Transcript
00:00:00This is Handy. It's an open source speech recognition and dictation app that runs
00:00:06completely offline with no cloud, no subscription, no account, and it lets you choose your own
00:00:12favorite dictation model. And I honestly think that this has now become my favorite speech-to-text
00:00:18dictation tool. So in this video, we'll take a look at Handy, see how it works, and we'll run a
00:00:24few tests to see how it performs and how it ranks up against heavyweights like the commercial
00:00:29whisper flow. It's gonna be a lot of fun, so let's dive into it. Up until now, I was kind of skeptical
00:00:39of speech dictation apps, and I just couldn't find one that I wanted to use. I didn't want to use any
00:00:45commercial products and pay for dictation, and I was also worried about the fact that using an open
00:00:51source tool with a local dictation model might hog up my CPU. But Handy actually changed my mind on both
00:00:58those issues. So first off, let's talk about how Handy is actually built and how it works. Handy was built by
00:01:04the developer CJ Pace with Tari, so it uses Rust for the backend and React and TypeScript for handling the
00:01:12settings UI on top. Transcribe CPP runs the whisper family models in GGM and GGUF format. Transcribe RS
00:01:20runs Parakeet. The CPAL library handles the cross-platform audio input. RDEV owns the global keyboard
00:01:28shortcuts, and it uses a library called Rubato, which provides real-time audio streaming and takes care
00:01:33of resampling. And this combination makes it super easy and super lightweight to use it even on my Macbook.
00:01:40And the flow itself is dead simple from the user side. You set a custom keyboard shortcut and then
00:01:46either toggle it on or off or hold it down for push to talk. And while you're holding it, there's a utility
00:01:52called Cilero VAD, which is quietly filtering out the silence in the background so you're not burning
00:01:58compute transcribing complete silence. And when you let go, Handy sends the audio to whichever model you
00:02:04picked and then pastes the transcribed text directly into whatever input is focused at that moment. And to be
00:02:11fair, Handy is not unique in this way. A lot of other dictation applications on the market work in a similar way.
00:02:17But the thing I love the most is that you can choose your own preferred model and each of the models
00:02:22listed in the model section has a speed and accuracy meter. So it's kind of like choosing your favorite
00:02:28character in Mario Kart. You can decide which metric is the priority for you. In my case, I found that
00:02:34using the Parakeet Unified 0.6 billion parameter model works very well for me because Parakeet is a CPU
00:02:42only model. And by their own numbers, it runs at around five times real time speed on a mid range
00:02:48i5 chip. And my M2 Max is way past that spec. So I don't even notice it when I'm running it in the
00:02:54background. And there are so many other options to choose from. They even have models that are specific
00:02:59to one language. For example, we have some Russian ones, some Ukrainian ones. And I love the honesty
00:03:05coming from Handy's mission statement. Handy isn't trying to be the best speech to text application.
00:03:11It's trying to be the most forkable one. And you can even drop your own fine-tuned Whisper
00:03:17GGML models into the apps models folder, restart it, and then it just shows up as a custom option.
00:03:23And I think the biggest reason why I would choose Handy over something like Whisperflow any time of the
00:03:28day is because Handy just works offline and I don't need to set up an account and there's no subscription.
00:03:34And I know for sure that all my dictation audio samples stay on my machine and they're not being
00:03:40secretly used to train other voice models. So I'm super happy that we now have an MIT
00:03:46licensed dictation tool like this. So all that sounds great, but let's actually put Handy up to the test
00:03:52and see how it performs. I'll be comparing Handy's performance against Whisperflow and also against
00:03:58Google's own transcription service, which is available here in Google Docs. So here on the Google Docs page,
00:04:05I'm going to activate Google's own voice typing mode. And at the same time, I will be holding down
00:04:11Handy's transcription shortcut. And then I'm just going to go on a random tangent. And by the end of it,
00:04:18we'll see what kind of result each of the service gives us and compare the two.
00:04:26So yesterday I went to pick up some groceries and on my way back home, I suddenly remembered that I have to
00:04:34head back and hook up my SQLite database to a COBOL service that is running on my custom GPU cluster.
00:04:43I didn't actually know how to do this. So I had to ask for help. And the only tool I could think of which could help me
00:04:55was an LLM called Grok. And then I also got a bit of help from chat GPT. And by the end of it,
00:05:04we got a working application running, but it was a really big hassle. Let's separate these two.
00:05:12So as you can see here, the Google voice typing service is a lot worse. It doesn't add any punctuations.
00:05:20It doesn't know how to split your sentences. It didn't get COBOL correctly. It didn't get the SQLite
00:05:27database correctly. It didn't even get the word LLM correctly. It thought that Grok was rock. And this
00:05:35is my favorite one. It thought that chat GPT was Chad GPT. Chad GPT. Nice. So yeah. So you can clearly
00:05:45see that handy is a lot better because it does give you punctuation. It does recognize certain terms,
00:05:52and it got most of the technical terms correctly. And now let's also compare this result with
00:05:58whisper flow. So I'm going to try to repeat the same text as closely as possible to be fair to this test.
00:06:08So yesterday I went to pick up some groceries and on my way back, I suddenly remembered that I have to
00:06:16head back home and hook up my SQLite database to a COBOL service that is running on my custom GPU cluster.
00:06:26Now, I didn't actually know how to do this. So I had to ask for help. And the only tool I could think of
00:06:33which could help me was an LLM called Grok. And then I also got a bit of help from Chad GPT.
00:06:41And by the end of it, we got a working application running, but it was a really big hassle.
00:06:48Okay. And you can clearly see that whisper flow did the best job. It separated the sentences correctly,
00:06:55and it got all of the terms right. So because this is a commercial application, you would expect that
00:07:03it's going to be better than the open source tool. But honestly, there isn't that much of a difference
00:07:09between this result and whisper flows. So I still kind of prefer using handy just because it's free
00:07:17and private and open source. But whisper flow did a very good job here. So it is a solid service nonetheless.
00:07:25So there you have it, folks. That is handy in a nutshell. I wasn't joking when I said that
00:07:30this has truly become my favorite dictation tool. And I'm honestly afraid that I might forget how to
00:07:36type on my keyboard if I start overusing it too much. But those are just my thoughts and my observations.
00:07:42But what do you think about Handy? Have you tried it? Will you use it? Let us know in the comments
00:07:47section down below. And folks, if you like these types of technical breakdowns, please let me know by
00:07:52smashing that like button underneath the video. And also don't forget to subscribe to our channel.
00:07:57This has been Andres from BetterStack and I will see you in the next videos!