleagues report that their streaming on-device model obtained an 8 percent relative improvement in word error rate over a conventional model at the same latency, and that it was more than 400 times smaller ⁵. An 8 percent relative gain is not a large absolute change when the error rate is already low, so read it as evidence that on-device models can be competitive, not as a general ranking.

Apple's advanced dictation in the iOS 27 beta shows a similar pattern in practice. 9to5Mac reports that this dictation runs only on the on-device model, which gives the same transcription quality whether or not the phone has a data connection ⁶. This is a report on a beta feature, and it describes consistency between online and offline use rather than a score against a named cloud service.

Where cloud recognition still has the edge

Cloud recognition still leads in three places.

• Largest models. A server is not limited by a phone's memory, so it can run larger models than a phone can hold.
• Less common languages. Apple lists 63 language and regional variants for Mac Dictation, and 15 of them are not on its on-device and modeless list ⁷. If your language is one of those, the cloud option may be the only one with good support. Check the list for your language first.
• Recordings you have not tested. Microsoft describes its online speech recognition in Windows as using Microsoft cloud-based services for more accurate recognition ⁸. This is the vendor's own description of its design, not a measured comparison, but it is a fair reason to test the cloud option on your audio before you rule it out.

Noise, accents, names and specialist vocabulary are not covered by the facts above, and they can change the result for any engine. Measure with recordings that have those features.

It depends: what changes the answer

The answer shifts with four things.

1. The language. A major language on a recent engine is the best case for on-device accuracy. A less common language may only work well in the cloud.
2. The audio. Short, clear dictation is the easiest input. Long recordings, several speakers and loud rooms can make every engine worse.
3. The model size you can run. A phone or laptop runs a smaller model than a data center, so the best local engine and the best cloud engine are not always the same models.
4. The vocabulary. Names, product terms and jargon are where a general model slips. A custom vocabulary can help on either side, if the engine supports one.

What to do

1. Pick your test audio. Use 10 to 15 minutes of the speech you actually produce, with the noise and names it contains.
2. Transcribe it by hand to make a reference text, then score each engine with word error rate.
3. Compare on the same audio, the same language and the same settings. A fair test needs the same input for every engine.
4. Choose the lowest error rate you can run within your limits on memory, battery and privacy.
5. Recheck every few months, because engines change. A result from one year may not hold the next.

Yap, which we make, is an on-device dictation app for Mac and Windows, so we have a stake in this answer. We have tried to state where the evidence is weak. For the free options on Mac, see offline dictation on Mac, and for the Windows download, see the download page.

Related pages

• Can you run speech recognition on a Raspberry Pi? covers the hardware limits of small boards.
• What is Speech Recognition and Synthesis from Google? explains the Android engine that Google publishes under that name.
• Is speech recognition more accurate than humans? looks at the comparison with people.
• Cloud dictation and on-device speech recognition define the two designs.