SenseVoice vs Qwen3-ASR: Which Is Better for Chinese, Japanese and Korean?

The accuracy numbers are our own. We ran SenseVoice Small, Qwen3-ASR and Whisper Large V3 Turbo on the same FLEURS read sentences, about 30 minutes per language, with the models Whisper Notes ships. The speeds are the ones the app shows on each model's card, from a separate test on short clips. Everything else here comes from how the Mac app works today.

How much more accurate is Qwen3-ASR than SenseVoice?

Character error rate for Chinese, Japanese and Korean, word error rate for English; lower is better. Green marks the lowest in each row. Our own run on the FLEURS test set, about 30 minutes and 139 to 180 read sentences per language, on a Mac. SenseVoice with the language set. Clean read speech only: noisy rooms, meetings and long recordings are not covered.

Qwen3-ASR makes fewer mistakes than SenseVoice in every language we tested. In Chinese and Japanese the gap is about one point: 6.49% (Qwen3-ASR) against 7.69% (SenseVoice) in Chinese, and 5.74% against 6.78% in Japanese. Korean has the biggest gap: SenseVoice gets more than twice as many characters wrong, 7.85% against 3.54%. In English, Qwen3-ASR gets about one word in 22 wrong and SenseVoice about one in 12.

Whisper Large V3 Turbo is close to Qwen3-ASR. It was the most accurate of the three in Japanese and second in Korean and English, but slightly behind SenseVoice in Chinese.

How much faster is SenseVoice than Qwen3-ASR?

Green marks the faster model. These are the speeds shown on the model cards in the app: engine processing time only, without loading the model, measured on short clips in a separate everyday-use test on a Mac. SenseVoice ran at 115.2× with the language set and 131× with automatic detection; Qwen3-ASR measured 7.65× before rounding. The one-hour times are worked out from these speeds and rounded.

SenseVoice is more than 15 times faster. For an hour-long recording that is about half a minute against about eight minutes. If you transcribe long recordings, or want the text right away, that difference matters more than a point of accuracy. If you can wait a few minutes and want fewer corrections, it doesn't.

Set the language for a little more accuracy

With the language set, SenseVoice made fewer mistakes than with Auto in all four languages: 7.69% against 7.90% in Chinese, 6.78% against 6.87% in Japanese, 7.85% against 8.09% in Korean and 8.46% against 8.51% in English. The gaps are small, but they all point the same way. We did not run Qwen3-ASR or Whisper with Auto, so we have no numbers for them; the app's own advice is that, except for Parakeet, choosing a language is usually more accurate than Auto.

Parakeet is the exception: in Whisper Notes, telling it the language made it less accurate, which is why it always detects the language itself (see Parakeet Ultra vs Parakeet V3). On Mac, the setting is Language under Transcription in Settings.

How long do long recordings take?

In an earlier test with a real recording, on an Apple M4 Pro, SenseVoice transcribed a 27-minute episode of a Chinese interview podcast in 13.83 seconds. Whisper Large V3 Turbo took 2 minutes 4 seconds for the same file. Qwen3-ASR was not part of that test. That test ran on an earlier version of the app; the speed matches today's model card, about 117× real time against 115–131×.

As a rough estimate scaled up from that test, an hour-long meeting or lecture would take SenseVoice about half a minute and Whisper Large V3 Turbo about four and a half minutes on that Mac. From the speeds above, Qwen3-ASR would need roughly eight minutes.

Neither model reads a long file in one go. In the Mac app, SenseVoice splits the audio into parts of up to 30 seconds, cut at pauses, and shows the text of each part as soon as it is done. Qwen3-ASR also works in parts of up to 30 seconds, cut at the quietest moment near each limit, and its text appears while it is being written. Either way the transcript fills in part by part, and you can start reading before the whole file is finished.

Which one can I use on my device?

Green means available, amber means not available.

On iPhone, SenseVoice is the fast choice for Chinese, Japanese and Korean, with Whisper Small as the alternative.

Why two Mac versions? The Mac App Store version works within the limits Apple sets for App Store apps, so a few features are only in the Mac app from our website: Fn voice typing, the on-device AI assistant and the command-line tool (the background). Both have the same speech models.

What each needs on a Mac

Download sizes and memory labels as shown in the current Mac app from our website. Amber marks a requirement to check before you download. Peak memory is from the same everyday-use test as the speeds above; GiB is about 7% more than GB.

Neither model is built in. Each downloads once, from its card in Settings, and then works offline. SenseVoice is the lighter of the two by a wide margin: about a fifth of the download and well under half the memory of Qwen3-ASR. On a Mac with less than 16 GB of memory, the Qwen3-ASR card shows its memory note in orange as a warning. For comparison, Whisper Large V3 Turbo is about a 2 GB download, and its card says about 2.4 GB of memory.

Which should you use?

• Fewest mistakes, on Mac: Qwen3-ASR 1.7B, in both Mac versions. It beat SenseVoice in all four languages of our test. In Japanese, Whisper Large V3 Turbo was slightly more accurate still, 5.30% against 5.74%.
• Speed, long recordings, a Mac with less memory, or iPhone: SenseVoice Small. An hour of audio takes about half a minute instead of about eight, and it needs a fraction of the download and memory.
• Japanese or Korean accuracy without Qwen3-ASR 1.7B (it needs macOS 15): Whisper Large V3 Turbo, 5.30% against 6.78% for SenseVoice in Japanese and 4.25% against 7.85% in Korean. It is much slower than SenseVoice.
• Other languages: Parakeet, the built-in model, for English and 24 other European languages (see Parakeet Ultra vs Parakeet V3), and Whisper for the rest. Qwen3-ASR 1.7B against Whisper across 30 languages is in our Qwen3-ASR vs Whisper benchmark. Every model's error rate in 42 languages is side by side in our per-language model comparison.

What else should you know?

• Speaker labels, timestamps and subtitles work with both. SenseVoice and Qwen3-ASR both give timestamps, so speaker labels and SRT subtitle export work the same with either model.
• Simplified or Traditional Chinese follows your system language. After transcription, Whisper Notes converts Chinese text to Simplified or Traditional characters on your Mac, following your system language. If you set the language to 简体中文 (Simplified Chinese) or 繁體中文 (Traditional Chinese), that choice takes precedence.
• Dialects are untested. Qwen's own model card lists 30 languages plus 22 Chinese dialects and accents. Whisper Notes has no separate dialect setting: you choose Chinese or Auto, and the model handles the rest. We have not tested dialects, so we can't say how well either model handles them.
• Qwen3-ASR is in Beta. If it ever stops unexpectedly, Whisper Notes turns it off at the next launch and switches to another model, so the app still opens. You can choose Qwen3-ASR again in Settings to retry.
• Our numbers are for clean read speech. FLEURS sentences are read aloud in quiet conditions. Meetings, noisy rooms, far-away microphones and heavy accents can change the gaps in either direction.
• Mixed Chinese and English is untested. We have not measured how either model handles sentences that switch between Chinese and English, so try a typical recording of yours before you settle on one.

How do I switch models?

On Mac, open Settings and scroll to Transcription Model. Each model has a card with its speed, what it is best for, how many languages it covers, how much memory it needs and its download size. To download a model, click its card or the download arrow on it; once it is installed, click the card to switch to it. The languages link on each card (Supported Languages, or 30 supported languages on the Qwen3-ASR 1.7B card) lists every language with our error rate for it.

Settings → Transcription Model in the Mac app from our website

Then set the Language under Transcription instead of leaving it on Auto.

What is SenseVoice?

From the FunAudioLLM paper: SenseVoice-Small has 234 million parameters and processes 10 seconds of audio in 70 ms; Whisper-Large-V3 has 1,550 million and takes 1,281 ms. Both on a data center GPU, not on a Mac.

Qwen3-ASR comes from Alibaba's Qwen team under the Apache 2.0 license. In Whisper Notes it runs on the Mac's GPU through MLX, at 8-bit. Its results against Whisper in 30 languages are in our Qwen3-ASR benchmark.

Frequently Asked Questions

Try it

SenseVoice is in Whisper Notes on iPhone and Mac, and Qwen3-ASR (Beta) is on Mac. Use the buttons below to get the iPhone app from the App Store or the Mac app from our website. Questions? Email support@whispernotes.app.

Download for iOS Download for macOS