Whisper is the wrong model to benchmark against, or rather, there are better models that are state of the art now like Nemotron and Parakeet both by Nvidia, as well as Mistral's Voxtral and Cohere Transcribe.
However, what's funny is, RIP to a lot of the paid apps that simply wrap Whisper, I'm sure Apple will make a native GUI such as a recorder app for macOS that obviates the need for these wrappers, which everyone seems to be vibe coding these days.
Also, this test is English-only, while a strong point of other models is to understand different languages without first having to say which one (so you don't need 3 different keyboard shortcuts if you wanna dictate in 3 languages day-to-day)
Reminds me of the time my neighbours must have wondered if I was having some kind of a breakdown when trying out really basic MacOS voice recognition back in the early 2000s. There was a keyboard shortcut and you could say something like "phone number for firstname lastname" and it would theoretically show you that phone number. Thing is it didn't seem to like a British accent, so I spent a good hour trying out different accents, rotating through various US accents, Australian, South African, Canadian and so on. It seemed to respond best to some kind of a melange of Californian / Australian.
Not too far off what happened, although thankfully I wasn't actually trying to do anything other than test it. Going to take the opportunity afforded by Scottish TV comedy here, and make a very tenuous link to intercultural exchange so I can post my favourite Rab C Nesbitt scene, hands across the sea indeed: https://www.youtube.com/watch?v=uKxPH_QH940
Interesting - I don’t think I’ve ever seen anyone from the UK refer to talking in a “British accent” before since we are normally aware of the wild regional variations.
Fair point! I think it's a tic from being English and having lived in Scotland for quite a while so I autocorrect "English" to "British", but I've over-corrected here. (Also perhaps something to do with "British English").
I use Systran/faster-whisper-medium for real-time subtitling, but you need to get used to the context it's used it and the weirdness it translates into. Parakeet has great mandarin>CN text, but running that + a translation model has been tricky and I never got it fast.
As an Australian, Apples voice models have always sucked. I've tried using stt (again) more recently and its improved, but i'm so tired of having to Americanize my voice to get it to figure out what the hell i'm saying.
As a Texan first, American second, I sympathize with this statement. Siri can't understand me probably 25% of the time. I use STT for iMessage while in the car, and half the time it will take 3+ times to either get it right or me give up, and hope to remember to text them by hand when I next stop.
The Two Yoots problem. Do you use d's in place of t's such as dees/dems/dose/dere? I have a heavy queens accent so you'll hear me say things like "deres tree uh dem ova dere."
> there are better models that are state of the art now like Nemotron and Parakeet both by Nvidia
Is parakeet state of the art? It always transcribes speech fragments for me, like if I stutter and say "m-m-m-map" parakeet will dutifully transcribe "m m m map". Which I guess could be a good thing or a bad thing depending on what you want. Whisper does not do that however.
I think that's parakeet doing its job there. That is a closer reflection of what you've actually said. The trick is then throwing that output through some additional deterministic and non-deterministic steps to tidy it up however you prefer. It's exactly what I do with my free and open source dictation app (dictator.robgough.net) for Mac+iOS. And of course, everything stays entirely on-device. Gemma E4B is really wonderful for that second step, it's great at language – but takes up 6-7GB RAM.
Parakeet is certainly faster on my machine (m3 max), but I can't stand using it over Whisper for dictating my prompts. It makes a lot more mistakes, possibly because (like you mentioned) large portions of the speech will pause / stutter as I think about what to include.
With whisper v3 turbo, I can almost always live with the few mistakes because the overall stream-of-thought word-salad I provide is still clear at a high level. The bits and pieces of context seem to help, that I might leave out if typing and focused more on traditional conciseness / clean writing. With parakeet, I needed to do frequent editing even for shorter bits of speech.
I realize some applications prioritize the latency.
It sounds like post processing should be the job of an LLM. I would like the voice model to be faithful to what was said and then that output can be smoothed over or postprocessed as needed for the use case
To be clear, I'm talking about high word error rate with parakeet vs whisper, not post processing and cleaning up my speech. Re: being faithful to what was said, one small example, Whisper will often put ellipses when I pause.
Whisper v3 is still the best (by far) when it comes to poor quality input (say background audio from a security camera), though remains more susceptible to hallucination so it's a bit of a tradeoff.
From my experience, Speech-to-text falls way short of Wispr flow and I would assume the ones that are said to be better than that. It lacks context awareness and formatting
Yeah, apple will be optimizing a model to work on ANE and then turn it into a native app. My only hope is that it has a reasonable api so that I can use that as a generic input source across iOS / macOS that’s equivalent to the ubiquity of the keyboard.
> RIP to a lot of the paid apps that simply wrap Whisper
I started using a few open source apps for transcription and eventually subscribed to a paid one...
On paper, it's not hard to compete, but for this use case, a few rough edges make it really frustrating to use. Like a keyboard that sometimes doubles the letter "e"
Automatic dictionary, seamless language switch, no issues with accents, etc... Putting the effort in the last mile makes a world of difference.
If anyone has better options, I'm willing to have a look. The best open source solution I found was Handy, and I currently use Wispr Flow
I built my own because I was frustrated with a lot of the free options. Largely because a lot of them had an upsell to be able to do the secondary post-processing step with an LLM. And it wouldn't pick up things like emojis properly or say numbers. Because of that, I left quite a lot of options in there for customising and adding additional steps, etc. Feel free to take a look: dictator.robgough.net
My initial Mac version actually had three additional steps that you could toggle, obviously at the cost of some speed. That is what the website talks about, although nowadays for my own use I've reduced that to just one step and found that it's pretty great. I've got a new version in test to tidy that up, but still lets you define as many steps as you want.
I don’t know how Apple divides computation between the GPU and the Neural Engine, but one major benefit, especially for real-time transcription on laptops, is the improved power and thermal efficiency. I noticed better accuracy after switching my app to SpeechAnalyzer, and I suspect part of that improvement for me came from the microphone no longer having to compete with jet-engine fan noise.
Thanks, was looking at a better diarization model.
Even for those sorts of apps, MacParakeet which I've been using is FOSS so no payment needed. In reality these days with AI the ability to spin up a free and/or OSS competitor falls to zero.
I'm using it offline. But it's much faster than realtime so it should be usable for streaming. I just asked Codex / Sol to integrate FireRedVAD with Whisper...
Thanks, I didn't see vanch007 version at first (only ~30 downloads), I usually look at mlx-community. For the size I was looking at the wrong model (TTS not transcribe-diarize), thanks for the corrections.
The canadian government will provide lots of historical data for curious citizens, many of which are recordings of interviews from decades and decades ago. For a book project this allows me to make a hours of audio searchable through a GUI application I have developed that has a voxtral backend.
> However, what's funny is, RIP to a lot of the paid apps that simply wrap Whisper, I'm sure Apple will make a native GUI such as a recorder app for macOS that obviates the need for these wrappers, which everyone seems to be vibe coding these days.
What's insane to me is that you have all of these low-quality me-too apps, and literally no one could bother to read the damn Human Interface Guidelines or follow iOS design conventions.
Doing so is literally LESS WORK than trying to make your own custom awful iOS UI.
Not if your app is a Web wrapper, which so many of these are.
If you use SwiftUI (the native recommendation by Apple), it severely penalizes you, if you want to paint outside the lines (which is a big reason that I don't use SwiftUI for shipping apps). It's insanely easy to write a native app that is 100% in line with HIG.
This particular product used Whisper, so that was obviously the right model to compare it against. Further this is explicitly on device, and Nemotron 3.5, as one example, is 2.5GB for the model.
And if someone were broadly comparing all on-device models (instead of just looking at how this new on-device ones compares to what a specific product uses), Nemotron 3.5's WER are actually a bit higher than what they report for SpeechAnalyzer, for both tests.
Just ran it against Whisper-Large-V2 on a math lecture (my primary use case for ASR is subtitling math lectures), and it was substantially faster and only slightly worse. Very usable for live transcription though I'll probably stick with whisper for the time being since I don't really need the subtitles to be generated in real time.
Been using it for a podcast app I have been developing for half a year lol (I hope I publish it by version 27) and I can confirm it’s real fast.
Splitting the audio in multiple segments and firing it up without hitting the maximum limit of concurrent decoding streams makes it blazing fast. Fair enough you loose the cut, but it’s good enough for just podcast. In one minute it chews through one hour of audio. This on an iPhone 17 Pro.
You could perhaps run over the segment splitting points (plus a few seconds back and forward) in a second batch then merge the results in the end so you don't miss anything.
Nothing really, except that I get to play with SpeechAnalyzer APIs, foundation models, translations. It’s basically my playground where to try all things. Been listening a lot of Chinese podcasts lately, transcribed and translated by local models.
Edit: all that said, the app is irrelevant. What I want to say is that live transcripts on iOS using Apples frameworks works very well. Only thing I miss is diarization support.
I took a swing at bringing this into Handy.computer if anybody's interested: https://github.com/cjpais/Handy/discussions/1031 . Looks like there has been past demand for someone to implement it, but no proposed PRs. This article was inspiring.
I will plug Willow for mac recording. IMO it's basically to me a "better than perfect transcription" as it cleans things up and is almost instant. I liked Superwhisper but switched to Willow as it was a big difference.
Its so good that I'm not sure that it's possible to get any better. Speech to text seems like basically a solved problem, if not now then definitely in 5 years. I don't know if any of these speech to text businesses will work in the long run, but for consumers they are great. My guess is the 2030 version of Apple's SpeechAnalyzer will be so good that nobody will need to use 3rd party software.
What isn't solved is domain-specific jargon with these tools. When i talk to my coding agent, i want to be able to speak the names of symbols and files and have it be aware of that stuff, like having LSP integration.
If I say 'useSuspenseQuery' I want it to come out as useSuspenseQuery not 'use suspense query'. Even if I had to say 'symbol useSuspenseQuery' to give a hint that i'm referencing a symbol, that would be fine.
> Apple's new SpeechAnalyzer is the most accurate on-device speech engine we tested. It beat every Whisper model we ship, including Whisper Small, on both the clean and the noisy half of LibriSpeech, while running roughly three times faster than Small.
Vs Voxtral would be a better comparison. No other model, open or closed, has been able to hit such a low AER (Acronym Error Rate ;)) for my meeting transcripts. Seems to understand/infer all the technobabble I use at work. Never have to edit anything. Whisper was catastrophically bad.
I typically disable autocorrect on Apple products because of this, cautiously optimistic about their improved speech models, but definitely worried that it's going to 'correct' technical jargon to more common words.
Whisper small/tiny/base are almost four years old (they were not updated for Whisper v2 or v3). Is there really nothing better to benchmark against by now?
I have tried everything (that will run on a 12GB RTX 4070) and I have yet to find anything with better accuracy than Whisper V2 Large for my dataset (discord audio from TTRPG sessions, isolated per-speaker, mostly non-American accents)
And yea, Nvidia's Parakeet v3 is good enough for my own just local transcription most of the time.
When I need local transcription to be more reliable and I don't have the energy to proof read a long ramble, I still often just pop open chatGPT, dictate, cut, paste.
But we're pretty much already to the point where local transcription models can replace cloud ones for personal use. They're still a bit rough around the edges in terms of polish and latency, but plenty of people are fine with that to avoid yet another app subscription and not having to worry about wondering what's potentially happening with their data.
Any chance you can benchmark against whisper large and large v3 turbo? These run comfortably on older Macbooks and are still far more accurate in real life dictation compared to even the parakeet models( despite ASR leaderboards) with an RTF < 1.
Try MOSS-Transcribe-Diarize from a few days ago. I’m getting better results than those whisper models. And it’s very fast and small. Better suited to noisy audio too.
I run SuperWhisper on both my Mac (where is uses Whisper) and my iPhone (where it uses SpeechAnalyzer and have found that SA does indeed run faster and anecdotally more accurately. Super exciting!
Finally. I‘d be delighted though if they actually implemented language autodetection (like everywhere else) though. There’s little more frustrating in my day to day than having dictated half a page to find that it‘s complete gibberish because Apple forces you to select the right language first…
Same with the keyboard. Apple is completely incapable of taking context into account for the input mechanisms of the operating system.
If I start typing and the existing text is in Spanish, then a sensible default is to select the Spanish keyboard I have installed and let me adjust otherwise.
App developers should also be allowed to supply mini-dictionaries within a context to allow autocorrect to work correctly in that context, so for example in this thread [SpeechAnalyzer, API, Whisper, Parakeet, Nemotron] should be supplied so that these terms are autocorrected.
For my current purposes, I need a speech-to-text model/API to also emit word-level timestamps - for now, that makes ElevenLabs's Scribe v2 the best multiplatform, multi-language choice though it does look like this SpeechAnalyzer API provides them (although only for English).
I don’t like it written that way either, and it always seems like the type of number you put on a slide for a head of sales or something. It rankles because:
- it implies that error could be increased n-times, but a 15x _increase_ in 9% error would be an error rate of 135%, which is nonsensical.
- a reduction from 90% error to 20% error is clearly a bigger improvement in rightness to a reduction from 9% to 2%. One is “almost all wrong to almost all right”, the other is “more right”, but they are both a 4.5x reduction in error which means that the 4.5 quantity doesn’t have a constant meaning.
The answer is something like log odds ratios, but that introduces the additional need for a reader to know what that is, and that would be unusual.
Was intrigued by the Inscribe product so I downloaded it to test it out. I uploaded 1 voice file about 4 minutes long and was promptly given a message that I've reached the free plan limits.
Kind of a bait and switch. How can we test the product with such short time limits and what, exactly are you offering if all the processing is done on device by Apple?
Every single asr model I tested so far did not support timestamps properly though. Some use external aligner to create timestamp, but the accuracy is still much inferior than whipser in case the audio is noisy.
Cloud is so cheap and quick. I use local too but my api bill is like 3 quid a month. You would have to be very cheap or have compliance needs to tolerate the error gap
I make an iOS app that uses this API heavily for transcribing diverse audio of varying bitrate and recording quality. The audio often contains music, multiple speakers, sound effects. SpeechAnalyzer almost always gets it.
It can struggle with proper nouns but will return something phonetically similar.
My main gripe is that it requires a separate model download per language. I understand the why they did this (to save disk space). But it makes multi-lingual audio hard to transcribe unless you know ahead of time the languages in the audio.
As an app developer the biggest win from using Apple's model is I don't have to bundle it in my app so my app looks much smaller. If a user has many transcription apps each one could have their own model. If Apple's model is used only one copy is needed.
i wish they had benchmarked it against parakeet-unified-en-0.6b and cohere-transcribe-03-2026. i am using parakeet with https://handy.computer daily and it's amazing.
this is amazing. if i had a mac i would try to reverse engineer the code, extract the weights and port it to something that works on linux/windows like torch or burn. then put the code on github and weights on a torrent site. lifes too short to let apple keep their models exclusive.
This hasn't been tested in court. But there's a high chance that model weights are not copyrightable, only the code to generate them is.
Cloud models are usually protected by trade secret laws, leaking them would get you in trouble. However if the model is made available publicly, as long as you don't break the law to get them, anything after that would be fair game unless Apple can prove that humans have significant authorship over the weights, which hasn't been tested and is a significant burden to prove/disprove.
Copyright protects original forms of expression, not arbitrary data. It is very arguable whether it applies to model weights. However, it would likely constitute a license violation.
Aside from the legality of it, I think you are underestimating how complex it can be to do that. It is possible in theory but not something that will be a fun side quest like you are making it seem.
The Jedi Hand Wave-y nature of the way people talk about AI these days is going to make reigning in the AI superpowers nearly impossible. Because there are people out here who believe models of this quality are easily replicated or reverse engineered. Neither is really doable on any reasonable timeline by people who are not AI experts. Real AI experts. Not TF/PyTorch monkeys or Agent Slop Slingers.
And those people are already highly incentivized to not make anything performing better than SOTA models open source.
I stopped reading after seeing they compared only with Whisper Small, Base, Tiny
This is useless test and benchmark when you have these day Whisper-V3-Large and Whisper V3-Turbo that you can faster than realtime on 5 years old macbook on apple sillicon (ANE). They didn't even compared to parakeet v2 or parakeet v3. And only english language...
Im hoping Apple gets the new Siri working better on older phones. I was excited to use it but the latest beta / Siri runs too slow on my iPhone Pro Max 15.
Im looking for the same experience I have when talking to chatGPT. As for past two years or more talking to GPT within it's app and on my iPhone Pro Max 15 it runs smooth as butter :-). This is the experience I was and still am hoping with Apple, but Im thinking all the extra layers of privacy and security might be slowing them down?
Overall, Apple who is suing Open AI should just buy them and let me have the best conversational AI out there baked into my old ass iPhone. Because as so far the new Siri on my old phone (tho again GPT works great talking to it and for years) doesnt come close. It's the same old "Could you try that again," Siri. BOO!!!
Yeah, ChatGPT voice is great experience vs. Siri on that phone. In case you haven't done something like this already:
1. In Shortcuts app, make shortcut named "AI Voice Mode" (or whatever you want, YMMV)
2. Set it to run the ChatGPT action "Voice Mode" (requires at least the minimum paid tier, I think)
3. To trigger, say "Hey Siri, AI Voice Mode" (or whatever you called the shortcut)
This is a pretty slick integration, but yeah, if it were baked in that would be all the better.
Ridiculous that for past many years we can talk to GPT on our iPhones without any hiccups and this new Siri is still the same old horse crap (at least for me and this latest beta). Buy them already Apple or possibly be replaced by them as their path & trajectory (working on Ai devices now and they are stealing like Jobs did with Xerox) mirrors yours in the late 1970s.
Thanks for the tip and if Im not mistaken it's similar to asking Siri to ask chatGPT to ask XYZ?
> similar to asking Siri to ask chatGPT to ask XYZ
Effectively, it sort of does that, but really it just listens to the wakeword and opens/switches to the requested app & modality.
FWIW, I get a very different functional result using the Shortcut method vs. asking Siri to delegate natively. To compare, I asked Siri (non-beta here) now to "ask ChatGPT <x>" and I got a top-card with some fairly low quality SEO-ranked weblinks.
New Siri is impressive in that it answers satisfactorily now 80% of the time vs 10% with old Siri.
But it’s slow as shit. GPT, Claude, and Gemini can answer me in 5-10 seconds. Google AI Mode can answer in 2 seconds.
New Siri usually takes 25 seconds to respond to me. This morning it timed out (with strong network connection) when asked a simple multiplication question.
Author here. I ship both Apple speech engines plus WhisperKit side by side in a transcription app, which made it possible to run all five through identical production code on the same audio: LibriSpeech test-clean and test-other, 5,559 utterances, fully on-device on an M2 Pro.
Apple published no accuracy numbers for SpeechAnalyzer (or for SFSpeechRecognizer, ever, as far as I can tell), so the migration question has been guesswork. Short version: the new API cuts WER 3.5-4x vs the old one (2.12% vs 9.02% on test-clean), and it also beat Whisper Small on both splits at about 3x the speed. The old API came in last on clean speech, behind even Whisper Tiny.
On "why should I trust a vendor benchmark": the Whisper column reproduces OpenAI's published LibriSpeech WERs within +0.11 to +0.42 on all six measurements (same corpus, same normalizer, same scorer for every engine), and the raw per-utterance transcripts are downloadable from the article if anyone wants to rescore with their own normalizer.
Limitations worth stating up front: English only, read speech rather than meeting audio, one machine. Precise per-engine timing isn't in the article yet because the accuracy runs shared the machine with a dev workload; WER is load-independent, timing isn't.
Two things that might interest people migrating: SFSpeechRecognizer sends audio to Apple's servers unless you set requiresOnDeviceRecognition, and with SpeechAnalyzer, finishing your input stream is not enough to end a session. If you never call finalizeAndFinishThroughEndOfInput(), the results sequence never terminates and your await hangs forever. I found that one because it was shipping in my own app.
Happy to answer questions about the harness or the normalizer.
At this point, I would not recommend ignoring Parakeet TDT 0.6b v2/v3 (english-only versus multilingual). Those models have been out for a year, give or take, and they are both accurate and fast. I would choose Parakeet over Whisper in almost all situations these days. Parakeet works great even on my several year old iPhone 15 Pro Max, so if an app is going to ship a dedicated model, I strongly recommend investigating Parakeet.
On the more cutting edge front, Granite Speech 4.1 has proven to be a reliable workhorse for me, but it is larger than Parakeet. Cohere Transcribe is interesting, but how strong it is seems to vary more from task to task.
Parakeet Unified 0.6B came out a few months ago, combining both online streaming and offline transcription into one model, and that is one that I need to test more, but it seems promising.
As others have mentioned macOS 27/iOS 27 is supposed to have a new model, particularly on devices with 12GB of RAM or more. I have not actually seen the option to enable that new model yet, though, despite being on the beta on a device that meets the requirements. Maybe a benchmark would reveal that it is already active?
You can't reasonably expect generic ASR to infer tmux from "tee-mucks". "tee-em-you-ex" works reliably if you're ok with capitalisation for your use case.
If this isn't open source/weights and can't run locally, I don't see how this is a replacement for Whisper or other open models, e.g. within Home Assistant.
It's a local model so it's essentially open weight such that you could feasibly export it somehow since it's already on the laptop somewhere. Apfel is a wrapper app like ChatGPT but using Apple Foundation Models, I assume something similar will happen with this transcription model.
It's not open weight, but the point is to be an on device (and thus local, privacy preserving) option. The article mentions that as the caveat
> What this means if you just want good transcription
> If you are on a current iPhone or Mac, the best on-device transcription engine for English is already in the operating system, and the private option is no longer the compromise option
Presumably if you don't trust apple you wouldn't purchase their products and even if you were for example forced to use it via work or something you wouldn't use this feature ... so it doesn't really change the calculus as presented by this article - IF you ALREADY HAVE a MODERN Mac (and trust apple) this is your best option
You are being naive. An Apple device makes dozens of network requests every minute or so to Apple. It is neigh on impossible to verify what is being requested or sent. Also unplugging the Internet and verifying that something still works does not mean the app won't phone home behind your back when it can. These things are designed to fail silently.
There are many alternatives for trying to find out what’s going on. If you don’t want to bother, and most people don’t, well, what else is there to say?
It is generally a good idea to know what software is phoning home, if you can pinpoint it.
If you have any software recommendations, I’d be happy to know.
It's more complex than that. LittleSnitch or other application firewalls won't help when programs like package managers legitimately need to access the Internet. And figuring what what is being sent would require decrypting the traffic. There is no easy way to figure out if Apple or some rogue app is siphoning off your home directory.
The appeal is that users only have to download it once across all apps that use it. Instead of convincing a user to give a couple gigs for just your one app