lunixbochs

10 ms·

61.
by lunixbochs 5y ago
I would say not quite as convenient, because they lean on that UI to also make you constantly confirm top-1 commands that would've worked fine. As you can see in my Conformer demo video I can hit top-1 so reliably I don't even nee
62.
by lunixbochs 5y ago
Yeah for sure, breaks are really important. sconv was especially disappointing because it looked so good on metrics during my training, but the cracks really started to show once it entered user testing. Conformer has been so much less st
63.
by lunixbochs 5y ago
Are you referring to the test you mentioned in this thread? https://news.ycombinator.com/item?id=26784732 If so, December predates Conformer, so you're talking about the sconv model, which is the model I was complainin
64.
by lunixbochs 5y ago
> vs saying a whole ‘sentence’ in one go (fast and natural but you’re probably going to have to go back and edit/try again) "probably going to have to go back and edit" is generally not the case with my Conformer model, wh
65.
by lunixbochs 5y ago
There are some a similar techniques where you use evaluation metrics to decide which data to train on each epoch. I wonder if you could make that parameter trainable instead of using a hyperparameter search for it. For phonetic similarity I
66.
by lunixbochs 5y ago
At face value, actively altering grammar to words that you didn't say sounds like the language model is very heavily weighed. I'm curious if you mean the keyboard? Because that recently switched to on device I think, which means m
67.
by lunixbochs 5y ago
You could weigh insertions by how much perplexity they add (sum), deletions by how much perplexity they remove (-sum), and replacements by how big the ppl difference is in the replaced word (abs(sum)). And report this as a 4-part score (com
68.
by lunixbochs 5y ago
Interesting, maybe instead of my proposed perplexity metric, we measure the difference in both utterance and per-word perplexity between ground truth and output with a strong language model? Ideally it's low - the language model should
69.
by lunixbochs 5y ago
Maybe some kind of alignment metric, to measure how far off on timing notes tend to be? I can also imagine a generalized "local error rate" which measures how far away errors tend to be from each other. If errors tend to be cluste
70.
by lunixbochs 5y ago
I ship speech recognition to users for full computer control (mixed commands and dictation) with a very tight feedback loop. I get a lot of direct feedback about any common issues. One time I beta tested a new speech model I trained that sc
71.
by lunixbochs 5y ago
There's a user space network stack in gvisor, written in Go. And it's already included in wireguard-go. See also https://fly.io/blog/ssh-and-user-mode-ip-wireguard/ which has already done this for SSH th
72.
by lunixbochs 5y ago
I'm an ASR researcher shipping high quality English models trained on limited resources, and while I've needed to include other datasets to make the model more robust to different kinds of text, Common Voice is a substantial part
73.
by lunixbochs 5y ago
Copilot isn't a good example because it was trained differently. A better comparison would be GPT3 vs GPT-J. Note that there's still a pretty big difference in parameter size.
74.
by lunixbochs 5y ago
My reading of the generator diagram (figure 6) isn't that it is generating waveforms, but that it is generating phoneme probabilities. You can train a similar system to produce audio on the output of wav2vec, though it probably won&#x2
75.
by lunixbochs 5y ago
One addendum to the linked post's notes: > SoTa in low-resource setting Libri-light by a lot on WER clean test 100h labeled: others ~4 vs theirs ~2.5 > SoTa on high-resource noisy data (3.3 vs 3.4) close to SoTa on clean data Thi
76.
by lunixbochs 5y ago
> I can find literally nothing online, or even internally, about using Visual Studio with Dragon NaturallySpeaking My Talon project is capable of voice controlling Visual Studio, both at the interface level and at the code insertion&#x2F
77.
by lunixbochs 5y ago
> The participant was able to communicate at about 90 characters, or 18 words, per minute. > By comparison, able-bodied people close in age to the study participant can type on a smartphone at about 23 words per minute, the authors sa
78.
by lunixbochs 5y ago
If you can reproduce the original information based only on a state input, you have stored it in the state (in an encoded form or not). If your state is smaller than the original information, you have compressed it. If your reproduction is
79.
by lunixbochs 5y ago
I don't think that assessment is quite right. The hidden size is fixed - the second argument to Pytorch's nn.LSTM constructor is "hidden_size – The number of features in the hidden state h". A call to `y, hidden = layer.
80.
by lunixbochs 5y ago
Only if the entire loop audio fits into the RNN's hidden state? Or if you use some kind of external memory mechanism that gives the network access to previous audio.
81.
by lunixbochs 5y ago
I did tell you to try "command history" on the same line I told you about events.tail(). You can pass a string argument to events.tail() to filter it (which works even better in the beta/next version). I'm considering mo
82.
by lunixbochs 5y ago
First - online learning is generally not necessary with my Conformer model (which is generally >5x better at not making errors than the current public model). It's also quite good at accents. My overall goal is to ship speech models
83.
by lunixbochs 5y ago
I edited the parent since you started writing this post, maybe re-read the parent as I preempted some of the questions upthread. For browser stuff, I'm actually really disappointed with how the "actions implemented in talon files&
84.
by lunixbochs 5y ago
Talon doesn't require a $300 mic or a completely silent room. It works fine on my macbook air's builtin mic, and explicitly handles some interesting background noise scenarios (e.g. loud music, sitting next to a running dryer).
85.
by lunixbochs 5y ago
I recommend asking about hiccups on the Slack [1]. My basic analysis is you're missing a vscode-side plugin, you might need to restart Talon once after dropping in knausj_talon, and that your statement of "it doesn't really w
86.
by lunixbochs 5y ago
You don't really need to pause in Talon if you set it up right and practice: https://twitter.com/lunixbochs/status/1378159234861264896
87.
by lunixbochs 5y ago
Weird, I don't see those interstitials. Do you have a screenshot of what that looks like?
88.
by lunixbochs 5y ago
This was also a feature of the DEFCON 26 finals in doublethink
89.
by lunixbochs 5y ago
My final artifact is a dmg. I emit it as part of the app build, in my very first CI stage on Mac, and immediately upload for notarization before the build job ends. Then, in parallel to tests, I have a separate job that checks the dmg for c
90.
by lunixbochs 5y ago
I'm assuming it doesn't work well with nested bundle signing. As per my other thread it also seems to be picky about which subdirectories it signs, and there are lots of weird paths (LaunchDaemons, XPCServices, LoginItems, etc) yo

More ›