TL;DR — Our own FLEURS run, 30 shared languages
| Qwen3-ASR 1.7B (Beta) | Whisper Large V3 Turbo | |
|---|---|---|
| Languages | 30 named | 101 |
| Wins, of 30 shared languages | 15 | 15 |
| Wins outside the error bars | 5 | 11 |
| CER average — Chinese, Cantonese, Japanese, Korean, Thai | 5.30% | 13.44% |
| WER average — the other 25 | 13.99% | 10.16% |
| Cantonese | 6.00% CER | 37.97% CER |
| Download / peak memory in the run | ~2.5 GB / 2.43 GiB | ~1.6 GB / 1.87 GiB |
| Where you can run it | Mac Direct Download (DMG) 1.6.0+ only | Both Mac builds |
* Our own measurement, not an official leaderboard. Google FLEURS test split, frozen revision 70bb2e84, about 15 minutes of audio per language per model, one M5 MacBook Air (32 GB, macOS 26.5). FLEURS is short, clean, read speech: these numbers do not predict meetings, dictation, noise or long recordings. WER and CER are never averaged into one score.
Is there a local model that beats Whisper in your language? Since version 1.6.0 of the Mac Direct Download (DMG) build, Whisper Notes for Mac has a fourth engine you can try: Qwen3-ASR 1.7B, running as MLX 8-bit entirely on your Mac, in Beta. We ran it against Whisper Large V3 Turbo on the same machine, the same audio and the same scoring, in all 30 languages it names. The answer turned out to be "yes, in fifteen of them" — Whisper took the other fifteen — and the fifteen are not the ones we expected.
We are the people who ship the app, so we tried to make the comparison hard to fudge. The Whisper we ran is Large V3 Turbo because that is the largest Whisper the app ships. It is still our own measurement rather than an independent benchmark, and read speech rather than real recordings.
The short answer
- If you transcribe Cantonese, this is the largest single result in the run. Qwen3-ASR 1.7B scored 6.00% CER where Whisper Large V3 Turbo scored 37.97%, and SenseVoice Small scored 36.71% on the same clips. Cantonese is the one language where the new engine changes what is possible on the Whisper Notes Mac Direct Download build.
- If you transcribe Hindi, Thai, Vietnamese or French, Qwen3-ASR 1.7B was ahead by a margin wider than the error bars — Hindi by 16.7 points, Thai by 7.0, Vietnamese by 4.7, French by 1.4.
- If you transcribe Chinese, Japanese, Korean, English, German, Spanish or Italian, it was ahead by a hair, and a hair is all we will claim. Those gaps sit inside the 95% band, which means this run cannot separate the two models. Either is a reasonable pick. Persian, Arabic and Indonesian sit in the same band, as do Russian, Portuguese, Macedonian and Malay in Turbo's favour.
- If you transcribe Finnish, Hungarian, Greek, Czech, Swedish, Danish, Polish, Romanian, Filipino, Turkish or Dutch, stay on Whisper Large V3 Turbo. Turbo won all eleven by a clear margin, and Finnish and Hungarian by more than 19 points.
- If speed or disk space is what you care about, this is not your engine. Parakeet V3 and SenseVoice Small were an order of magnitude faster in the same harness, and both are a fraction of the download.
- If you are on iPhone or the Mac App Store build, this engine is not available to you at all. It is in the Mac Direct Download (DMG) build from 1.6.0 onwards and nowhere else.
What is Qwen3-ASR 1.7B, and what is it not?
Qwen3-ASR is Alibaba's speech recognition family, built on the audio understanding of their Qwen3-Omni foundation model. The 1.7B model card claims "state-of-the-art performance among open-source ASR models", and lists 52 languages and dialects: 30 named languages plus 22 Chinese dialects.
Alibaba also sells Qwen3-ASR as a cloud service. If you have read about a "Qwen3-ASR API", that is the DashScope Real-time API or the FileTrans API, where you upload audio to Alibaba's servers and receive text back.
That is not what Whisper Notes does. We take the open weights, convert them to Apple MLX at 8-bit, and run them on your Mac's own GPU. There is no account, no API key and no upload. Your audio never leaves the machine, in this engine exactly as in the other three. That property is the reason we run any model locally, and it is the one thing that does not vary between the engines in the picker.
Three things it is not — the other three engine families ship in every channel, but this one is not on iPhone, not in the Mac App Store build, and not out of Beta. Those three are Whisper, Parakeet V3 and SenseVoice; the Mac Direct Download (DMG) build 1.6.0 and later adds Qwen3-ASR 1.7B as a fourth.
Didn't SenseVoice replace Qwen3-ASR already?
It did, and that sentence is still true. In Direct Download (DMG) 1.5.0 we replaced Qwen3-ASR with SenseVoice Small. The model we removed was the 0.6B Qwen3-ASR, running autoregressively through sherpa-onnx on the CPU. It took 224 seconds on a 27-minute Chinese podcast. SenseVoice did the same file in 13.83 seconds.
What arrived in 1.6.0 is a different model. It is nearly three times the parameters, it runs on the GPU through MLX rather than on the CPU through sherpa-onnx, and it is scored separately in everything below. The 0.6B is not coming back. Treat these as two members of the same family rather than one model leaving and returning.
SenseVoice did not go anywhere either. It is still the fast pick for Chinese, Japanese and Korean, on every channel, and it is still the only one of the two you can run on an iPhone.
How did we measure this?
The run is a frozen protocol rather than a set of one-off timings, so here is enough of it to argue with.
- Models. Qwen3-ASR 1.7B as the MLX 8-bit build the app ships; Whisper Large V3 Turbo as the whisper.cpp ggml build the app ships, on Metal.
- Data. The test split of Google FLEURS, pinned to revision 70bb2e84, CC-BY-4.0.
- Sampling. A deterministic prefix of about 15 minutes of audio per language. Every model saw the same clips in the same order, so no model got an easier draw.
- Scoring. Word Error Rate for space-delimited languages, Character Error Rate for Chinese, Cantonese, Japanese, Korean and Thai. The two are never mixed into a single global number, because a character rate and a word rate do not mean the same thing.
- Error bars. A 10,000-resample bootstrap over items on every cell. FLEURS does not publish speaker cluster metadata, so the intervals are item-level and we label them as such.
- Machine. One M5 MacBook Air, 32 GB, macOS 26.5. Empty outputs and failures were counted and kept in the denominator. There were none for either model.
Now the part that matters more than the protocol. FLEURS is short, clean, read speech: someone reading a sentence into a good microphone. Your meetings are not that. Your dictation is not that. We have watched read-speech rankings fail to survive contact with real audio twice in our own testing, so the standing rule in our benchmark repo is that read-speech rankings do not extrapolate to long real-world recordings. Use the table below to answer "is this model plausible for my language", not "this is the accuracy I will get".
We have not finished a real-meeting evaluation of Qwen3-ASR 1.7B yet. When we do, it will be a separate post with its own methodology.
Where does each model actually win?
All 30 languages, sorted from Qwen's largest win to its largest loss. Lower is better in both columns. The last column says whether the gap is wider than the paired 95% interval — where it says no, this run cannot tell the two models apart, and you should read the row as a tie.
Our own FLEURS run, one M5 MacBook Air, read speech.
| Language | Metric | Qwen3-ASR 1.7B | Whisper Large V3 Turbo | Difference | Outside the error bars? |
|---|---|---|---|---|---|
| Cantonese | CER | 6.00 | 37.97 | -31.98 pp | Yes |
| Hindi | WER | 13.67 | 30.42 | -16.74 pp | Yes |
| Thai | CER | 5.92 | 12.95 | -7.04 pp | Yes |
| Vietnamese | WER | 5.07 | 9.79 | -4.72 pp | Yes |
| Persian | WER | 26.98 | 30.03 | -3.05 pp | No |
| Arabic | WER | 14.14 | 15.75 | -1.61 pp | No |
| Indonesian | WER | 4.97 | 6.50 | -1.54 pp | No |
| French | WER | 3.98 | 5.39 | -1.40 pp | Yes |
| Chinese | CER | 5.69 | 6.88 | -1.19 pp | No |
| English | WER | 5.07 | 5.92 | -0.85 pp | No |
| German | WER | 3.51 | 4.10 | -0.60 pp | No |
| Japanese | CER | 5.39 | 5.71 | -0.32 pp | No |
| Korean | CER | 3.51 | 3.70 | -0.18 pp | No |
| Spanish | WER | 3.60 | 3.76 | -0.15 pp | No |
| Italian | WER | 2.98 | 3.05 | -0.07 pp | No |
| Russian | WER | 5.98 | 5.51 | +0.47 pp | No |
| Portuguese | WER | 5.64 | 4.91 | +0.73 pp | No |
| Macedonian | WER | 17.81 | 16.64 | +1.18 pp | No |
| Malay | WER | 11.60 | 10.18 | +1.41 pp | No |
| Dutch | WER | 7.65 | 5.63 | +2.02 pp | Yes |
| Turkish | WER | 8.34 | 5.53 | +2.81 pp | Yes |
| Danish | WER | 20.12 | 14.92 | +5.20 pp | Yes |
| Polish | WER | 12.78 | 5.32 | +7.46 pp | Yes |
| Romanian | WER | 18.97 | 8.22 | +10.75 pp | Yes |
| Filipino | WER | 20.44 | 9.37 | +11.07 pp | Yes |
| Swedish | WER | 20.04 | 8.72 | +11.31 pp | Yes |
| Czech | WER | 23.92 | 11.15 | +12.77 pp | Yes |
| Greek | WER | 30.20 | 12.97 | +17.23 pp | Yes |
| Finnish | WER | 26.08 | 6.54 | +19.53 pp | Yes |
| Hungarian | WER | 36.35 | 13.66 | +22.69 pp | Yes |
Our own paired run on the FLEURS test split, revision 70bb2e84, about 15 minutes per language, one M5 MacBook Air. "pp" is percentage points. "Outside the error bars" means the paired 95% bootstrap interval for the difference does not contain zero.
Three things fall out of that table, and only one of them is flattering.
The scoreboard is a tie. Fifteen languages each. There is no version of this data where Qwen3-ASR 1.7B is a drop-in replacement for Whisper across the board, and we are not going to present it as one. Sixteen of the thirty gaps are wider than the error bars — five of them Qwen's, eleven Turbo's; the other fourteen are ties in everything but the decimal place.
The wins and the losses live in different places. Averaged over the five languages scored by character — Chinese, Cantonese, Japanese, Korean, Thai — Qwen scored 5.30% against Turbo's 13.44%. Averaged over the other twenty-five, scored by word, Qwen scored 13.99% against Turbo's 10.16%. Its clear wins are the character-scored languages plus Hindi, Thai and Vietnamese; it falls apart on the Nordic and Central European languages that Turbo handles comfortably, and on Filipino, where Turbo is 11 points ahead.
Cantonese is the row that changes a decision. 6.00% against 37.97% is not a tuning difference. We had also published SenseVoice as the Cantonese answer, and on these same clips SenseVoice scored 36.71% CER, close to Turbo. Neither of those clears the bar for Cantonese on this read-speech set, and we would rather add that qualifier than leave the older recommendation standing without one. If Cantonese is your language and you are on the Mac Direct Download build, try the Beta. If you are on iPhone or the Mac App Store build, SenseVoice remains the Cantonese engine you have, and it remains far better on Chinese, Japanese and Korean than its Cantonese row suggests.
What does that accuracy cost?
Accuracy is one axis. Here are the other three, measured in the same harness on the same machine, so the engines are at least comparable to each other.
| Engine | Languages | Median speed in this run (short clips) | Peak memory | Download | Channels |
|---|---|---|---|---|---|
| SenseVoice Small | 6 choices | 161.0x | 0.95 GiB | 827 MB | iPhone, both Mac builds |
| Parakeet TDT 0.6B v3 | 25 | 82.9x | 0.09 GiB | 465 MB | iPhone, both Mac builds |
| Qwen3-ASR 1.7B (Beta) | 30 named | 8.7x | 2.43 GiB | ~2.5 GB | Mac Direct Download (DMG) 1.6.0+ only |
| Whisper Small | 101 | 6.4x | 0.87 GiB | 600 MB | iPhone, both Mac builds |
| Whisper Large V3 Turbo | 101 | 2.4x | 1.87 GiB | ~1.6 GB | Both Mac builds |
Median realtime factor across each engine's languages in the same FLEURS run, one M5 MacBook Air. These are short-clip harness figures and they are lower than what the app does on a long recording. Use the column to compare the engines against each other inside this one run, and nothing else. Peak memory is the maximum process footprint observed across items.
Read that table as a ladder rather than a ranking. Parakeet V3 and SenseVoice Small are roughly an order of magnitude quicker than Qwen3-ASR in the same harness, and they are 465 MB and 827 MB against 2.5 GB. In the app itself, on a long file, Parakeet has run at 103x realtime on an M4 Pro and SenseVoice at 52x or better, which is a different measurement on different audio but points the same way.
Qwen3-ASR sits in the middle of that ladder. It was not the slowest engine in this short-clip run — Whisper Large V3 Turbo was — but it varies more than the others by language, from 2.7x on Greek to 12.4x on Japanese. And it is the heaviest thing in the app on every footprint axis: the largest download, the largest peak memory, and the only engine that raises the system requirement.
What else should you know before switching?
Five conditions come attached to this engine. None of them is a defect. All of them are the shape of a 1.7B model running on a laptop.
- It is a Beta. That is the label in the picker, and it means the behaviour and the defaults can still move between releases. The other four engines are not Betas.
- It needs macOS 15 or later, on Apple silicon, with 16 GB of memory recommended. The rest of the Mac app runs on macOS 14. This engine is the one thing in it that asks for more, and 2.43 GiB of peak process memory in our run is why; on an 8 GB Mac, Parakeet V3 and Whisper Small cover the same job in 465 MB and 600 MB.
- It is about a 2.5 GB download. It downloads inside the app on first use, resumes if you cancel it or lose the connection, and is verified before it is used. If disk space is tight, Whisper Small covers 101 languages in 600 MB.
- Voice Typing with this engine is accuracy-first, not live. Hold Fn, speak, release, and the finished text is pasted once. We tested a genuinely streaming version against this one on the same audio and the streaming path was worse: 9.25% against 3.13% WER on 30 English pairs, 22.60% against 2.93% on 30 German pairs. We took the accuracy. It does accept very short utterances, so dictating a one-line correction still works. The engine also has to be loaded before Voice Typing will accept the key, rather than recording you into a model that is not ready.
- Imported audio does show text as it goes. That is a different path: the file is transcribed offline in one pass and the words appear on the page as the model produces them. What you keep is the final result of that pass.
If none of that appeals, nothing about your setup has to change. Parakeet V3 is still the default and still the right answer for English and most of Europe, and the model comparison table for all the non-Beta engines lives in that post.
Which 30 languages, and how do you turn it on?
Qwen3-ASR 1.7B names 30 languages: Arabic, Cantonese, Chinese, Czech, Danish, Dutch, English, Filipino, Finnish, French, German, Greek, Hindi, Hungarian, Indonesian, Italian, Japanese, Korean, Macedonian, Malay, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Thai, Turkish and Vietnamese. Simplified and Traditional Chinese are two choices in the picker sharing one Chinese request, which is why you will count 31 checkmarks against 30 names on our language support page.
Every one of those 30 is also covered by Whisper, on every channel — what changes is the error rate, by as much as 32 points on Cantonese.
To enable it: open Whisper Notes 1.6.0 or later from the Mac Direct Download (DMG), go to Settings, then Transcription Model, and choose Qwen3-ASR 1.7B. The model downloads on first use. If the download is interrupted, reopening the picker resumes it rather than starting over.
The picker carries the same numbers as the table above. Select a model, open its language list, and each language shows its measured WER or CER with the lowest value across the engines in green, under the note "On-device results on a subset of FLEURS." We would rather you pick a model from measurements than from marketing, including ours.
It also works from the command line. The local CLI and MCP server accept "qwen", "qwen3" and "qwen3-asr" as names for this engine, and run it through the same app, the same language selection and the same chunking as the window does.
Frequently Asked Questions
Is Qwen3-ASR more accurate than Whisper?
It depends on the language, and the split is close to even. On our own FLEURS run inside Whisper Notes for Mac, across the 30 languages both models cover, Qwen3-ASR 1.7B was ahead in 15 and Whisper Large V3 Turbo in 15. Qwen led clearly on Cantonese (6.00% vs 37.97% CER), Hindi (13.67% vs 30.42% WER), Thai, Vietnamese and French. Turbo led clearly on Finnish (6.54% vs 26.08% WER), Hungarian, Greek, Czech, Swedish, Danish, Polish, Romanian, Filipino, Turkish and Dutch. Fourteen of the thirty gaps sit inside the error bars and should be read as ties. Choose Qwen3-ASR if your language is in its winning column and you can accept a Beta on the Whisper Notes Mac Direct Download (DMG) build; choose Whisper Large V3 Turbo for everything else, and for every language outside Qwen's 30, since Whisper reaches all 101 choices.
Can I use Qwen3-ASR on iPhone or from the Mac App Store?
No. Qwen3-ASR 1.7B is in the Mac Direct Download (DMG) build of Whisper Notes from version 1.6.0 onwards, and in no other channel. The iPhone app and the Mac App Store build have three engine families — Whisper, Parakeet V3 and SenseVoice — and every language Qwen recognises is also covered by Whisper on those channels. Counted as models, the Mac App Store build offers four and the Direct Download build offers five, because Whisper ships as both Small and Large V3 Turbo. If you need Cantonese on iPhone, SenseVoice is the engine to use there.
Qwen3-ASR or SenseVoice for Chinese, Japanese and Korean?
Both are close on accuracy and very far apart on speed. On our FLEURS run, Qwen3-ASR scored 5.69% CER on Chinese, 5.39% on Japanese and 3.51% on Korean; SenseVoice Small scored 7.06%, 6.85% and 7.82% on the same clips. SenseVoice ran at a median 161x realtime in that run against Qwen's 8.7x, downloads 827 MB against about 2.5 GB, and is available on iPhone and both Mac builds. Choose SenseVoice unless you are transcribing Cantonese, where Qwen3-ASR scored 6.00% CER against SenseVoice's 36.71% and the gap is large enough to be worth the wait.
What does "Beta" mean for this engine?
It means the engine ships and works, and that its behaviour and defaults can still change between releases of the Mac Direct Download (DMG) build. The other four models in Whisper Notes — Parakeet V3, SenseVoice Small, Whisper Small and Whisper Large V3 Turbo — are not Betas. Transcription is on-device in the Beta exactly as in the others: no account, no upload, works in airplane mode. If you would rather not run a Beta, Parakeet V3 remains the default and nothing changes for you.
What are the system requirements for Qwen3-ASR 1.7B?
An Apple silicon Mac on macOS 15 or later, with 16 GB of memory recommended, plus about 2.5 GB of disk for the model. That is higher than the rest of Whisper Notes for Mac, which runs on macOS 14 and later. In our benchmark run the model peaked at 2.43 GiB of process memory, the highest of the five engines. On an 8 GB Mac, Whisper Small covers the same 101-language list in 600 MB and Parakeet V3 covers 25 European languages in 465 MB.
Is this the same Qwen3-ASR that SenseVoice replaced in 1.5.0?
No, it is a larger model from the same family. The engine removed in Direct Download (DMG) 1.5.0 was Qwen3-ASR 0.6B, running autoregressively on the CPU through sherpa-onnx, which took 224 seconds on a 27-minute Chinese podcast. Qwen3-ASR 1.7B is roughly three times the parameters and runs on the GPU through Apple MLX at 8-bit. SenseVoice Small is still in the app on every channel and is still the faster pick for Chinese, Japanese and Korean.
Does the audio leave my Mac when I use Qwen3-ASR?
No. Alibaba also offers Qwen3-ASR as a cloud service through the DashScope Real-time and FileTrans APIs, where audio is uploaded to their servers. Whisper Notes does not use those. It runs the open weights locally as an MLX 8-bit model on your Mac's GPU, with no account and no API key, and transcription works with the network off. That is true of all four engine families in the app.
Why do these numbers not match the accuracy I get on my recordings?
Because FLEURS is short, clean, read speech, and your recordings are not. Our run is a calibration on a public dataset: about 15 minutes per language, one M5 MacBook Air, the same clips for every model. It is useful for asking whether a model is plausible for a language, and it is not a prediction of your accuracy on a meeting, on noisy audio, on accented speech or on a two-hour file. Read-speech rankings have failed to survive real audio twice in our own testing, so we publish them as calibration rather than as a verdict.
Try it
Qwen3-ASR 1.7B is in Whisper Notes for Mac 1.6.0 and later, Direct Download (DMG) only, as a Beta. Settings, then Transcription Model. The free trial covers 10,000 words, which is enough to run your own language through it and disagree with the table above.
If you find a language where our numbers do not match your experience, we would like to hear about it. Email mac@whispernotes.app. Full release notes: whispernotes.app/changelog.
Sources for everything above: our FLEURS run on the Google FLEURS test split at revision 70bb2e84, the Qwen3-ASR 1.7B model card, and the per-engine language table on our language support page, which is generated from the app's own source.
Download for macOS