anshul

@ALTIC_DEV @WisprFlow Ooo ill check it out today for sure

@luhsnaa @WisprFlow Check out FluidVoice if you’re looking to explore. It’s free, local, really accurate, and extremely fast, especially with the latest update ;) We got hella viral 2 weeks ago and have shot up to 6.5K+ github stars

i've been @WisprFlow maxxing for a few months now and im never going back to using a keyboard. BUT it isn't perfect and even with the screen reading capabilities it messes up tech terms i use. pydantic becomes by dantic LoRA becomes laura RAGAS becomes raga's

voice ai went from being a signal reconstruction problem to a token prediction problem. and now the same machinery that scaled text models can start eating audio too.

frame rate is the product knob > codec frame rate = audio tokens per second > higher frame rate = more tokens, more compute, more accumulated errors > lower frame rate = shorter sequence, easier generation > but you still need enough detail to reconstruct speech

another banger from @rumik_ai , voice ai has gotten exponentially better and not many know what really changed. i'll try to dumb this down even further for myself: > voice ai used to turn speech into mel spectrograms. > newer systems turn speech into discrete audio tokens. that change is why speech ai is starting to look a lot more like llms.

so i'm making a small open-source benchmark for programmer speech. the dataset will have real clips from my friends and i. i also wanna see if fine tuning means the model memorizes the terms or learns patterns. this expirement could totally not lead anywhere but im curious

this is why normal stt scores can be misleading. a transcript can be 95% right and still be useless for a programmer. if the missed 5% is the package name, eval metric, model name, cli flag, branch name, whatever, the average score hides the failure.

@realAkashAnand Research

At YC we have an internal platform called Bookface, which has better startup advice than anything you'll find on Google. But (a) only YC founders have access, and (b) it doesn't cover every topic. Plenty of times I've searched and it come up empty. Turns out Claude Fable 5 is really good at multi-step research. So I built a skill that helps anyone learn about a topic or figure out how to implement something at work. It finds essays and articles with immediate action points, written only by top-tier operators, founders, and experts who've actually done the thing and can tell you exactly what they did in the situation you're in. Fluffy SEO blogs are strictly rejected. And it does a MUCH better job than Google AI mode, Perplexity, or even Claude deep research. I’ve been using this every day, with great results. Here are some things I’ve researched this week: - How to define top-level KPIs and metrics for my startup - How to run a company all-hands with my employees - How to prompt Claude to pick the right fonts for video - How to measure the success of features (small, medium, and large) It’s useful for everything from sales, to marketing, to engineering, to product, to design, to general advice. If you want this skill, comment "RESEARCH" and I'll send it over. (You'll have to follow me so I can DM you.)

@kdrvrtk he went through some injury right? haven’t heard much from him since. Trying to stay relevant maybe?

Tell me a bigger disappointment than Saket Gokhale. I’ll wait. Bro went from inspiring people to get into fitness to endlessly yapping hate about science-based lifting, cuffed reps, and every other niche debate. Meanwhile, he barely touches free weights anymore. Miss the old Saket. All my homies hate Saket Gokhale.

@AlfieJCarter Agents

R.I.P. rebuilding your GTM stack from scratch every session. A complete Claude Skill Library can replace a $15,000/month agency retainer. It is not as easy as hiring someone else to do it. But if you start today, you can have 56 skills loaded into Claude covering SEO, content, outbound, sales, growth, analytics, strategy, ads, social, and CRM by end of this week. I usually charge $299 for access to this library but today, it's free. Like this post + comment 'Agents' and I'll DM you the entire skill library for free. (Must be following, or I can't message.) Taking this down in 48 hours.

@kdrvrtk @ShashankH_ @soni0696 congrats mann

My company decided to put up with me so i guess i am SDE-1 now🎉 Thanking @ShashankH_ @soni0696 for approving my AI generated slop code so that i could reach this stage

stereo vs mono is another common one. model expects mono. file is stereo. might error. might take one channel randomly. might downmix wrong. might feed 2× the expected shape into a pipeline that doesn’t care gracefully. check channels. downmix explicitly. log what you did.

if 8 kHz audio gets treated as 16 kHz without an explicit resample step, the model hears it at the wrong speed. features don’t match training. WER explodes. nobody touched the weights. looks like "ASR is bad on phone calls." might just be "we never validated the input."

debugging voice stacks taught me: read sample rate, channels, duration, and sample count before you run ASR. wrong format often fails without an explicit error.

the further up the stack you go from raw audio, vad → endpointing, the less the problem is technical and the more it's about who you think your user is and how they speak. tuning these parameters can be quite painful as you are forced to CHOOSE your primary user.

a voice agent optimized for fast back-and-forth will feel like it's cutting you off if you're dictating. a dictation tool with 3s patience will feel sluggish in a live conversation.

Voice Activity Detection (VAD) and endpointing are not the same thing. VAD: “Is there speech in this frame right now?” Endpointing: “Is the user done with this thought?” a good voice stack keeps these separate.