Best Open Source Subtitle Generator? Canary Qwen 2.5B + Whisper Full Guide

[FurkanGozukara]

Best Open Source Subtitle Generator? Canary Qwen 2.5B + Whisper Full Guide

Full tutorial: https://www.youtube.com/watch?v=4lAk6sf1qF8

NVIDIA NeMo Canary-Qwen-2.5B is an English speech recognition model that achieves state-of-the art performance on multiple English speech benchmarks. Canary model is the new king that dethroned famous Whisper.

Full tutorial for the Whisper TTS Premium speech-to-text app by SECourses with new NVIDIA Canary Qwen 2.5B support. In this video, I demo local subtitle generation, compare Canary Qwen 2.5B against Whisper Large V3, show output formats, batch processing, presets, YouTube URL and live microphone options, then install the app from scratch on Windows.

You will also see RunPod and Massed Compute notes, first-run model download, RTX 5000/CUDA 13 driver requirements, subprocess mode for preventing VRAM/RAM leaks, and when to use Whisper instead of Canary.

Links:

Download App and the source post: [ https://www.patreon.com/posts/whisper-webui-to-145395299 ]

Discord: [ https://discord.com/channels/772774097734074388/1079506787734134844 ]

Patreon app index: [ https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/Patreon-Posts-Index.md ]

Related RunPod/Massed Compute setup tutorial: [ https://youtu.be/ZRrzvD4wNys ]

In my tutorial-video tests, Canary Qwen 2.5B achieved 5.91% global WER and reached up to 46x faster than real-time transcription, making it my new recommended default for English speech-to-text. Whisper remains useful when you need broader spoken-language support or word-level timestamps.

Chapters:

00:00:00 Intro to the local open-source speech-to-text app and new Canary support

00:00:20 Quick demo setup with NVIDIA Canary Qwen 2.5B

00:00:33 Maximum-quality defaults and starting subtitle generation

00:00:48 Live transcription speed and accuracy preview

00:01:03 Chunk length settings for smaller or larger subtitle segments

00:01:18 Fast generation, supported exports, and restarting with all formats

00:01:31 Multiple subtitle file formats explained

00:01:47 Batch processing folders, output paths, subfolders, and overwrite mode

00:01:58 YouTube URLs, microphone/live transcription, translation, and BGM separation

00:02:09 Saving presets and using advanced parameters

00:02:24 Auto-optimized defaults for Whisper and Canary models

00:02:39 Canary Qwen 2.5B vs Whisper Large V3 comparison begins

00:02:54 Real-world WER benchmark and 5.91% Canary result

00:03:10 Why non-native English speech is harder to transcribe accurately

00:03:24 Canary speed advantage and 46x real-time transcription explained

00:03:43 Test averages across long and short tutorial videos

00:03:59 Cases where Whisper slightly wins and final Canary recommendation

00:04:14 Opening the output folder after transcription completes

00:04:27 VTT output, matching filenames, capitalization, and punctuation

00:04:44 Accuracy examples inside the generated transcript

00:04:58 TXT, TSV, SRT, LRC exports and word-level timestamp note

00:05:20 Download page, latest ZIP, and installation overview

00:05:31 Windows requirements: Python 3.11, Git, CUDA, and C++ notes

00:05:51 Choosing install location and keeping the app isolated in venv

00:06:04 Extracting the ZIP and running Windows install/update BAT

00:06:23 Automatic model downloads on first run

00:06:34 RunPod, Massed Compute, and Linux installation files

00:06:50 Where to learn RunPod and Massed Compute setup in the related guide

00:07:29 UV-powered Windows installation completes quickly

00:07:41 Starting the app with Windows start app BAT

00:07:58 Selecting video/audio input and generating subtitles on a fresh install

00:08:10 First-run Canary model download and 5GB model size

00:08:35 Easy setup goal and automatic fresh-install workflow

00:08:53 Discord, Patreon index, and 100+ SECourses applications

00:09:13 RTX 5000 support and updated NVIDIA driver requirement

00:09:35 Fresh-install transcription starts successfully

00:09:47 Automatic downloads for Canary, Whisper, diarization, and extra tools

00:10:16 Canary becomes the new default model recommendation

00:10:36 Subprocess mode to prevent VRAM and RAM leaks

00:10:51 Why running transcription as a subprocess is recommended

00:11:04 Switching back to Whisper models when needed

00:11:20 Whisper language coverage vs Canary and audio/video support

00:11:42 Real recording benchmark: 27 minutes transcribed in about 2 minutes

00:11:56 Model loading overhead and clean RAM/VRAM release

00:12:08 Final notes, subscribe reminder, and downloading the full transcript ZIP

#CanaryQwen #Whisper #SpeechToText #SubtitleGenerator #OpenSourceAI #LocalAI #NVIDIA #RunPod #MassedCompute #SECourses

Video Transcription

• 00:00:00 Greetings everyone. Today I am going to  show you state of the art, open source,  

• 00:00:06 locally runnable speech-to-text application.  It is called as Whisper TTS Premium application  

• 00:00:13 by SECourses. However, with today's update,  now we are supporting Canary model as well,  

• 00:00:20 NVIDIA Canary Qwen 2.5 billion parameters. Let's make a quick demonstration. So this  

• 00:00:26 is my latest tutorial video. I have drag and drop  it. The default parameters are all set for maximum  

• 00:00:34 quality. So generate subtitle, and let's watch  live how it is being transcribed. It is starting.  

• 00:00:43 Okay, and the transcription started quickly. It  is really, really fast and it is the best quality,  

• 00:00:50 best accuracy speech to text model right now. So  as we can see live, it is currently transcribing  

• 00:00:59 my video fully automatically. Currently it  is transcribing as 40 second chunks, but if  

• 00:01:06 you want smaller chunks for better subtitles or  whatever reason, you can go to advanced parameters  

• 00:01:14 and change the chunk length to whichever  parameter you like, like 20, 30, 10, 50, it is  

• 00:01:21 up to you. And the file is about to be generated. So this application is amazing. It is supporting  

• 00:01:27 so many features like you can get with multiple  file formats, but we should have started that  

• 00:01:34 way. Okay, let's cancel generation and start again  because it is fast. And generate subtitle file. So  

• 00:01:43 it will generate all these formats and save with  all of them. It supports full batch processing, so  

• 00:01:50 you can give a folder path and output folder path,  it will process all the videos and audio files  

• 00:01:55 inside that. You can also include subdirectories,  overwrite existing files. It supports YouTube  

• 00:02:01 URLs, microphone, live transcription,  some translation but it is not very good,  

• 00:02:06 and BGM separation. So you can also save and  load your presets from here. Fully saving and  

• 00:02:13 loading presets. So we support all of the advanced  parameters you can also set, but these parameters  

• 00:02:20 are automatically set to best defaults for both  Whisper models and also Canary model. For example,  

• 00:02:30 if I change the model from here, Canary  to Whisper, the advanced parameters will  

• 00:02:36 be accordingly changed, according to the model. Meanwhile it is transcribing again, let's see the  

• 00:02:43 difference between NVIDIA Canary Qwen 2.5 billion  versus Whisper Large version 3, which is the  

• 00:02:51 latest version of the Whisper model. So currently  Canary Qwen 2.5 has 5.91 percentage global word  

• 00:03:02 error rate. I have tested these on my tutorial  videos, so this is a real test. I am not a  

• 00:03:08 native English speaker, therefore it is harder  to accurately transcribe speech to text my voice.  

• 00:03:16 However, it did an excellent job. You see the  global word error rate of the Whisper is much  

• 00:03:22 bigger than the Canary. Moreover, the Canary  speech to text speed is much faster. This is  

• 00:03:29 46 times faster than real-time. Which means 46  minutes of speech is transcribed in 1 minute.  

• 00:03:39 You can see the values. So the Canary model is  much better at every statistic compared to the  

• 00:03:47 Whisper. And there are individual test results as  well, as you are seeing. So I did run big tests  

• 00:03:55 like 1 hour, like 5 minute, like 13 minute, 55  minutes, and I did get the overall averages. In  

• 00:04:02 few cases, the Whisper did perform very slightly  better as you can see. So this new model is my  

• 00:04:10 recommendation, use it as a default model. Moreover, it is also default set in our  

• 00:04:18 application. So let's look at the results, because  the transcription has been completed. When you  

• 00:04:24 click to open outputs folder, it will open you  the outputs folder, you will see the transcription  

• 00:04:31 names are same as your input file name. So we  have VTT format. You can see VTT format. And as  

• 00:04:39 you can see, it is also having proper uppercase  and lowercase characters as well. Also having the  

• 00:04:48 accurate syntax or punctuation as well. I mean,  you see installers zip file. This is amazing.  

• 00:04:55 It also supports TXT format as you can see, or  TSV format as you can see, our classic subtitle  

• 00:05:03 format SRT, or LRC format. So we support  all of the formats. Currently this model  

• 00:05:11 doesn't have word-level timestamps, but Whisper  also has it if you need word-level timestamps. 

• 00:05:19 So this application is the ultimate, very best  application right now. How you are going to  

• 00:05:24 install and use it? We have an excellent page  as usual. The link will be in the description  

• 00:05:29 of the video. Download the latest zip file.  Before starting installation, make sure that  

• 00:05:34 you have read and follow the requirements. So  the requirements are here, Windows requirements.  

• 00:05:41 We need Python 3.11. Moreover, you need Git. CUDA  and C++ tools probably won't be necessary since I  

• 00:05:49 have pre-compiled all the libraries and all of my  applications install into a virtual environment.  

• 00:05:56 So move the latest zip file into wherever you  want to install. So let's install into our  

• 00:06:02 Q drive T2. Do not have space characters in your  installations. Extract files here. Then all I need  

• 00:06:10 to do is Windows install and update dot bat file.  So it will fully automatically install everything  

• 00:06:16 for me. It will generate a virtual environment  and install all the libraries inside here,  

• 00:06:21 so it is fully isolated. The models will  be automatically downloaded into models  

• 00:06:26 folder when you first time run them. All of  them will be fully automatically downloaded. 

• 00:06:31 For RunPod we have RunPod Simple Pod  instructions. Please follow these files.  

• 00:06:36 It has all the instructions as you can see.  We also support Massed Compute. Please follow  

• 00:06:41 this. If you have a Linux users, use the Massed  Compute install sh file. If you don't know how  

• 00:06:47 to use RunPod or Massed Compute, our latest  tutorial in details explain how to use them.  

• 00:06:55 So you can follow the ultimate guide to AI voice  cloning tutorial. This tutorial has full chapters,  

• 00:07:02 so you can go to video description and find  the Simple Pod RunPod part if you don't know,  

• 00:07:09 or Massed Compute part. For example, let's search  for Massed Compute. So the Massed Compute part  

• 00:07:15 starts here and the RunPod compute part starts  here. So you can follow this tutorial to learn  

• 00:07:20 how to install and use this application or all  of my applications on RunPod and Massed Compute. 

• 00:07:27 The installation on Windows will be very  fast as you can see because we are using  

• 00:07:31 UV installation and it is already completed.  Now all you need to do is just double click  

• 00:07:36 Windows start app dot bat file and it will start  the application and it will be right away ready  

• 00:07:43 to use. So the application started. Let's make  a test and see how it downloads automatically  

• 00:07:49 the model. So I have selected my video. This  supports both videos and audio files. Canary is  

• 00:07:56 selected. It is all default. You can also set  all of the options that you want and generate  

• 00:08:01 subtitle file. This is a fresh installation,  so first it need to download model. Okay,  

• 00:08:07 as you can see the model download started. So it  will download the model, it is around 5 gigabytes  

• 00:08:13 for this new model. It is downloading pretty fast,  it depends on your internet connection. Once the  

• 00:08:19 download has been completed, it will start. I  can see that it is using my entire internet right  

• 00:08:24 now. It is using 1 gigabits per second. So after this download we will see the  

• 00:08:31 transcription started on a fresh installation. I  am making all of my applications extremely easy  

• 00:08:39 to install, set up, and use. You can find our  Discord server in the link in the source post.  

• 00:08:48 You see we have Patreon exclusive post index,  plus we have SECourses Discord. When you click  

• 00:08:55 this link you will get to our Discord page.  And in the index page you can see all of  

• 00:09:02 our applications. We have more than 100 unique  applications and we are adding more and keeping  

• 00:09:07 all of them up to date and working. All of my  applications works on RTX 5000 series as well,  

• 00:09:13 just make sure that you have updated NVIDIA driver  because CUDA 13 requires updated NVIDIA driver.  

• 00:09:20 For example, when I type NVIDIA-smi, I should  see an updated driver like this and I should  

• 00:09:26 see CUDA version 13 or bigger on this screen when  I type NVIDIA-smi. Make sure that it is like this. 

• 00:09:35 Alright, so our transcription started with  the fresh installation. As you can see,  

• 00:09:40 it is working amazingly. This model requires a lot  of requirements, but I have made it so easy and  

• 00:09:47 so fluent to install and use, start using right  away. The models are automatically downloaded,  

• 00:09:53 both for Canary, Canary from NVIDIA with Qwen 2.5,  and both for Whisper. Even the other models like  

• 00:10:03 diarization and some background remove or other  features that we support, all of them is fully  

• 00:10:11 automatically downloaded and done. Transcription  almost completed. You can see the speed,  

• 00:10:16 it is just amazingly fast. I like this model, so  this model is my now new default model, but always  

• 00:10:23 you can also use other models to test. If you  encounter any issues, just message me from Patreon  

• 00:10:30 or from Discord or from YouTube, leave a comment.  Check out all the features of the application. 

• 00:10:36 One another thing that is, we are starting the  process as a subprocess, therefore after the  

• 00:10:43 transcription has been completed, there will  be absolutely zero VRAM leak or the RAM leak.  

• 00:10:51 So this is a very big advantage. If you don't  want as a subprocess, you can uncheck that box,  

• 00:10:57 but I recommend to run as a subprocess because it  is extremely effective. The transcription speed is  

• 00:11:04 amazing with this new Canary model, but always  you can switch to the Whisper and use all of  

• 00:11:12 these Whisper models. One advantage of Whisper is  that it supports 100 different spoken languages.  

• 00:11:20 However, if you are going to transcribe English  audio, I recommend you to use Canary. Our  

• 00:11:29 interface, as I have said, supports both audio  and video files. I also show preview like this,  

• 00:11:35 this is a feature that I have coded. So this  application is extremely improved by me. It has so  

• 00:11:42 many new features, you won't find this application  anywhere else. And you can see that while I am  

• 00:11:47 recording a video, the transcription speed was 12  times real-time. So 27 minute took only 2 minutes  

• 00:11:56 to transcribe. Most of the time is spent for  loading the model into the VRAM and starting  

• 00:12:02 transcription. We can also see that it is entirely  cleared. The RAM, VRAM, everything is cleared. 

• 00:12:08 So thank you so much for watching. Hopefully  I will start making more videos. Please open  

• 00:12:13 the bell, subscribe. You can also download  the transcription from here. You see it is  

• 00:12:18 downloaded as a zip file and it includes all of  the transcription files. This is very useful when  

• 00:12:24 you are working with RunPod or Massed Compute.  Thank you so much. Hopefully see you later.