YouTube2Text

Qwen3-TTS Tutorial: Open-Source Voice Design & Cloning — Transcript

by Thorsten-Voice · 2,628 words · 448 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Good, today is February 1st, 2026 and
  2. 0:04I'm taking a hands-on look at Qwen 3
  3. 0:07TTS, the newly open-source
  4. 0:09text-to-speech family from the Qwen
  5. 0:12team. Qwen 3 TTS brings together voice
  6. 0:16design, a 3-second voice cloning, and
  7. 0:19instruction-based voice control with
  8. 0:21support for
  9. 0:23>> [clears throat]
  10. 0:23>> Chinese, English, Japanese, Korean,
  11. 0:26German, French, Russian, Portuguese,
  12. 0:29Spanish, and Italian.
  13. 0:31In this tutorial, I'll take a look at
  14. 0:33the GitHub repository and play around
  15. 0:36with their Hugging Face demo space.
  16. 0:39To make everything easy to follow, I've
  17. 0:41created a Google Colab notebook that you
  18. 0:44can use as a starting point and adapt
  19. 0:47for your own experiments.
  20. 0:49In it, I will show you how to start
  21. 0:52WebUI, generate voices from descriptions
  22. 0:55in English and German, and clone voices
  23. 0:58from short reference samples.
  24. 1:01If you enjoy practical hands-on AI voice
  25. 1:04tutorials like this one, consider liking
  26. 1:07the video and subscribing to the
  27. 1:09channel. And do not forget to share my
  28. 1:12channel with other open-source voice
  29. 1:14technology enthusiasts.
  30. 1:17Wow, that's all for today.
  31. 1:20This is all for the introduction and now
  32. 1:22let's let's go. Let's start taking a
  33. 1:25look to their GitHub page and as always,
  34. 1:28all the links in the description box
  35. 1:29below. So, first of all, the license, it
  36. 1:32is Apache 2 licensed and they gained
  37. 1:366,000 stars already since their release
  38. 1:401 week ago. So, let's take a look to
  39. 1:43their commit history
  40. 1:45and you can see their initial commit has
  41. 1:48just been
  42. 1:51last week. So, just 1 week, 6,000 stars
  43. 1:56on GitHub.
  44. 1:58I mean, I'm not really jealous because
  45. 1:59it took me over 3 years to gain 10,000
  46. 2:03subscribers on my YouTube channel, but
  47. 2:06so they got 6,000 stars
  48. 2:09in 1 week.
  49. 2:10The code is purely Python-based, as you
  50. 2:13can see here.
  51. 2:14On top of the read me, you can see some
  52. 2:17links to hugging face model information
  53. 2:19blocks, a demo space. We'll take a
  54. 2:22closer look in a moment.
  55. 2:24And we have this one news here, so from
  56. 2:26January 22nd, where they released their
  57. 2:30initial release, and this contains two
  58. 2:33models. We have a 0.6
  59. 2:36billion model parameters model and a 1.7
  60. 2:40billion parameters model.
  61. 2:42Here you can see Qwen text-to-speech
  62. 2:44covers 10 major languages, so
  63. 2:47Chinese, English, Japanese, Korean,
  64. 2:50German, French, Russian, Portuguese,
  65. 2:51Spanish, and Italian.
  66. 2:53And here you can see intelligent text
  67. 2:55understanding and voice control. So,
  68. 2:57this is this type of voice design where
  69. 3:00you can describe
  70. 3:03attributes for your voice, timbre,
  71. 3:05emotion,
  72. 3:06prosody, so these types of information
  73. 3:10to design and create
  74. 3:12a text or textually described AI voice.
  75. 3:16In this table, you can see a comparison
  76. 3:18of the features depending on their
  77. 3:21models. Again, we have this 1.7
  78. 3:24billion parameters models, and we have
  79. 3:27these 0.6 billion parameters models
  80. 3:31here. In the quick start section, they
  81. 3:34create a Python 3.12 virtual environment
  82. 3:37using conda,
  83. 3:39and you can install it pretty simple
  84. 3:41using Qwen TTS Python package. We will
  85. 3:44come to that in a moment when we take a
  86. 3:46look to my Google Colab notebook. In
  87. 3:50general, they provide lots of Python
  88. 3:53code snippets and examples for all their
  89. 3:56use cases, voice design and voice
  90. 3:58cloning. So,
  91. 4:01pretty good documentation here on their
  92. 4:03readme.
  93. 4:04I have to say, I guess they deserve the
  94. 4:066,000 stars in 1 week.
  95. 4:09Before we hop onto their hugging face
  96. 4:12demo space to play a little bit around
  97. 4:14with,
  98. 4:15um let's take a short look to their blog
  99. 4:17post here. Let's scroll down to the
  100. 4:20audio samples. You can see here some
  101. 4:24control instructions mentioned here. So,
  102. 4:27here's control instruction, gender male,
  103. 4:30pitch low pitch, so describing your
  104. 4:33design voice.
  105. 4:34And we have this text here, so let's
  106. 4:37just listen to this sample.
  107. 4:39>> Nine different exciting ways of cooking
  108. 4:42sausage. Incredible. There were three
  109. 4:45outstanding deliveries in terms of the
  110. 4:47sausage being the hero. The first dish
  111. 4:50that we want to dissect, this
  112. 4:52individual's
  113. 4:54>> Okay, let's listen to a Chinese sample,
  114. 4:57which I cannot understand at all.
  115. 5:12Okay, I do not understand
  116. 5:15any word, but I understand or I can I
  117. 5:17can feel the emotion, which is pretty
  118. 5:20impressive.
  119. 5:21So, feel free to take a look to their
  120. 5:23blog post including lots and lots of
  121. 5:27audio samples. At first, let's stick
  122. 5:30with the default and use the first
  123. 5:32option here, voice design,
  124. 5:34>> [snorts]
  125. 5:34>> with the default text language English
  126. 5:37and just hit generate.
  127. 5:40Generation took about 20 seconds here,
  128. 5:42and let's listen to the result.
  129. 5:46>> It's in the top drawer. Wait, it's
  130. 5:49empty? No way, that's impossible. I'm
  131. 5:51sure I put it there.
  132. 5:55>> Okay, I can hear this hint of panic in
  133. 5:57the voice. So, pretty impressive, in my
  134. 6:00opinion.
  135. 6:01In my opinion, please let me know what
  136. 6:03you think. And now, let's take another
  137. 6:06sample. I've entered the text. This is
  138. 6:08just a short example. Even if it is a
  139. 6:10neutral text, I am curious if we can
  140. 6:12still hear the panic as requested in
  141. 6:15voice description. Took around 27
  142. 6:18seconds, and let's listen to the result.
  143. 6:21>> This is just a short example.
  144. 6:23Even if it is a neutral text, I'm
  145. 6:25curious if we can still hear the panic
  146. 6:27as requested in voice description.
  147. 6:30>> I'm impressed.
  148. 6:31I'm really impressed.
  149. 6:33Good job.
  150. 6:35>> [snorts]
  151. 6:35>> So, let's switch to my mother tongue.
  152. 6:38So, let's use a German text. So, I just
  153. 6:42entered a German text phrase here. Das
  154. 6:44ist nur ein ganz langweiliger, neutraler
  155. 6:46und ziemlich inhaltsleerer Satz.
  156. 6:48Hier bin ich auf die Panik in der Stimme
  157. 6:50gespannt. Which basically translates to
  158. 6:52this is a boring, neutral, and
  159. 6:54empty phrase. I'm curious on the panic
  160. 6:58in this AI voice. And let's hit
  161. 7:00generate. The voice description is still
  162. 7:03in English, so no change in the voice
  163. 7:05description. See what we got.
  164. 7:08>> Das ist nur ein ganz langweiliger,
  165. 7:10neutraler und ziemlich inhaltsleerer
  166. 7:13Satz.
  167. 7:14Hier bin ich auf die Panik in der Stimme
  168. 7:16gespannt.
  169. 7:17>> This is not as good as it is in English.
  170. 7:20So, in English, the panic, I could hear
  171. 7:22the panic. In German, it is not really
  172. 7:26panic.
  173. 7:27Let's take a look to the voice clone tab
  174. 7:29here.
  175. 7:30So, you can upload your reference audio.
  176. 7:34And then the reference text, so what you
  177. 7:36have spoken obviously in the reference
  178. 7:38audio, and then you can enter the text
  179. 7:41your cloned AI voice should speak then.
  180. 7:44I could do this here, but I would like
  181. 7:47to hop on to my Google Colab notebook
  182. 7:50because I've added this into my notebook
  183. 7:52as Python code. So, we'll play around
  184. 7:55with the Google Colab notebook now. If
  185. 7:57you go on my Torsten Voice GitHub
  186. 8:00repository, you can see a YouTube folder
  187. 8:04and inside that folder, you can see the
  188. 8:07TWQNTTS
  189. 8:09notebook. Just open it here and you can
  190. 8:11see the code of the Google Colab. Copy
  191. 8:14the URL of this notebook
  192. 8:17and go to colab.google.com
  193. 8:22and hit file
  194. 8:24open notebook
  195. 8:27GitHub here and paste the GitHub URL
  196. 8:30here.
  197. 8:33This will directly open this Google
  198. 8:36Colab notebook I've created for you in
  199. 8:38the Google Colab environment. So, you
  200. 8:42can see on the right upper corner, we
  201. 8:44have the T4, so a GPU that you can
  202. 8:47choose from in the
  203. 8:49Google Colab environment. Once
  204. 8:51connected, let's go through the notebook
  205. 8:54step by step. So, the first step is to
  206. 8:56import torch on the Python dependency
  207. 8:59level and install a system socks package
  208. 9:02to do some audio magic. So, to run this
  209. 9:06special step, here's this play button.
  210. 9:08Whoa.
  211. 9:10I'm old and I wear glasses, but this is
  212. 9:12really
  213. 9:13big icon. So, big screen soon. But,
  214. 9:16let's hit this big play button here.
  215. 9:19While running the task, it is rotating
  216. 9:21that circle. Okay, took 13 seconds and
  217. 9:25you can see
  218. 9:26the output here. So, the installation
  219. 9:29and dependency stuff and all of this.
  220. 9:31So, let's go to the next
  221. 9:34uh script block here.
  222. 9:36Uh I've added this to do some
  223. 9:38GPU checks and give out or print out
  224. 9:40some cuter information. Let's hit play.
  225. 9:44And we can see we have one GPU in this
  226. 9:47notebook called Tesla T4
  227. 9:51with a total memory of around 14 GB of
  228. 9:54memory.
  229. 9:56And now the actual magic, let's run pip
  230. 9:58install QwenTTS.
  231. 10:01Task finished and you can see the
  232. 10:03typical output when you install Python
  233. 10:05packages using the Python package
  234. 10:07management pip.
  235. 10:09And next step is according to the readme
  236. 10:13to reduce GPU memory requirements, you
  237. 10:16can install this flash attention. So
  238. 10:19let's run this task. Step completed.
  239. 10:23Before we run the next task, so proxy
  240. 10:26localhost port to external network,
  241. 10:28let's go one step further. And the next
  242. 10:31step will be to run the web UI and this
  243. 10:34will be bind bound to some sort of
  244. 10:38localhost or 000 port, not 172 001, but
  245. 10:44to a not reachable
  246. 10:47port when you run it on Google Colab.
  247. 10:49And the port will be 8000 and this task
  248. 10:53means this magical fairy dust code block
  249. 10:57is to proxy that's local port 8000 to
  250. 11:01external available URL. So let's hit
  251. 11:05play and output will be some sort of
  252. 11:08cryptical
  253. 11:09um URL which is reachable then from my
  254. 11:13local system.
  255. 11:15And next task is to open the QwenTTS
  256. 11:18demo. You can run it with the 1.7
  257. 11:22billion parameters model or with the
  258. 11:240.
  259. 11:266 billion model. Let's run this task.
  260. 11:30And this will obviously download some
  261. 11:31model and model related checkpoints and
  262. 11:34configuration files. So, once you can
  263. 11:36see this running on local URL
  264. 11:39as the output, let's go up to our proxy
  265. 11:42and open this URL in a new tab.
  266. 11:46Okay, so don't ask me why.
  267. 11:49>> [laughter]
  268. 11:50>> During my testing, this cool code
  269. 11:52snippet here using this external
  270. 11:55proxying
  271. 11:57worked like a charm.
  272. 11:58Now,
  273. 12:00it did not work. I've no idea why.
  274. 12:03So, I did some testing and played a
  275. 12:05little bit around and found out that it
  276. 12:08is not required at all.
  277. 12:11So, what you can do instead is and I
  278. 12:14will update my Google Colab on my GitHub
  279. 12:17repository on that. You can see in the
  280. 12:20command running the Q and web UI with
  281. 12:24the port 8,000 or 8,200
  282. 12:29which I use for testing now.
  283. 12:31You can add minus minus share
  284. 12:34which will automatically
  285. 12:37not just running it on a local
  286. 12:39interface, but on a public available
  287. 12:42URL.
  288. 12:43So, now you can see already public URL
  289. 12:46and now let's hit this one. Woohoo! Here
  290. 12:50we are. So, finally we have the model.
  291. 12:53So, you can see the Qwen 3 TTS demo with
  292. 12:58the same fields we have seen on the
  293. 13:00official Hugging Face space. So, text to
  294. 13:02synthesize, a language to choose from,
  295. 13:06and some instructional or additional and
  296. 13:09optional instructions.
  297. 13:11Oh,
  298. 13:12okay. Now, as we have got this, let's go
  299. 13:16back to the
  300. 13:18to the Jupyter notebook. Stop
  301. 13:21the running web UI process
  302. 13:26and go one step further. And now we have
  303. 13:29voice design in English.
  304. 13:32This is
  305. 13:33the original
  306. 13:34block or code block from the GitHub
  307. 13:36repository. So, what you can see here,
  308. 13:39we have this Qwen 1.7
  309. 13:427 1.7 billion parameters model.
  310. 13:46And we have
  311. 13:47the language English
  312. 13:50and a text we would like to synthesize.
  313. 13:53So, this is a short test using English
  314. 13:55language with additional instructions
  315. 13:58young male speaker with a happy tone.
  316. 14:01And this will be saved as
  317. 14:04output voice design.wave.
  318. 14:07You can hit on the left navigation bar,
  319. 14:09there is files, and you can see, let's
  320. 14:12refresh once, there's just this sample
  321. 14:14data output file. Now, let's run this
  322. 14:18task here.
  323. 14:19Again, it will download some
  324. 14:22model-related files. The task is
  325. 14:24finished. And when we refresh the file
  326. 14:27there are listing here,
  327. 14:29we can see output voice design.wave. So,
  328. 14:32let's just
  329. 14:34download the file and let's listen.
  330. 14:37>> This is a short test using English
  331. 14:38language.
  332. 14:39>> So, what do you think? Is it a young
  333. 14:42male speaker with a happy tone?
  334. 14:44>> [snorts]
  335. 14:45>> So, let me know in the comments.
  336. 14:46And then next step, let's do this step
  337. 14:49with a German phrase. Let's hit play.
  338. 14:53And I'm using the same 1.7 billion
  339. 14:56parameter model, and we have this text
  340. 14:58to be spoken. Das ist nur ein kleiner
  341. 15:00Test. So, this is just a short test.
  342. 15:04And this time in German language and
  343. 15:07with an old female speaker with a happy
  344. 15:09tone. Result has been saved as
  345. 15:13voice design minus DE for German. So,
  346. 15:16let's download this file, too.
  347. 15:18>> [sighs and gasps]
  348. 15:19>> Das Das ein kleiner Test.
  349. 15:22>> Okay, it's a female voice. She she
  350. 15:23speaks
  351. 15:25German,
  352. 15:26but
  353. 15:27it's a heavy tone of English accent.
  354. 15:32>> Das ist nur ein kleiner Test.
  355. 15:34>> I'm not sure if I made something wrong
  356. 15:36or it's my mistake. Please let me know
  357. 15:38in the comments
  358. 15:40if you know what I'm doing wrong here,
  359. 15:43but the happy tone is missing in my
  360. 15:45opinion completely. So, this
  361. 15:49beginning is not really happy.
  362. 15:51>> [sighs]
  363. 15:52>> Again, you can use this Google Google
  364. 15:54you can use this Google Colab notebook
  365. 15:57as base and do some modifications for
  366. 15:59your own use case.
  367. 16:01Now, let's go and see the voice cloning
  368. 16:05functionality.
  369. 16:06And choose the model again and here you
  370. 16:10have a reference audio text,
  371. 16:13which you could record on the Hugging
  372. 16:15Face official demo space,
  373. 16:17or you can use it here from some
  374. 16:20download source.
  375. 16:22Let's take this one. I've created a
  376. 16:24German and an English reference short
  377. 16:26reference phrase
  378. 16:28to be downloaded here. This is a simple
  379. 16:30reference phrase for voice cloning tests
  380. 16:33in multiple open source text-to-speech
  381. 16:35projects. And this is the transcript
  382. 16:37here.
  383. 16:38And this is the phrase I would like to
  384. 16:41be synthesized from my direct voice
  385. 16:44clone. So, let's run this task. Task is
  386. 16:47finished. Let's refresh directory
  387. 16:49listing and here we have the voice
  388. 16:51clone.
  389. 16:52And again, download and listen. I'm
  390. 16:54curious how close this cloned voice will
  391. 16:56be compared to my original reference
  392. 16:58phrase.
  393. 17:01>> [snorts]
  394. 17:01>> And I guess this is pretty close. So, if
  395. 17:04we see this is the text to be spoken,
  396. 17:07I'm curious in how close the voice
  397. 17:10and let's just compare the original
  398. 17:12reference phrase. So, this the is the
  399. 17:14phrase I've recorded manually. This is a
  400. 17:16simple reference phrase for voice
  401. 17:18cloning tests in multiple
  402. 17:21I'm curious how close this cloned voice
  403. 17:23will be compared to my original
  404. 17:25reference phrase.
  405. 17:27I really like this one. So, let's do
  406. 17:29this in German language. So, the same
  407. 17:33code block,
  408. 17:34we'll use
  409. 17:36German reference phrase. So, let's just
  410. 17:38download
  411. 17:40the ref in DE.
  412. 17:50Okay, so this is the reference
  413. 17:52audio and the transcribed text. And the
  414. 17:56text to be spoken I
  415. 18:00So, I'm curious how close uh my AI voice
  416. 18:04is compared to my original voice.
  417. 18:06Uh language German and should be saved
  418. 18:09as voice clone DE. Oh, let's hit save.
  419. 18:12No, play.
  420. 18:14And while this is downloading and um
  421. 18:17doing its magic, and before you grab
  422. 18:19yourself something to drink, do not
  423. 18:21forget to give this video a thumbs up,
  424. 18:24subscribe to my channel, and share my
  425. 18:26Thorsten Voice YouTube channel with
  426. 18:28other voice technology open source
  427. 18:31enthusiasts. So, thank you for all your
  428. 18:33support.
  429. 18:35So, task is finished and we have our uh
  430. 18:38voice clone DE wave audio. Let's listen.
  431. 18:50What do you think? Obviously, I cannot
  432. 18:52be objective in this case. It's my
  433. 18:53personal voice. Let's do direct
  434. 18:56comparison between
  435. 18:57my originally manual recording and the
  436. 19:01cloned version.
  437. 19:14>> Impressive, in my opinion.
  438. 19:17But again, please let me know in the
  439. 19:18comments on what you think on this voice
  440. 19:21cloning, voice designing, and on the
  441. 19:24capabilities of Qwen TTS in general.
  442. 19:29So, and for now, that's it. I really
  443. 19:32hope you like this video, you enjoyed
  444. 19:34it. And
  445. 19:36yeah, I'd be happy to welcome you to my
  446. 19:38next video on this channel. And until
  447. 19:40then, have a nice day. Thank you, and
  448. 19:43bye-bye.

About this transcript

This page contains the full transcript of Qwen3-TTS Tutorial: Open-Source Voice Design & Cloning by Thorsten-Voice, generated from the public captions YouTube serves with the video. The transcript has 2,628 words across 448 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.