Show HN: Aqua Voice 2 – Fast Voice Input for Mac and Windows

Hey HN - It’s Finn and Jack from Aqua Voice (https://withaqua.com). Aqua is fast AI dictation for your desktop and our attempt to make voice a first-class input method.

Video: https://withaqua.com/watch

Try it here: https://withaqua.com/sandbox

Finn is uber dyslexic and has been using dictation software since sixth grade. For over a decade, he’s been chasing a dream that never quite worked — using your voice instead of a keyboard.

Our last post (https://news.ycombinator.com/item?id=39828686) about this seemed to resonate with the community - though it turned out that version of Aqua was a better demo than product. But it gave us (and others) a lot of good ideas about what should come next.

Since then, we’ve remade Aqua from scratch for speed and usability. It now lives on your desktop, and it lets you talk into any text field -- Cursor, Gmail, Slack, even your terminal.

It starts up in under 50ms, inserts text in about a second (sometimes as fast as 450ms), and has state-of-the-art accuracy. It does a lot more, but that’s the core. We’d love your feedback — and if you’ve got ideas for what voice should do next, let’s hear them!

withaqua.com

・

139 points

・

the_king

・

9 days ago

89 comments

idk1 ・ 9 days ago

I’ve been using this for some time and I have to say it is fantastic. I’m intentionally not writing this with Aqua but by hand and it is taking so much longer. This to me feels like what Apple Intelligence could be, it is so much better than stuff all of the big tech is doing. For example, if you tell Siri voice dictation to go back and delete something what Siri will do is just write out “go back and delete something“ also if you tell Siri to go back and spell a name differently all Siri will do is write out the letters that you said to go back and type out. Honestly, for voice dictation software it feels like travelling to another planet in terms of improvement.

niel ・ 9 days ago

Real-time text output à la Apple Dictation with the accuracy of Whisper is something I've been looking for recently - I'll definitely give Aqua a spin.

MacWhisper [0] (the app I settled on) is conspicuously missing from your benchmarks [1]. How does it compare?

[0]: https://goodsnooze.gumroad.com/l/macwhisper

[1]: https://withaqua.com/blog/benchmark-nov-2024

  • the_king ・ 9 days ago

    We're more accurate and much faster than Mac Whisper, even their strongest model (Whisper Cpp Large V3).

    For that benchmarking table, you can use Whisper Large V3 as a stand-in for Mac Whisper and Super Whisper accuracy.

    • pbowyer ・ 9 days ago

      I've been using Superwhisperapp for over a year and I get nothing like the error level your comparison table suggests. Which model were you using with it?

      Aqua looks good and I will be testing it, but I do like that with superwhisper nothing leaves my computer unless I add AI integrations.

aylmao ・ 9 days ago

This is super impressive, great job!

Side-comment of something this made me think of (again): tech builds too much for tech. I've lived in the Bay before, so I know why this happens. When you're there, everyone around you is in tech, your girlfriend is in tech, you go to parties and everyone invariably ends up talking about work, which is tech. Your frustrations are with tech tools and so are your peers', so you're constantly thinking about tech solutions applicable to tech's problems.

This seems very much marketed to SF people doing SF things ("Cursor, Gmail, Slack, even your terminal"). I wonder how much effort has gone into making this work with code editors or the terminal, even though I doubt this would a big use-case for this software if it ever became generally popular. I'd imagine the market here is much larger in education, journalism, film, accessibility, even government. Those are much more exciting demos.

  • the_king ・ 9 days ago

    thanks!

    I share the same sentiment. I remember thinking in college how annoying it was that I was reading low-resolution, marked-up, skewed, b&w scans of a book using Adobe Acrobat while CS concentrators were doing everything in VS Code (then brand new).

    but we do think voice is actually great with Cursor. It’s also really useful in the terminal for certain things. Checking out or creating branches, for example.

fxtentacle ・ 9 days ago

This looks like it'll slurp up all your data and upload it into a cloud. Thanks, no. I want privacy, offline mode and source code for something as crucial to system security as an input method.

"we also collect and process your voice inputs [..] We leverage this data for improvements and development [..] Sharing of your information [..] service providers [..] OpenAI" https://withaqua.com/privacy

  • FloatArtifact ・ 9 days ago

    Local inference only is an absolute requirement. It's not even really all that accessible if it's online only. I can say this as someone that's used over 20000 hours worth of voice dictation and computer control.

  • canada_dry ・ 9 days ago

    First thing I looked for and read: the FAQ.

    No mention of privacy (or on prem) - so assumed it's 100% cloud.

    Non-starter for me. Accuracy is important, but privacy is more so.

    Hopefully a service with these capabilities will be available where the first step has the user complete a brief training session, sends that to the cloud to tailor the recognition parameters for their voice and mannerisms... then loads that locally.

    • oulipo ・ 8 days ago

      A similar but offline tool is VoiceInk, it's also open-source so you can extend it

  • pokstad ・ 9 days ago

    This should be on the FAQ. I was trying to find out if it was 100% processed locally.

  • jmcintire1 ・ 9 days ago

    fair point. offline+local would be ideal, but as it stands we can't run asr and an llm locally at the speed that is required to provide the level of service we want to.

    given that we need the cloud, we offer zero data retention -- you can see this in the app. your concern is as much about ux and communications as it is privacy

    • fxtentacle ・ 9 days ago

      The problem if you actually need the cloud is that it kind of completely destroys your business model. OpenAI is bleeding money every month because they massively subsidize the hosting cost of their models. But eventually they will have to post a profit. And then if they know that your product is completely dependent on their API, they can milk you until there's no profits left for you.

      And self-hosting real-time streaming LLMs will probably also come out at 50 cents per hour. Arguing a $120/month price for power users is probably going to be very difficult. Especially so if there is free open-source alternatives.

    • mrtesthah ・ 9 days ago

      MacWhisper does realtime system-wide dictation on your local machine (among other things). Just a one-time fee for an app you download -- the way shareware is supposed to be. Of course it doesn't use MoE transcription with 6 models like Aqua Voice, but if you guys expect to be acquired by Apple (that is your exit strategy, right?), you're going to need better guarantees of privacy than "we don't log".

      • shinycode ・ 9 days ago

        I downloaded the turbo whisper model optimized for Mac, created a python script that get the mic input and paste the result. The python script is LLM generated and it works with pushing a key. For 80% of the functionality for free and done locally.

  • toddmorey ・ 9 days ago

    And man it's another monthly subscription. I'm not mad at them for finding a gap in the market and putting a business around it. I'm mad at Apple for leaving that gap... hopefully built in voice dictation improves quickly.

    • FireBeyond ・ 9 days ago

      Is there a gap in the market? It's being rapidly filled with the likes of MacWhisper, etc., which offer local-only, one-off pricing.

    • pablopeniche ・ 8 days ago

      "hopefully built in voice dictation improves quickly." I would not hold my breath on that one lol

  • jackthetab ・ 9 days ago

    Agreed.

    This is where I bounce (out of this discussion).

  • thmsmlr ・ 9 days ago

    I totally agree, I created BetterDictation (.com) exactly because of that. Offline was a super important requirement for me.

jrvarela56 ・ 8 days ago

Feedback: I use MacWhisper and Tiny wisperkit model (english only) is way faster than any cloud service on my M1 macbook pro.

I’d say local is necessary for delightful product experience and the added bonus is that it ticks the privacy box

alxlu ・ 9 days ago

I’ve been using this for a while now and I really enjoy it. I ran into a semi-obscure bug and emailed them and they basically fixed it the same day.

I do wish there was a mobile app though (or maybe an iOS keyboard). It would also be nice to be able to have a separate hotkey you can set up to send the output to a specific app (instead of just the active one).

  • the_king ・ 9 days ago

    thanks! We're working on iOS, but it's tough to get the ergos right given all of Apple's restrictions and neglected APIs.

rkagerer ・ 9 days ago

You mentioned it "lives on your desktop". How does licensing work, and can you install and use it on a machine without internet access?

rickydroll ・ 8 days ago

I've been using Aqua since it was announced on HNN. I've survived the teething pains by using a mixture of Aqua and Dragon, depending on what I was doing. With this new Windows app, I've given up using Dragon for anything.

Things I've learned are:

1. It works better if you're connected by Ethernet than by Wi-Fi.

2. It needs to have a longer recognition history because sometimes you hit the wrong key to end a recognition session, and it loses everything.

3. Besides the longer history, a debugging mode that records all the characters sent to the dictation box would be useful. Sometimes, I see one set of words, blink, and then it's replaced with a new recognition result. Capturing would be useful in describing what went wrong.

4. There should be a way to tell us when a new version is running. Occasionally, I've run into problems where I'm getting errors, and I can't tell if it's my speaking, my audio chain, my computer, the network, or the app.

5. Grammarly is a great add-on because it helps me correct mis-speakings and odd little errors, like too many spaces caused by starting and stopping recognition.

When Dragon Systems went through bankruptcy court, a public benefits corporation bid for the core technology because it recognized that Dragon was a critical tool for people with disabilities to function in a digital world.

In my opinion, Aqua has reached a similar status as an essential tool. Well, it doesn't fully replace Dragon for those who need command and control (yet). The recognition accuracy and smoothness are so amazing that I can't envision returning to Dragon Systems without much pain. The only thing worse would be going back to a keyboard.

Aqua Guys, don't fuck it up.

replete ・ 8 days ago

Product/UI looks good. Nice job. I would pay for a completely offline version of this, cloud voice data is non-starter for me though unfortunately

  • voltaireodactyl ・ 8 days ago

    Check out MacWhisper which is one time payment and does this among many other things.

willwade ・ 9 days ago

You’re real market you need to go hard on is the assistive tech market. You know the biggest companies in this space are those solving problems for dyslexia where govt grants in eg UK fund pretty much all their work? I had an access to work assessment and they recommend like sweets stuff from texthelp. It’s then paid for by the government following these assessments. But it’s crap. It literally is a crap tool for adhd or dyslexia because these users literally CANT remember or deal with barriers like learning how to dictate correctly. Aqua voice solves this. I’m your biggest fan. I recommend it in my AT assessments all the time :)

  • waveringana ・ 8 days ago

    yes I really hope a lot of these ML startups check out the history of ML tech a bit more because so many accessibility tools are built via ML but theyve been abandoned

adamesque ・ 9 days ago

I was very delighted by Aqua v1, which felt like magic at first.

But I’ve noticed/learned that I can’t dictate written content. My brain just does not work that way at all — as I write I am constantly pausing to think, to revise, etc and it feels like a completely different part of my brain is engaged. Everything I dictated with Aqua I had to throw away and rewrite.

Has anyone had similar problems, and if so, had any success retraining themselves toward dictation? There are fleeting moments where it truly feels like it would be much faster.

  • SCdF ・ 9 days ago

    I use my (work) computer entirely with my voice, and it takes a lot of effort to work out what to actually write and to not ramble. Like you I've found that it's better to throw out words in sort of half sentence chunks, to give your brain time to work out what the next chunk is.

    It's very hard, and I wouldn't do it if I didn't have to.

    (which is why I'm always perplexed by these apps which allow voice dictation or voice control, but not as a complete accessibility package. I wouldn't be using my voice if my hands worked!)

    It's also critically important (and after 3-4 years of this I still regularly fail at this) to actually read what you've written, and edit it before send, because those chunks don't always line up into something that I'd consider acceptably coherent. Even for a one sentence slack message.

    (also, I have a kiwi accent, and the dictation software I use is not always perfect at getting what I wanted to say on the page)

    • e12e ・ 9 days ago

      Curious about your current setup, and if maybe adding a macro/functionality to clean up input via an LLM would help?

      In my experience LLM can be quite forgiving when given some unfinished input and asked to expand/clean up?

  • noahjk ・ 9 days ago

    Same here. My two biggest hurdles are:

    1. like you mentioned, the second I start talking about something, I totally forget where I'm going, have to pause, it's like my thoughts aren't coming to me. Probably some sort of mental feedback loop plus, like you mentioned, different method of thinking.

    2. in the back of my mind, I'm always self-conscious that someone is listening, so it's a privacy / being judged / being overheard feeling which adds a layer of mental feedback.

    There's also not great audio clues for handling on-the-fly editing. I've tried to say "parentheses word parentheses" and it just gets written out. I've tried to say "strike that" and it gets written out. These interfaces are very 'happy path' and don't do a lot of processing (on iOS, I can say "period" and get a '.' (or ?,!) but that's about the extent).

    I have had some success with long-form recording sessions which are transcribed afterwards. After getting over the short initial hump, I can brain-dump to the recording, and then trust an app like Voice Notes or Superwhisper to transcribe, and then clean up after.

    The main issue I run into there, though, is that I either forget to record something (ex. a conversation that I want to review later) or there is too much friction / I don't record often enough to launch it quickly or even remember to use that workflow.

    I get the same feeling with smart home stuff - it was awesome for a while to turn lights on and off with voice, but lately there's the added overhead of "did it hear me? do I need to repeat myself? What's the least amount of words I can say? Why can't I just think something into existence instead? Or have a perfect contextual interface on a physical device?"

  • the_king ・ 9 days ago

    I think Aqua v1 had two problems:

    1. The models weren't ready.

    2. The interactions were often strained. Not every edit/change is easy to articulate with your voice.

    If 1 had been our only problem, we might have had a hit. In reality, I think optimizing model errors allowed us to ignore some fundamental awkwardness in the experience. We've tried to rectify this with v2 by putting less emphasis on streaming for every interaction and less emphasis on commands, replacing it with context.

    Hopefully it can become a tool in the toolbox.

    • adamesque ・ 8 days ago

      Looking forward to giving it another try!

  • jmcintire1 ・ 9 days ago

    Imo it is a question of right tool for the right job, adjusted for differences between people. For me, the use case that made our product click was prompting Cursor while coding. Then I wanted to use it whenever I talked to chatgpt -- it's much faster to talk and then read, and repeat.

    Voice is great for whenever the limiting factor to thought is speed of typing.

  • cloogshicer ・ 9 days ago

    I'm exactly the same. Aqua is so incredible and I really tried to like it, but I just can't get my brain to think of what I want to say first, I have to pause to think constantly.

SCdF ・ 9 days ago

I currently use Talon, which I note is not in your benchmarks.

I can't find any documentation on how Aqua works, or how it compares, so I'm not sure it's meant to be a replacement / competitor to Talon? What are you configuring? How are you telling it that you like "genz" style in Slack? Can I create custom configurations / macros?

One thing I like about Talon is it's not magic. Which maybe is not what you're going for. But I am giving it explicit commands that I know it will understand (if it understands my accent obvs), as opposed to guessing and constructing a human language vague sentence and hope that an llm will work it out. Which means it feels like something I can actually become fast with, and build up muscle memory of.

Also that it's completely offline, so I can actually run it on a work computer without my security folks freaking out.

  • the_king ・ 9 days ago

    We're building something different, but there is some overlap. Aqua is built for max speed, while keeping accuracy high. To achieve that, inference runs in a datacenter (for now).

    You can customize Aqua using custom instructions, similar to ChatGPT custom instructions, and get some Talon functionality from it:

    In my own, I have:

    1. Breaking the paragraphs with three or four sentences.

    2. Don't start a sentence with "and".

    3. Use lowercase in Slack and iMessage.

    4. Here are some common terminal commands...

  • willwade ・ 9 days ago

    Aqua voice is nothing like talon. I wouldn’t bother trying to compare. It’s a dictation tool. Just entry. Not commands. But it’s bloody impressive. You don’t need to learn anything - you just talk like you would talk to someone across the way from you

    • SCdF ・ 9 days ago

      Oh, from the video I got the impression it was more than that, based on it recognising app contexts and the like. I guess that's mostly just icing on the cake for the core dictation part.

      • pablopeniche ・ 9 days ago

        >recognizing app contexts

        Users have different preferences on the text format they input into different apps. Aqua is able to pick up on these explicit and implicit preferences across apps – but no "open XYZ app" commands, yea

TylerE ・ 9 days ago

I will have to look into this. I am currently in the process of going on disability as I cannot work due to (amongst other things) carpal and cubical tunnel in both arms.

roland_kovacs ・ 8 days ago

Hey guys, great idea! Most of the apps already have voice recognition. Do you think about serving a niche where this feature is not existing? Also, the data protection part is unclear to me, I don't want everything uploaded to a cloud where I don't know what is happening with it.

oulipo ・ 9 days ago

Interesting!

A nice open-source alternative is VoiceInk, check it out: https://github.com/Beingpax/VoiceInk

do you also plan to open-source part of your platform?

  • razemio ・ 9 days ago

    I just tried it on a M4 Max MacBook Pro. When you have such a processor, it seems to be even faster than Aqua Voice 2, does more, optional supports open router AND is open source? Thank you so much for the recommendation!

    • pablopeniche ・ 8 days ago

      >it seems to be even faster >runs locally

      This is obviously a lie. If this was true, all the inference provider companies would go to zero. I support open-source as much as the next guy here, but it's obvious that the local version will be slower or break more often. Like, come on guys. Be real.

      To illustrate this, M4 Max chips do 38 TOPS in FP8. An NVIDIA H100 does 4,000 TOPS.

      Prakash if you're going to bot our replies, at least make it believable.

      • razemio ・ 7 days ago

        I am not Parkash. Just check my profile. I am not a bot. github.com / razem-io. I checked his youtube videos. He seems to lack presentation skills but his app is very usable at its current state.

        I have both apps open. The STT seems to be faster with VoiceInk. Like it is instant. I can send you a video if you want.

        I am sorry. I did not want your product to look bad. You are right you still need to offload the llm part to openrouter and the like if you want this to be fast too. However, having the ability to switch AI on/off on demand and context aware with custom prompts is perfect. It can use ollama too. Yes this will be much slower but local. Best off both worlds. No subscription, even if you use cloud ai.

      • joshjob42 ・ 8 days ago
        ・ 3 more

        I mean, it's just doing STT, it's not trying to run inference on a frontier model, and if you have to connect to a remote datacenter vs your own machine that could add latency that slows down the experience. There are also tradeoffs in deployment in terms of latency vs throughput that doesn't occur on your local machine that only ever is doing STT on one thing at a time.

        I'm skeptical of the claim that local is literally faster, but it's not impossible in the way you suggest.

        • pablopeniche ・ 8 days ago

          If you watch the video you'll see that Aqua does more than just STT and why we need the frontier models. I think if you want just STT of what you said verbatim, then yea any local version of Whisper is great. If you want the enhancements shown in the video, you need the frontier models. And doing that locally is slow.

    • oulipo ・ 8 days ago

      And it processes stuff locally on your device!

  • pablopeniche ・ 9 days ago

    Just tried it and it crashed

    • oulipo ・ 8 days ago

      Then submit a report, but for 100s of people using the open-source, free and extensible version, it works great!

      For me with a small (but very accurate) whisper model it can be even faster than AquaVoice, and works locally (which is really important)

somberi ・ 8 days ago

I use MacWhisper and it works well enough for me to stop looking for options.(MBA M2 24GB Ram - Large V3 English)

I wouldn't feel comfortable if someone were looking over my shoulder while I'm typing at a coffee shop.

I am not your customer.

bemmu ・ 9 days ago

I would like to try this, but I use Synergy to use two computers with the same keyboard. I have Aqua Voice now on the server computer, would be great if I could input text to the client computer using it as well.

qntmfred ・ 9 days ago

I use the built-in voice typing in Windows and am pretty happy with it. How would you say this compares (presuming most of your comparisons are mac-centric)

  • the_king ・ 9 days ago

    Aqua is in another league when it comes to accuracy. I just ran them side by side on a simple q to ChatGPT and here were the results...

    Aqua Voice

      What is the first recorded eclipse in human history? I'm not asking when the first one occurred, but the first written record we have of an eclipse.
    Windows Voice Typing (v11 24H2, Dell XPS 13 9340)

      What is the first recorded eclipse in human history i'm not asking 1 like the first occurred but the first ridden record we have of an eclipse
    Windows mistakes were:

    -"1" should be "when"

    -"ridden" should be "written"

    -No punctuation

aminsadeghi ・ 9 days ago

Is there going to be Linux support at some point?

  • the_king ・ 9 days ago

    We can do that, with some help from the community.

    • emacsen ・ 9 days ago

      I'm a paying customer and I signed onto Aqua Voice shortly after your demo on HN.

      My experience with it has been overall positive but mixed. I enjoy using it for dictation, but I found that issuing editing commands and having them recognized/executed often took a lot longer than making an edit myself (which I can't do while in dictation mode).

      But as a paying customer, seeing you go in this direction is somewhat sad/frustrating. You're abandoning the product I use, and you're saying that if I want to see my platform supported, I or someone from the community has to provide it- for a fully proprietary paid application.

      I understand that I'm a minority user, but it's a bit disappointing to read this.

      • the_king ・ 8 days ago

        Totally understand, thanks for being a customer. I'm sorry we weren't able to make the web version as smooth as we wanted to.

        We do plan to support Linux. This was probably a little bit of a blind spot for us - not realizing that anyone running a Linux desktop doesn't even have system voice tool to fall back on.

    • inciampati ・ 9 days ago

      What support do you need?

tomblomfield ・ 9 days ago

I recently started using Aqua and it's great. The team really improved the latency in the last few weeks.

hu3 ・ 9 days ago

How does it compare to https://wisprflow.ai ?

btw, grats!

  • the_king ・ 9 days ago

    Thanks!

    We're faster, more accurate, and have a streaming option. Aqua can go from key-up to paste in as little as 450ms. Flow was closer to 1000 in our tests.

    Overall, you'll notice we make a few more tweaks to the output than Wisprflow.

    For example, Aqua + Cursor is very powerful - we syntax highlight your transcript. The easiest way to see this is to use streaming mode (double press Fn) + deep context + cursor and try asking it to change something.

    This also works in other "context rich" environments.

    • redcanvas ・ 9 days ago

      Hey, love the what you are building in this category. I've been using a competing product which you know very well about. They advertised about how you can improve your work per minute by dictation, which was the main draw for me because I do a founder. There's a lot of managerial work that I'm doing.

      It has been a godsend in terms of increasing my productivity because I no longer have to type. I think your product's accuracy and latency shortening just make this even better. I often use it and then find out, "Hey, I need to make some changes," and I need to re-edit some of the stuff, which reduces the WPM productivity amount. So I think accuracy is definitely key here. Key metric to differentiate a product.

      I am pushing this to other colleagues to get them to adopt. One challenge people are saying is that. One is that some people may not be as organized (you know, they might be a lot more organizationally structured in their mind). So for them, they're having trouble - they'd like to write things out, and by the time if things go out of their mouth, you know it's already formulated logical thought. Whereas you know people like me are a lot more verbal vomit type of person. For me it's huge because I say a lot of um like in all the other things I just dump stuff out and then organize it later.

      Whereas other people organize stuff in their brain and then dump the information out. So people who do a lot of coordination and just you know so I feel like this could be two different segments to take into account.

      Another one that's been fantastic is that we have multilingual colleagues who are speaking in Mandarin or something else and then they speak it and then ask Flow to be sent to translate it to a different language. That part I think has been fantastic.

      I think the ability to edit what you wrote with AI is going to be the next key feature. Providing the context in the window is all wiithin the conversation right? For example, you just ask after because what you write out is not the final and you need to do a lot of editing and formatting. Sometimes when you say too much stuff, it's just like a huge jumble paragraph with a lot of fluff words. Make it clear, concise, trim non-effective words. I think those are a key feature because it's not about your productivity, it's about other people being able to ingest your information efficiently. At least that's what I look at from a managerial perspective.

      To give you an example, everything I laid out above came from dictation. You can see how this is inefficient. There's a lot of inefficiencies here.

      • redcanvas ・ 9 days ago

        A feature that would be great is similar to how you can write snippets in all the other tools where you can say "calendar" or "cal" and then it gives you the link. If this is something possible, I think that would make this fantastic.

        Another feature that would be great is actually being able to have a conversation with an AI model first and then refine the output iteratively until you're ready and then pipe all that over. The ability to have a chat is very good or do this all through voice.

    • hu3 ・ 9 days ago

      Thanks! I'm tempted to try Speech to Text + Cursor/Copilot for development. It's probably the future since most people can speak faster than they can type.

      • jmcintire1 ・ 9 days ago

        This use case is great, even for people who haven't been interested in dictation before

bklyn11201 ・ 9 days ago

Music playing on Youtube in Chrome, Airpods in, the desktop and the sandbox/demo just don't work.

  • the_king ・ 9 days ago

    Shoot, might be due to AirPods mic init latency. AirPods work well on Desktop (though their mic quality isn't the best).

iAMkenough ・ 8 days ago

Tried viewing the Pricing link in the footer, but it requires a Google account to view.

waveringana ・ 8 days ago

will we ever see local, open source models? they are very important for accessibility reasons which this product can fit into, but wont because of it being cloud based (and proprietary).

hasperdi ・ 8 days ago

Anyone can recommend a good dictation app on Linux?