Qwen3-TTS Tutorial: Open-Source Voice Design & Cloning — Transcript
Full transcript
- 0:00Good, today is February 1st, 2026 and
- 0:04I'm taking a hands-on look at Qwen 3
- 0:07TTS, the newly open-source
- 0:09text-to-speech family from the Qwen
- 0:12team. Qwen 3 TTS brings together voice
- 0:16design, a 3-second voice cloning, and
- 0:19instruction-based voice control with
- 0:21support for
- 0:23>> [clears throat]
- 0:23>> Chinese, English, Japanese, Korean,
- 0:26German, French, Russian, Portuguese,
- 0:29Spanish, and Italian.
- 0:31In this tutorial, I'll take a look at
- 0:33the GitHub repository and play around
- 0:36with their Hugging Face demo space.
- 0:39To make everything easy to follow, I've
- 0:41created a Google Colab notebook that you
- 0:44can use as a starting point and adapt
- 0:47for your own experiments.
- 0:49In it, I will show you how to start
- 0:52WebUI, generate voices from descriptions
- 0:55in English and German, and clone voices
- 0:58from short reference samples.
- 1:01If you enjoy practical hands-on AI voice
- 1:04tutorials like this one, consider liking
- 1:07the video and subscribing to the
- 1:09channel. And do not forget to share my
- 1:12channel with other open-source voice
- 1:14technology enthusiasts.
- 1:17Wow, that's all for today.
- 1:20This is all for the introduction and now
- 1:22let's let's go. Let's start taking a
- 1:25look to their GitHub page and as always,
- 1:28all the links in the description box
- 1:29below. So, first of all, the license, it
- 1:32is Apache 2 licensed and they gained
- 1:366,000 stars already since their release
- 1:401 week ago. So, let's take a look to
- 1:43their commit history
- 1:45and you can see their initial commit has
- 1:48just been
- 1:51last week. So, just 1 week, 6,000 stars
- 1:56on GitHub.
- 1:58I mean, I'm not really jealous because
- 1:59it took me over 3 years to gain 10,000
- 2:03subscribers on my YouTube channel, but
- 2:06so they got 6,000 stars
- 2:09in 1 week.
- 2:10The code is purely Python-based, as you
- 2:13can see here.
- 2:14On top of the read me, you can see some
- 2:17links to hugging face model information
- 2:19blocks, a demo space. We'll take a
- 2:22closer look in a moment.
- 2:24And we have this one news here, so from
- 2:26January 22nd, where they released their
- 2:30initial release, and this contains two
- 2:33models. We have a 0.6
- 2:36billion model parameters model and a 1.7
- 2:40billion parameters model.
- 2:42Here you can see Qwen text-to-speech
- 2:44covers 10 major languages, so
- 2:47Chinese, English, Japanese, Korean,
- 2:50German, French, Russian, Portuguese,
- 2:51Spanish, and Italian.
- 2:53And here you can see intelligent text
- 2:55understanding and voice control. So,
- 2:57this is this type of voice design where
- 3:00you can describe
- 3:03attributes for your voice, timbre,
- 3:05emotion,
- 3:06prosody, so these types of information
- 3:10to design and create
- 3:12a text or textually described AI voice.
- 3:16In this table, you can see a comparison
- 3:18of the features depending on their
- 3:21models. Again, we have this 1.7
- 3:24billion parameters models, and we have
- 3:27these 0.6 billion parameters models
- 3:31here. In the quick start section, they
- 3:34create a Python 3.12 virtual environment
- 3:37using conda,
- 3:39and you can install it pretty simple
- 3:41using Qwen TTS Python package. We will
- 3:44come to that in a moment when we take a
- 3:46look to my Google Colab notebook. In
- 3:50general, they provide lots of Python
- 3:53code snippets and examples for all their
- 3:56use cases, voice design and voice
- 3:58cloning. So,
- 4:01pretty good documentation here on their
- 4:03readme.
- 4:04I have to say, I guess they deserve the
- 4:066,000 stars in 1 week.
- 4:09Before we hop onto their hugging face
- 4:12demo space to play a little bit around
- 4:14with,
- 4:15um let's take a short look to their blog
- 4:17post here. Let's scroll down to the
- 4:20audio samples. You can see here some
- 4:24control instructions mentioned here. So,
- 4:27here's control instruction, gender male,
- 4:30pitch low pitch, so describing your
- 4:33design voice.
- 4:34And we have this text here, so let's
- 4:37just listen to this sample.
- 4:39>> Nine different exciting ways of cooking
- 4:42sausage. Incredible. There were three
- 4:45outstanding deliveries in terms of the
- 4:47sausage being the hero. The first dish
- 4:50that we want to dissect, this
- 4:52individual's
- 4:54>> Okay, let's listen to a Chinese sample,
- 4:57which I cannot understand at all.
- 5:12Okay, I do not understand
- 5:15any word, but I understand or I can I
- 5:17can feel the emotion, which is pretty
- 5:20impressive.
- 5:21So, feel free to take a look to their
- 5:23blog post including lots and lots of
- 5:27audio samples. At first, let's stick
- 5:30with the default and use the first
- 5:32option here, voice design,
- 5:34>> [snorts]
- 5:34>> with the default text language English
- 5:37and just hit generate.
- 5:40Generation took about 20 seconds here,
- 5:42and let's listen to the result.
- 5:46>> It's in the top drawer. Wait, it's
- 5:49empty? No way, that's impossible. I'm
- 5:51sure I put it there.
- 5:55>> Okay, I can hear this hint of panic in
- 5:57the voice. So, pretty impressive, in my
- 6:00opinion.
- 6:01In my opinion, please let me know what
- 6:03you think. And now, let's take another
- 6:06sample. I've entered the text. This is
- 6:08just a short example. Even if it is a
- 6:10neutral text, I am curious if we can
- 6:12still hear the panic as requested in
- 6:15voice description. Took around 27
- 6:18seconds, and let's listen to the result.
- 6:21>> This is just a short example.
- 6:23Even if it is a neutral text, I'm
- 6:25curious if we can still hear the panic
- 6:27as requested in voice description.
- 6:30>> I'm impressed.
- 6:31I'm really impressed.
- 6:33Good job.
- 6:35>> [snorts]
- 6:35>> So, let's switch to my mother tongue.
- 6:38So, let's use a German text. So, I just
- 6:42entered a German text phrase here. Das
- 6:44ist nur ein ganz langweiliger, neutraler
- 6:46und ziemlich inhaltsleerer Satz.
- 6:48Hier bin ich auf die Panik in der Stimme
- 6:50gespannt. Which basically translates to
- 6:52this is a boring, neutral, and
- 6:54empty phrase. I'm curious on the panic
- 6:58in this AI voice. And let's hit
- 7:00generate. The voice description is still
- 7:03in English, so no change in the voice
- 7:05description. See what we got.
- 7:08>> Das ist nur ein ganz langweiliger,
- 7:10neutraler und ziemlich inhaltsleerer
- 7:13Satz.
- 7:14Hier bin ich auf die Panik in der Stimme
- 7:16gespannt.
- 7:17>> This is not as good as it is in English.
- 7:20So, in English, the panic, I could hear
- 7:22the panic. In German, it is not really
- 7:26panic.
- 7:27Let's take a look to the voice clone tab
- 7:29here.
- 7:30So, you can upload your reference audio.
- 7:34And then the reference text, so what you
- 7:36have spoken obviously in the reference
- 7:38audio, and then you can enter the text
- 7:41your cloned AI voice should speak then.
- 7:44I could do this here, but I would like
- 7:47to hop on to my Google Colab notebook
- 7:50because I've added this into my notebook
- 7:52as Python code. So, we'll play around
- 7:55with the Google Colab notebook now. If
- 7:57you go on my Torsten Voice GitHub
- 8:00repository, you can see a YouTube folder
- 8:04and inside that folder, you can see the
- 8:07TWQNTTS
- 8:09notebook. Just open it here and you can
- 8:11see the code of the Google Colab. Copy
- 8:14the URL of this notebook
- 8:17and go to colab.google.com
- 8:22and hit file
- 8:24open notebook
- 8:27GitHub here and paste the GitHub URL
- 8:30here.
- 8:33This will directly open this Google
- 8:36Colab notebook I've created for you in
- 8:38the Google Colab environment. So, you
- 8:42can see on the right upper corner, we
- 8:44have the T4, so a GPU that you can
- 8:47choose from in the
- 8:49Google Colab environment. Once
- 8:51connected, let's go through the notebook
- 8:54step by step. So, the first step is to
- 8:56import torch on the Python dependency
- 8:59level and install a system socks package
- 9:02to do some audio magic. So, to run this
- 9:06special step, here's this play button.
- 9:08Whoa.
- 9:10I'm old and I wear glasses, but this is
- 9:12really
- 9:13big icon. So, big screen soon. But,
- 9:16let's hit this big play button here.
- 9:19While running the task, it is rotating
- 9:21that circle. Okay, took 13 seconds and
- 9:25you can see
- 9:26the output here. So, the installation
- 9:29and dependency stuff and all of this.
- 9:31So, let's go to the next
- 9:34uh script block here.
- 9:36Uh I've added this to do some
- 9:38GPU checks and give out or print out
- 9:40some cuter information. Let's hit play.
- 9:44And we can see we have one GPU in this
- 9:47notebook called Tesla T4
- 9:51with a total memory of around 14 GB of
- 9:54memory.
- 9:56And now the actual magic, let's run pip
- 9:58install QwenTTS.
- 10:01Task finished and you can see the
- 10:03typical output when you install Python
- 10:05packages using the Python package
- 10:07management pip.
- 10:09And next step is according to the readme
- 10:13to reduce GPU memory requirements, you
- 10:16can install this flash attention. So
- 10:19let's run this task. Step completed.
- 10:23Before we run the next task, so proxy
- 10:26localhost port to external network,
- 10:28let's go one step further. And the next
- 10:31step will be to run the web UI and this
- 10:34will be bind bound to some sort of
- 10:38localhost or 000 port, not 172 001, but
- 10:44to a not reachable
- 10:47port when you run it on Google Colab.
- 10:49And the port will be 8000 and this task
- 10:53means this magical fairy dust code block
- 10:57is to proxy that's local port 8000 to
- 11:01external available URL. So let's hit
- 11:05play and output will be some sort of
- 11:08cryptical
- 11:09um URL which is reachable then from my
- 11:13local system.
- 11:15And next task is to open the QwenTTS
- 11:18demo. You can run it with the 1.7
- 11:22billion parameters model or with the
- 11:240.
- 11:266 billion model. Let's run this task.
- 11:30And this will obviously download some
- 11:31model and model related checkpoints and
- 11:34configuration files. So, once you can
- 11:36see this running on local URL
- 11:39as the output, let's go up to our proxy
- 11:42and open this URL in a new tab.
- 11:46Okay, so don't ask me why.
- 11:49>> [laughter]
- 11:50>> During my testing, this cool code
- 11:52snippet here using this external
- 11:55proxying
- 11:57worked like a charm.
- 11:58Now,
- 12:00it did not work. I've no idea why.
- 12:03So, I did some testing and played a
- 12:05little bit around and found out that it
- 12:08is not required at all.
- 12:11So, what you can do instead is and I
- 12:14will update my Google Colab on my GitHub
- 12:17repository on that. You can see in the
- 12:20command running the Q and web UI with
- 12:24the port 8,000 or 8,200
- 12:29which I use for testing now.
- 12:31You can add minus minus share
- 12:34which will automatically
- 12:37not just running it on a local
- 12:39interface, but on a public available
- 12:42URL.
- 12:43So, now you can see already public URL
- 12:46and now let's hit this one. Woohoo! Here
- 12:50we are. So, finally we have the model.
- 12:53So, you can see the Qwen 3 TTS demo with
- 12:58the same fields we have seen on the
- 13:00official Hugging Face space. So, text to
- 13:02synthesize, a language to choose from,
- 13:06and some instructional or additional and
- 13:09optional instructions.
- 13:11Oh,
- 13:12okay. Now, as we have got this, let's go
- 13:16back to the
- 13:18to the Jupyter notebook. Stop
- 13:21the running web UI process
- 13:26and go one step further. And now we have
- 13:29voice design in English.
- 13:32This is
- 13:33the original
- 13:34block or code block from the GitHub
- 13:36repository. So, what you can see here,
- 13:39we have this Qwen 1.7
- 13:427 1.7 billion parameters model.
- 13:46And we have
- 13:47the language English
- 13:50and a text we would like to synthesize.
- 13:53So, this is a short test using English
- 13:55language with additional instructions
- 13:58young male speaker with a happy tone.
- 14:01And this will be saved as
- 14:04output voice design.wave.
- 14:07You can hit on the left navigation bar,
- 14:09there is files, and you can see, let's
- 14:12refresh once, there's just this sample
- 14:14data output file. Now, let's run this
- 14:18task here.
- 14:19Again, it will download some
- 14:22model-related files. The task is
- 14:24finished. And when we refresh the file
- 14:27there are listing here,
- 14:29we can see output voice design.wave. So,
- 14:32let's just
- 14:34download the file and let's listen.
- 14:37>> This is a short test using English
- 14:38language.
- 14:39>> So, what do you think? Is it a young
- 14:42male speaker with a happy tone?
- 14:44>> [snorts]
- 14:45>> So, let me know in the comments.
- 14:46And then next step, let's do this step
- 14:49with a German phrase. Let's hit play.
- 14:53And I'm using the same 1.7 billion
- 14:56parameter model, and we have this text
- 14:58to be spoken. Das ist nur ein kleiner
- 15:00Test. So, this is just a short test.
- 15:04And this time in German language and
- 15:07with an old female speaker with a happy
- 15:09tone. Result has been saved as
- 15:13voice design minus DE for German. So,
- 15:16let's download this file, too.
- 15:18>> [sighs and gasps]
- 15:19>> Das Das ein kleiner Test.
- 15:22>> Okay, it's a female voice. She she
- 15:23speaks
- 15:25German,
- 15:26but
- 15:27it's a heavy tone of English accent.
- 15:32>> Das ist nur ein kleiner Test.
- 15:34>> I'm not sure if I made something wrong
- 15:36or it's my mistake. Please let me know
- 15:38in the comments
- 15:40if you know what I'm doing wrong here,
- 15:43but the happy tone is missing in my
- 15:45opinion completely. So, this
- 15:49beginning is not really happy.
- 15:51>> [sighs]
- 15:52>> Again, you can use this Google Google
- 15:54you can use this Google Colab notebook
- 15:57as base and do some modifications for
- 15:59your own use case.
- 16:01Now, let's go and see the voice cloning
- 16:05functionality.
- 16:06And choose the model again and here you
- 16:10have a reference audio text,
- 16:13which you could record on the Hugging
- 16:15Face official demo space,
- 16:17or you can use it here from some
- 16:20download source.
- 16:22Let's take this one. I've created a
- 16:24German and an English reference short
- 16:26reference phrase
- 16:28to be downloaded here. This is a simple
- 16:30reference phrase for voice cloning tests
- 16:33in multiple open source text-to-speech
- 16:35projects. And this is the transcript
- 16:37here.
- 16:38And this is the phrase I would like to
- 16:41be synthesized from my direct voice
- 16:44clone. So, let's run this task. Task is
- 16:47finished. Let's refresh directory
- 16:49listing and here we have the voice
- 16:51clone.
- 16:52And again, download and listen. I'm
- 16:54curious how close this cloned voice will
- 16:56be compared to my original reference
- 16:58phrase.
- 17:01>> [snorts]
- 17:01>> And I guess this is pretty close. So, if
- 17:04we see this is the text to be spoken,
- 17:07I'm curious in how close the voice
- 17:10and let's just compare the original
- 17:12reference phrase. So, this the is the
- 17:14phrase I've recorded manually. This is a
- 17:16simple reference phrase for voice
- 17:18cloning tests in multiple
- 17:21I'm curious how close this cloned voice
- 17:23will be compared to my original
- 17:25reference phrase.
- 17:27I really like this one. So, let's do
- 17:29this in German language. So, the same
- 17:33code block,
- 17:34we'll use
- 17:36German reference phrase. So, let's just
- 17:38download
- 17:40the ref in DE.
- 17:50Okay, so this is the reference
- 17:52audio and the transcribed text. And the
- 17:56text to be spoken I
- 18:00So, I'm curious how close uh my AI voice
- 18:04is compared to my original voice.
- 18:06Uh language German and should be saved
- 18:09as voice clone DE. Oh, let's hit save.
- 18:12No, play.
- 18:14And while this is downloading and um
- 18:17doing its magic, and before you grab
- 18:19yourself something to drink, do not
- 18:21forget to give this video a thumbs up,
- 18:24subscribe to my channel, and share my
- 18:26Thorsten Voice YouTube channel with
- 18:28other voice technology open source
- 18:31enthusiasts. So, thank you for all your
- 18:33support.
- 18:35So, task is finished and we have our uh
- 18:38voice clone DE wave audio. Let's listen.
- 18:50What do you think? Obviously, I cannot
- 18:52be objective in this case. It's my
- 18:53personal voice. Let's do direct
- 18:56comparison between
- 18:57my originally manual recording and the
- 19:01cloned version.
- 19:14>> Impressive, in my opinion.
- 19:17But again, please let me know in the
- 19:18comments on what you think on this voice
- 19:21cloning, voice designing, and on the
- 19:24capabilities of Qwen TTS in general.
- 19:29So, and for now, that's it. I really
- 19:32hope you like this video, you enjoyed
- 19:34it. And
- 19:36yeah, I'd be happy to welcome you to my
- 19:38next video on this channel. And until
- 19:40then, have a nice day. Thank you, and
- 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.