I Cancelled Wispr Flow & Built This Instead (Free Tool) — Transcript

Full transcript

1. 0:00Whisper flow.
2. 0:00>> Whisper flow. Whisper flow. Whisper
3. 0:01flow.
4. 0:02>> Whisper flow.
5. 0:02>> Whisper flow.
6. 0:03>> Whisper flow.
7. 0:03>> Whisper flow.
8. 0:04>> Whisper flow. Whisper flow is, I'm just
9. 0:06going to say it, completely unnecessary.
10. 0:08They just raised $280 million at a $2
11. 0:10billion valuation for exactly what I'm
12. 0:12doing right now using a tool I built in
13. 0:1420 minutes with Claude code. Look at
14. 0:16that perfect dictation. So, in this
15. 0:18video, I'm going to show you exactly how
16. 0:19to build your own version of Whisper
17. 0:20Flow, how to run it completely on your
18. 0:22own machine, so nothing you say ever
19. 0:23leaves your computer, how to make it
20. 0:25look like a real app instead of a script
21. 0:26in a terminal, and how to make it even
22. 0:28faster than the one you're already
23. 0:30paying $15 a month for. This tutorial is
24. 0:32designed for a complete beginner. All
25. 0:34we're going to be doing is talking to
26. 0:35Claude Code, the natural language. By
27. 0:37the end, you'll have your own version of
28. 0:39Whisper Flow. So, not wasting any more
29. 0:41time, let's dive right into the build.
30. 0:42And we're going to start with this
31. 0:43prompt into Claude Code. And for this
32. 0:45prompt, we're actually going to keep it
33. 0:46really simple. All I'm going to say is I
34. 0:49want to clone whisper flow. Look up what
35. 0:51this is. I basically want it to be a
36. 0:55pushto talk dictation app. What would
37. 0:57you recommend for architecture based on
38. 1:00my machine? We'll start with the
39. 1:03skeleton, then build this out into a
40. 1:06proper MacOss application and add some
41. 1:09cool branding. So, let's just take that,
42. 1:11copy it, and I'm going to open up cursor
43. 1:13here. I'm going to open up Claude in
44. 1:15here and I'm just going to paste this
45. 1:16prompt so you can see just how easy this
46. 1:18is. All I'm doing is I'm going I I don't
47. 1:19know anything about the architecture or
48. 1:20anything like that. All I'm doing is I'm
49. 1:22saying this is what I want to build.
50. 1:23Look this up, come up with the
51. 1:25architecture, and we'll just get to a
52. 1:26working prototype really quickly and
53. 1:28then we can layer on things like
54. 1:29building out an actual Mac OS
55. 1:31application, updating some of the
56. 1:32branding, you know, the waveform that
57. 1:34shows up, all that kind of stuff. So
58. 1:35Claude has a plan for us now, and it's
59. 1:38laid out all these things. shell, HUD,
60. 1:40hotkey, audio, ST, cleanup, injection.
61. 1:43Now, you really don't need to know what
62. 1:44any of this stuff means, but for the
63. 1:46sake of this being a proper tutorial, I
64. 1:48just threw an explanation of these in
65. 1:50this deck here. So, this is what the
66. 1:51architecture is. So, first the shell,
67. 1:53this is just the app itself. This is the
68. 1:55going to be the container that
69. 1:56everything lives in. So, we're building
70. 1:57a proper Swift application, how we're
71. 2:00actually going to build the interface.
72. 2:01But this is just the heads up display.
73. 2:03So, when I actually just like you saw
74. 2:05when I was transcribing initially, that
75. 2:06little waveform that's going to pop up.
76. 2:08So we can see that it's actually being
77. 2:09transcribed in real time. Hotkey CG
78. 2:11event tap. So this is just what button
79. 2:13we're going to hold to actually activate
80. 2:15the transcription. In this case, we're
81. 2:16just going to use Fn. Audio is just
82. 2:18actually the microphone that's going to
83. 2:19be picked up when we are talking. Then
84. 2:21of course, we have ST, speech to text.
85. 2:23This is the model that turns your voice
86. 2:24into words. In this case, we're just
87. 2:26using Axe native speech transcriber.
88. 2:29This is relatively new. I wanted to
89. 2:30experiment with this. It's built
90. 2:32directly into Mac OS, so it's 100%
91. 2:34local. You're not calling an API. You're
92. 2:36not paying for anything. and you're not
93. 2:37even downloading a local model just
94. 2:39directly out of the box through Mac OS.
95. 2:41And then just as an alternative, just in
96. 2:44case speech transcriber is not good
97. 2:45enough, we have Parakeet, which is
98. 2:47Nvidia's local model that you can
99. 2:48download really easily. I'll show you
100. 2:50how to do that through hugging face.
101. 2:51We're going to just have that as a
102. 2:52backup in case speech transcriber is
103. 2:54just not that good. And then cleanup as
104. 2:56well. This one's completely optional,
105. 2:58but we're going to have a small local
106. 2:59model that's just going to add
107. 3:00punctuation, cut the ums, do that kind
108. 3:02of stuff that Whisper Flow, I know, does
109. 3:04really well. Like I said, we absolutely
110. 3:06do not need to do this, but just to see
111. 3:08how good it can actually get to Whisper
112. 3:09Flow, we're adding the second small
113. 3:11model. And then finally, we have
114. 3:12injection. So, this is just making sure
115. 3:13that the text is injected properly where
116. 3:15our cursor is. So, with that, let's
117. 3:17check in on the latest in the
118. 3:19architecture build. So, it looks like it
119. 3:21is done, I think. Yeah. So, you can see
120. 3:23here it's just explaining Mac OS 26
121. 3:26ships speech analyzer/spech
122. 3:28transcriber. Apparently, Parakeet V3
123. 3:31beats it on English accuracy. So, it did
124. 3:33some research on that. We're going to
125. 3:34start with just speech transcriber just
126. 3:36to see how good it is. And then we can
127. 3:38test parakeet later on. And then it has
128. 3:40basically all the things that I already
129. 3:41explained. The HUD hotkey, CG event tab,
130. 3:44FN button, and then the only thing it
131. 3:45looks like I need to do is just grant
132. 3:48permissions. So system settings, privacy
133. 3:49and security, accessibility, and then
134. 3:51add what we're calling murmur, this
135. 3:53application that the agent just built.
136. 3:55So we'll do that now. I'll just go to
137. 3:56privacy and security, accessibility, and
138. 3:58then minor confusing thing, I already
139. 3:59have a whisper flow clone that I've been
140. 4:01using for a couple months now, also
141. 4:03called Murmur. We're just going to
142. 4:04change the name of this. We'll just call
143. 4:05it Murmur YouTube. Change name. Also, I
144. 4:08just gave permissions. Do you have full
145. 4:12access? I just need to open this system
146. 4:14settings. Open system settings. Okay.
147. 4:17Remember YouTube and change the name.
148. 4:19All right. Okay. Should have full access
149. 4:21now. Please test. Hey, would like to
150. 4:23access the microphone. Allow. Try now.
151. 4:25All right. So, after allowing those
152. 4:27permissions, it looks like this is
153. 4:28working. Collad drove this with a
154. 4:30synthetic right control hold and
155. 4:32confirmed everything is working in the
156. 4:33logs started listening finishing idle
157. 4:35speech analyzer starts HUD renders etc
158. 4:38etc so what do I need to do right
159. 4:41control testing oh look at this testing
160. 4:43hey uh can you hear me oh look at this
161. 4:45oh that's a nice little touch uh just
162. 4:47continued to test here I just want to
163. 4:50see how much you can pick up here and
164. 4:53how fast you are uh if I just keep
165. 4:55ranting like this um I like the waveform
166. 4:57too it's it's a fancy it's A fancy
167. 4:59little waveform. Okay. Where'd it go?
168. 5:00Where'd it go? Did it capture that
169. 5:02testing again? Can you hear me? All
170. 5:03right. It's close, Claude. It picks up
171. 5:06on my voice, I think. Do you see this in
172. 5:10the logs? It's just not inserting it
173. 5:13where my cursor is. Please look at logs
174. 5:16and fix. Okay, the logs show it's
175. 5:18working further than you'd think. Okay,
176. 5:20transcription and injection both
177. 5:22succeeded. Then why is it not working?
178. 5:23Let's see. What does it say? That's the
179. 5:24classic AX. Uh yeah, classic AX silent
180. 5:27failure. Return success and Electron
181. 5:29apps. Oh, interesting. Okay, so it's a
182. 5:31cursor thing when I'm trying to type in
183. 5:32cursor. All right, so it's fixing that
184. 5:34now. All right, rebuilt and armed. This
185. 5:36should work. Testing. Hello. Can hear
186. 5:38me? Hey, look at that. That was fast,
187. 5:41too. Dang, that was really fast. Let's
188. 5:43do some long rant here. Uh this is just
189. 5:45a long rant here. Uh just to test how
190. 5:47well this is working. Let's just keep
191. 5:49talking. I do like this little thing on
192. 5:51the on the right here where it just auto
193. 5:53transcribes even though it's kind of
194. 5:54irrelevant because you can't even really
195. 5:56see it. But anyway, this is uh this is
196. 5:58pretty good. Claude did a pretty good
197. 6:00job for literally a one shot more or
198. 6:02less. There was a couple there's a
199. 6:04couple, you know, permissions that need
200. 6:05to be granted and this ax this classic
201. 6:09AX error. Okay, let's see how fast.
202. 6:11Whoa, that is fast. Holy I bet that's
203. 6:14faster than parakeet. That is crazy. So,
204. 6:17I I haven't I haven't been using this
205. 6:18new Mac OS model, so I might need to
206. 6:20switch my own Whisper Flow clone to
207. 6:22this. Uh, testing this again. Let's just
208. 6:24see how fast this goes again. I'm going
209. 6:25to talk really fast and see what
210. 6:26happens.
211. 6:28Dude, that's crazy. [laughter]
212. 6:30So, holy crap, that actually worked way
213. 6:32better than I even thought. You can see
214. 6:33how easy that was, right? That was like
215. 6:34one and a half turns. We told Claude we
216. 6:36wanted to create a whisper flow clone.
217. 6:37We granted some permissions. We had one
218. 6:39minor issue with just inserting into
219. 6:41Electron apps like cursor. And then now
220. 6:42this thing is fully working. We have a
221. 6:44Whisper Flow clone that we can use,
222. 6:45which by the way, I'll include as a repo
223. 6:47in the link below if you want to just
224. 6:48take this and run with it. But I mean,
225. 6:49pretty freaking cool. So, what we're
226. 6:51going to do next is we're just going to
227. 6:52show you how to download a local model
228. 6:55just in case you want to do this. But it
229. 6:56looks like this Mac, this new speech
230. 6:58transcriber is very fast and works
231. 7:00really well. And I'm just going to tell
232. 7:02Claude to do that with our new Murmur
233. 7:04YouTube. So, good job, Claude. This
234. 7:06looks great. It's very fast. We're using
235. 7:08the Mac Mac OS 26 speech transcriber.
236. 7:10It's It's really good. I want to do a
237. 7:12sideby-side test with Parakeet somehow.
238. 7:15Um, I don't know how you can do this,
239. 7:17but just figure out how to run a test.
240. 7:19Maybe just come up with a quick little
241. 7:20dashboard and I'll just record
242. 7:21something. We'll see how quick and
243. 7:22accurate each of these are. I want you
244. 7:24to explain and just show how we grab
245. 7:26this parakeet model via hugging face
246. 7:28too. So, okay, it's doing two things.
247. 7:30The first is it's accessing this model
248. 7:32via hugging face. I have the hugging
249. 7:35face MCP server. Quick random aside,
250. 7:37that makes it easier to access and find
251. 7:39these models, but you might not have
252. 7:41that. I I don't want to get into setting
253. 7:42up the MCP server. So I just said don't
254. 7:44use the MCP server. So it's just using
255. 7:46plain curl to access this model via
256. 7:48Hugging Face's public API. So anyone can
257. 7:50do this. I do highly recommend that you
258. 7:51download the Hugging Face MCP server.
259. 7:53But that is outside the scope of this
260. 7:54demo. It's just going to access that via
261. 7:56Hugging Face and then just download this
262. 7:58parakeet model and then we're going to
263. 7:59do a sideby-side test. Okay, cool. So we
264. 8:01have something built here. Cloud just
265. 8:02needs to run this. What I'd expect so
266. 8:04you can judge the results. Parakeet
267. 8:06should win on raw RTF by wide margin.
268. 8:08Oh, interesting. Okay, we'll see about
269. 8:10that. Uh, cool. Just build this. I don't
270. 8:13know what else to say. I don't know what
271. 8:14it wants from me. All right. So, Claude
272. 8:15just got done putting together this
273. 8:16engine comparison. We're going to
274. 8:17compare Mac speech transcriber versus
275. 8:19parakeet, the local model that we just
276. 8:21downloaded from Hugging Face. We have
277. 8:22this handy little window here, engine
278. 8:25comparison. And all I need to do is just
279. 8:26hold control and test to see how fast
280. 8:29and accurate each of these are. So, I'm
281. 8:31just going to say, okay, just testing
282. 8:33this record now with our Murmur YouTube
283. 8:36whisper flow clone. And I'm going to
284. 8:38talk for a minute here and kind of go on
285. 8:39a rant. I may have to cut this down
286. 8:42because this is kind of ridiculous to
287. 8:44spend like 30 seconds to probably a
288. 8:46minute talking. Uh complete nonsense.
289. 8:49Okay. Wow. Jeez, those are freaking both
290. 8:52of them are really fast. Okay. So, looks
291. 8:53like parakeets's a little bit faster.
292. 8:550.32 seconds. 0.58. Let's see what
293. 8:58actually came out in the transcript.
294. 9:00Okay. Just testing this with remember
295. 9:02YouTube whisperflow clone. Yep, that all
296. 9:04looks all right. Okay. Interesting. So
297. 9:06AL called it cologne which is wrong. We
298. 9:09won't be able to tell quickly. You will
299. 9:11actually pick up some testing that I
300. 9:12honestly can't even believe. Let me test
301. 9:14something else too. Also want to see how
302. 9:15well you pick up on words like claw code
303. 9:18for whatever reason. Claude code is very
304. 9:20hard to get right. All these speech text
305. 9:23models don't seem to get clawed code
306. 9:25right. Let's see how well it does.
307. 9:26Parakeet still faster. They get they
308. 9:28can't get clawed code. Claw code. Claw
309. 9:31code. All right. This is what we're
310. 9:32going to do. We're just going to go with
311. 9:33Apple just because it's native to Mac
312. 9:34OS. It's a little easier to set up. You
313. 9:36don't need to go through the trouble of
314. 9:37downloading a local model via hugging
315. 9:39face. It'll just be available pretty
316. 9:41much out of the box in the GitHub repo
317. 9:42below. But what I will do because I know
318. 9:44of course that means if you're a Windows
319. 9:45user, this won't work. So I'll include a
320. 9:47separate GitHub repo below that will use
321. 9:49parakeet and make our own version of
322. 9:51Whisper Flow available to you. I'll
323. 9:52include instructions too on how to
324. 9:53actually download that local model and
325. 9:55you'll be able to just copy the repo,
326. 9:56give it to your agent, say figure this
327. 9:58out. All right, so that being said, our
328. 10:00WhisperFlow clone is fully working. But
329. 10:02the question is, how does this compare
330. 10:03to Whisper Flow? Is Whisper Flow faster
331. 10:06or better in any way? Let's just quickly
332. 10:08find that out. I'm going to pull up
333. 10:09Whisper Flow and compare it with our
334. 10:11clone somehow and see how accurate and
335. 10:13fast it is relative to our clone. All
336. 10:15right, so Claude added something to our
337. 10:18comparison engine where it's pulling
338. 10:19from Whisper Flow's database. So I have
339. 10:21Whisper Flow open here and in our
340. 10:23engine, if I open this comparison view,
341. 10:25I can record all three now. So it's
342. 10:27going to record Parakeet, it's going to
343. 10:29record Max Speech Transcribe, and then
344. 10:31it's going to record Whisper Flow. and
345. 10:32we're going to see how quickly these
346. 10:34transcribe and how accurate they are.
347. 10:36So, let's go record all three. Okay,
348. 10:37testing this now with Whisper Flow, with
349. 10:40Mac Transcribe, with Parakeet. I'm just
350. 10:42going to go on another rant. This is
351. 10:44going to be a quicker rant than the last
352. 10:46rant, but I'm just going to keep talking
353. 10:47so we can really test the speed of all
354. 10:49three of these models. See if Whisper
355. 10:51Flow is really that much better at this
356. 10:53point. It's going to be so marginal that
357. 10:55it doesn't really matter in any way.
358. 10:57Like, what is Whisper Flow's
359. 10:59differentiator here? Let's see.
360. 11:03All right. So, it looks like Parakeet
361. 11:05was the fastest, 0.27 seconds. Apple was
362. 11:07second, and then Whisper Flow was third.
363. 11:10So, okay, let's look at the transcript.
364. 11:12Just going to go into the rant. This is
365. 11:13going to be quicker. See if Whisper Flow
366. 11:15Whisper Flow got its own spelling
367. 11:17correct. Apple's missing some things
368. 11:18like saying rent instead of rant. Apple
369. 11:21got that question mark at the end.
370. 11:22That's pretty good. All right, so there
371. 11:23we go. Both of these models are
372. 11:24officially faster than Whisper Flow. And
373. 11:26our little clone is completely free. And
374. 11:28you can see at least in this example
375. 11:30that it is just as accurate in the
376. 11:31actual text transcription. Now the one
377. 11:33caveat and Claude calls this out here is
378. 11:35there is a little bit of latency because
379. 11:37we're actually pinging Whisperflow's
380. 11:39database to get how long the
381. 11:40transcription takes. So there's likely
382. 11:43some latency. And so that's why that
383. 11:44number is showing up much slower. But
384. 11:47even if it was, so if we look at the
385. 11:49numbers, even if it was a second slower,
386. 11:51it would still take 0.91 seconds versus
387. 11:530.48 versus 0.27. And there probably is
388. 11:56a way to be more methodical and get
389. 11:57accurate numbers between Whisperflow's
390. 11:59model parakeet and Mac speech
391. 12:01transcribe, but the point being the
392. 12:02difference in speed and accuracy is
393. 12:04marginal at best. Now, we could stop
394. 12:06right here and just have our little
395. 12:07waveform HUD. But let's get this close
396. 12:10to Whisper Flow and build a proper
397. 12:12interface, a guey as the pros call it.
398. 12:15That way, we can start and close the app
399. 12:17and just have it function very similar
400. 12:18to Whisper Flow. So, what I'm going to
401. 12:20ask Claude to do is have this dictation
402. 12:23history. We're not going to do notetaker
403. 12:24or anything like that. And then we're
404. 12:26going to add a dictionary, too, and an
405. 12:27ability to add words. So, this is the
406. 12:29prompt I'm going to give to Claude.
407. 12:31Basically, what I'm saying, we'll just
408. 12:32open up in this doc here. I'll include
409. 12:34this in the link below as well if you
410. 12:36need it, but I'm just saying turn this
411. 12:37into a proper Mac OS application, not a
412. 12:39menu bar. The main window holds the past
413. 12:41transcriptions, start a dictionary,
414. 12:43exactly what Whisper Flow contains. By
415. 12:45the way, too, to get this prompt, I just
416. 12:47went to a separate Claude agent that was
417. 12:48familiar with this project, and I said,
418. 12:49"What would be a good prompt to give to
419. 12:51this agent?" And then I also wanted it
420. 12:52to lean into some kind of design
421. 12:54direction. So I said the direction is a
422. 12:551980s tape recorder. Plot just listed
423. 12:58out some brands. And then I just said
424. 13:00don't use neon vapor wave synth wave
425. 13:02purple and pink gradients. I feel like
426. 13:03these are just too overused. So I really
427. 13:05want to just like lean into some kind of
428. 13:06design style and have be kind of quirky
429. 13:08and fun and a little bit different than
430. 13:10these other voice transcribers. So we'll
431. 13:11get Claude working on that. I'm guessing
432. 13:13it will be built relatively quickly. All
433. 13:16right. So Claude has built our Murmur
434. 13:18YouTube interface. We've got the
435. 13:19application. We've got transcription
436. 13:21history. We've got dictionary. So, let's
437. 13:23open it up and check it out. So, here it
438. 13:26is. It doesn't look bad. It's a little
439. 13:28plain, but I guess it technically got
440. 13:30the recorder kind of 80s feel. Could
441. 13:32have done a better job, but let's see if
442. 13:33it actually works. So, I'm just going to
443. 13:35hit record. Okay. Testing this record
444. 13:38now to see if our murmur YouTube guey is
445. 13:41actually working. How is it going to
446. 13:43transcribe guey? Okay. Didn't get gooey,
447. 13:45right? Goey, but okay. Testing this
448. 13:47record now. Murmured YouTube. Okay.
449. 13:49Well, there we go. We have transcription
450. 13:51history. So, let's go to dictionary. I
451. 13:53just had claude add some random phrases.
452. 13:55So, claude code. Let's see if it'll pick
453. 13:57up on clad code. Okay. Testing this now
454. 13:59by saying the word claude code. Let's
455. 14:01see if it will actually correctly
456. 14:04transcribe claude code. Boom. There we
457. 14:07go. Cool. Oh, and it even shows like a
458. 14:09correction. Corrected. Wow. That really
459. 14:11does that very fast. Like corrects it in
460. 14:13real time like that. Okay. Testing this
461. 14:15cloud code. Cloud code. Boom. Just like
462. 14:17that. All right. So, we have dictionary.
463. 14:19We can add different words here for the
464. 14:20model to pick up on that transcription
465. 14:22history. I mean, that's basically all
466. 14:24you need. Don't need anything fancy. And
467. 14:26then you can just close this and have it
468. 14:27running in your menu bar and then just
469. 14:29be recording stuff. So, I could say like
470. 14:31I could open that Google doc again. Just
471. 14:33say testing this again or murmur YouTube
472. 14:37transcribing. See if this actually works
473. 14:39in this Google doc. Boom. There we go.
474. 14:42Beautiful. So, that's going to do it.
475. 14:43That's how to clone Whisper Flow or
476. 14:45really any of these voice dictation apps
477. 14:47in just a couple prompts with Claude
478. 14:49Code. And don't get me wrong, too. I
479. 14:50have a lot of respect for Whisper Flow.
480. 14:51They've built a great brand. I mean,
481. 14:53their marketing is fantastic. And if you
482. 14:54want something that just works
483. 14:55immediately, it is a great option. But
484. 14:57if you're trying to save money, if
485. 14:58you're trying to own your own data, if
486. 15:00you're trying to customize this to your
487. 15:01liking, you can see just how easy it is
488. 15:03to do. Not to mention, if you don't want
489. 15:04to build any of this yourself, I'll
490. 15:05leave the repo below. Literally, just
491. 15:07copy it from GitHub, give it to Claude
492. 15:09or Codex or whatever you use, and just
493. 15:11say set this up on my own machine. So,
494. 15:12appreciate you watching. There are so
495. 15:14many apps like this, by the way, that
496. 15:15could just be cloned in really just a
497. 15:17few prompts with cloud code. So, if you
498. 15:18have any ideas, if you have any apps
499. 15:19that you use daily that you're paying a
500. 15:21lot of money for, let me know and I can
501. 15:23just put these together, make the GitHub
502. 15:24repos available, and give you the
503. 15:25ability to have fully custom software.
504. 15:27All right, I'll see you in the next one.