Users Say IndexTTS 25 Voice Cloning Beats ElevenLabs Full Windows Tutorial

Users Say IndexTTS 2.5 Voice Cloning Beats ElevenLabs | Full Windows Tutorial

Full tutorial link > https://www.youtube.com/watch?v=jf5IlYaJvdc

Make your own voice speak words you never recorded. This complete IndexTTS 2.5 Windows tutorial starts with real generated speech, then takes you through a fresh installation, reference-based voice cloning, longer narration, expression controls, timed video, folder batches and training a voice from original recordings. Follow the actual local application from a short sample and script to saved audio and a trained voice setup.

Links:

Download IndexTTS 2.5 and read the full feature guide: [ https://www.patreon.com/SECourses/posts/indextts-2-5-and-139297407 ]

Windows Requirements tutorial: [ https://youtu.be/DrhUHnYfwC0 ]

Updated Windows Requirements instructions: [ https://www.patreon.com/posts/windows-AI-requirements-tutorial-111553210 ]

You will learn how to find the current ZIP, locate it in Chrome, move it to a permanent folder and extract the whole package. Complete the linked Windows requirements first. Choose a path using English characters with no spaces, then run Windows_Install_or_Update.bat and Windows_Start_App.bat normally. Double-click the BAT files or select one and press Enter; do not run them as administrator. Keep the app command window open while you use its local browser interface. Load a clean voice reference, enter a short script, generate speech, play the complete result and find the output files. The reference can come from audio or video; use the range controls to choose the useful speech. Smart sentences groups longer scripts, while explicit pause tags and pronunciation dictionary entries help shape the result.

Main topics include speaker reference versus emotion reference, emotion vectors and text-driven expression, speaking rate and timing modes, candidate generation, saved outputs and reusable presets. Create a complete video from an SRT file and background image, with each caption tied to its intended time. Generate separate named audio files from a folder of scripts. For training, start with raw recordings and matching captions, prepare clips, audit the dataset and cache features. Keep entire source recordings separate for training, validation and the final test. Configure DoRA training, follow its loss and speech checks, inspect the saved checkpoint recommendation and compare Base with selected checkpoints using the same text, reference, strength and seed. The voice decoder has its own validation decision. A saved voice profile can carry reference, pace and decoding choices back into Voice Generation. Generate new words with the trained voice, inspect its line-length and pause guidance, and save the speaking rate for that adapter. Models and Performance covers GPU memory tiers, standard and INT8 ConvRot weights, block swapping, auxiliary model residency and the isolated memory benchmark. Model-file verification, unloading, Help, Changelog and the normal update route finish the Windows workflow.

Use the chapters to follow the installation in order or jump to voice cloning, expression, caption video, batches, training, checkpoint comparison and GPU memory controls.

Chapters:

00:00:00 Make your own voice read new words

00:01:17 Choose reference cloning or a trained voice

00:01:51 Find the installer and Windows requirements

00:02:46 Move and extract the installer package

00:03:39 Install and open IndexTTS

00:04:45 Make and save your first voice clone

00:06:08 Build a clean reference from audio or video

00:07:08 Write longer narration and control the pauses

00:09:50 Save pronunciations for technical words

00:10:59 Give the same voice a different delivery

00:13:08 Adjust pace and fit a production slot

00:14:55 Change the useful quality controls

00:16:30 Export and reuse the finished speech

00:17:44 Turn subtitle cues into speech and video

00:19:27 Generate several scripts in one batch

00:20:43 Save and reopen your working setup

00:21:44 Start voice training from raw recordings

00:23:20 Create complete speech clips

00:24:56 Check the data and build the feature cache

00:26:25 Configure a new LoRA or DoRA voice

00:27:48 Save checkpoints and watch the training

00:29:43 Read the training result and automatic comparisons

00:30:55 Compare voices fairly and choose a checkpoint

00:32:39 Generate with the trained voice and its profile

00:34:10 Match the runtime to your GPU memory

00:35:28 Keep the app ready for the next project

00:36:20 Use the finished voice and choose your next project

The tutorial demonstrates v6.22, with automatic model downloads, a prepared environment and prebuilt acceleration libraries. Read the training recommendations, listen to the complete checkpoint comparisons, and choose the voice setup that works for your project.

The package also supports RunPod, SimplePod and Massed Compute. This video teaches the Windows route; the linked post explains the other supported platforms and current downloads. The app link is also in the pinned comment. For installation questions, feedback and feature requests, use the post and video comments.

Video Transcription

• 00:00:06 That voice came from this text and the demo  sample bundled with the app. Give it a short  

• 00:00:12 voice sample, write something new, and turn those  words into speech. This is IndexTTS 2.5 Premium.  

• 00:00:20 It turns written text into speech in the voice  you choose. Our application brings voice cloning,  

• 00:00:26 long narration, subtitle timing and voice training  into 1 workspace. Here is the same workflow making  

• 00:00:33 a narrated video from subtitles. Here are several  scripts turned into separate audio files. You  

• 00:00:39 can build a whole project and keep each result  organized. And you can train your own voice from  

• 00:00:53 recordings, then compare the saved versions side  by side. I will show you how to choose a useful  

• 00:00:59 checkpoint and actually use it for new speech.  We have also prepared the installation, automatic  

• 00:01:04 model downloads and graphics card presets. I will  start with a fresh Windows installation, make  

• 00:01:10 the 1st voice, then take you through the advanced  workflows step by step. You can start with just a  

• 00:01:16 clean sample of your voice. Training is optional.  The reference tells the model who is speaking,  

• 00:01:22 while the text tells it what to say. For a voice  you will use repeatedly, we also have a complete  

• 00:01:28 training route. Your recordings become speech  clips, the app checks them, and a small adapter  

• 00:01:33 learns that voice for later use. Start with the  1st clone and get a useful result. Then add the  

• 00:01:40 features you need: expression, longer scripts,  timed video, batches or your own trained voice.  

• 00:01:46 Let us get the application installed. The link to  this post will be in the description of the video  

• 00:01:52 and the pinned comment. Start here at the top.  Read the post as well, because it contains the  

• 00:01:58 download, explanations and useful examples. Scroll  down to the Windows download. This zip contains  

• 00:02:05 the installer and the start files. Just below  it is the Windows Requirements section, which  

• 00:02:10 you should follow before running the installer.  If you have not completed these requirements,  

• 00:02:15 open the linked tutorial and follow that setup  first. The updated post is linked here too. We  

• 00:02:22 are starting after those requirements, so we can  focus on installing and using this application.  

• 00:02:27 Now return to the zip link and click it.  Chrome saves the file in Downloads. Open  

• 00:02:32 the downloads list and choose Show in folder, so  we can move the package to its final location.  

• 00:02:43 I will put the application on my G drive  in this folder. You can use another drive.  

• 00:02:48 Choose a short path with English characters and  no spaces, because spaces or other characters  

• 00:02:54 can cause installation problems. Paste the  downloaded zip into this folder. Moving it  

• 00:03:00 first keeps the installer and the application  together, and makes it easy to find the start  

• 00:03:05 file next time. Right-click the zip and  choose Extract All. Check the destination,  

• 00:03:11 then extract the complete package here. We need  all of these files together, so keep the whole  

• 00:03:18 extracted folder. These are the Windows files  we will use. The install or update file sets up  

• 00:03:25 the application. The start file opens it after  installation. Run the files by double-clicking,  

• 00:03:31 or select one and press Enter. Never run them  as administrator. Double-click Windows Install  

• 00:03:38 or Update. A command window opens and begins  the installation. Keep that window open while  

• 00:03:44 it works, and let the installer complete each  stage. The installer creates the application's  

• 00:03:51 environment and installs the packages for you.  Our package includes PyTorch 2.13 with CUDA 13,  

• 00:03:58 plus precompiled acceleration libraries. You  do not need to build those libraries yourself.  

• 00:04:05 It also downloads the required models  automatically. You can see the download  

• 00:04:09 and verification progress in this window. Let the  process finish before you start the application.  

• 00:04:16 The installer has now reached its completion  prompt. Press a key to close this window. Back  

• 00:04:23 in the extracted folder, double-click Windows  Start App to open the application. The start  

• 00:04:31 file opens the local address in your browser. Keep  its command window running while you use the app.  

• 00:04:37 Next time, you can start here again without  repeating the installation. On the 1st start,  

• 00:04:43 the app selects a preset for your graphics card's  memory. These presets coordinate the settings  

• 00:04:49 across the application. I will use the detected  preset to make our 1st voice. Start with Reference  

• 00:04:56 Voice on the left. Click the upload area and open  the bundled demo voice from the reference audios  

• 00:05:02 folder. You can use your own clean recording  later. This sample gives the voice its identity.  

• 00:05:09 Now write the words you want it to say in the  text box. I will use English. You can also choose  

• 00:05:14 Chinese, Japanese, Arabic or Spanish to match your  script. Click Generate voice. the 1st generation  

• 00:05:24 loads the models. The progress area shows what the  app is doing, and the finished audio appears here  

• 00:05:30 when the task completes. The generated speech is  ready. Press Play to hear the complete sentence.  

• 00:05:37 Notice that the words match our text while the  speaker comes from the reference we supplied.  

• 00:05:52 Open outputs folder to find the saved files.  Each task keeps its audio and settings together.  

• 00:05:59 Keep the wave file for editing, and keep the  task information when you want to reproduce  

• 00:06:05 that result. You can also load an existing  audio or video file by its full path. This  

• 00:06:12 is convenient when the recording is already on  your computer. I will use the bundled demo voice.  

• 00:06:23 Time ranges let you take just the section you  want. Enter a start and end time, then extract the  

• 00:06:29 range. Choose a complete, clean phrase from the  same speaker, with a natural beginning and ending.  

• 00:06:36 You can join more than 1 range, or clear the  range and choose another section. The reference  

• 00:06:41 library also lets you reopen saved samples. Use a  reference that represents the voice and delivery  

• 00:06:47 you want. There is a microphone recorder here too,  if you want to record a reference directly. For  

• 00:06:54 this walkthrough I am keeping the same saved  sample, so the comparisons all start from the  

• 00:06:59 same speaker. Let us turn a longer script into  narration. I will paste this short story into  

• 00:07:12 the text box. The live preview underneath shows  how the app divides the text into sections before  

• 00:07:18 generation. Smart sentences groups complete  sentences into useful speaking sections. Every  

• 00:07:28 sentence gives each sentence its own section.  Watch the preview change as I switch between them,  

• 00:07:34 then return to Smart sentences for this story.  The token budget limits how much text can fit in a  

• 00:07:41 section. If you use a trained voice, the automatic  option can match the typical line lengths in its  

• 00:07:48 recordings. We will see that voice profile after  training. Sentence pause controls the gap between  

• 00:07:54 separately generated sentences. Maximum pause  controls the long gaps inside generated speech.  

• 00:08:00 Start with the voice's defaults, then change these  when your narration needs a different rhythm. For  

• 00:08:10 an exact break, write a pause tag in the text.  This tag asks for 0.5 seconds of silence at this  

• 00:08:17 point. The preview shows it as a pause, so you  can check its position before generating. Generate  

• 00:08:26 the story, then listen across that break. You can  build chapters this way, with readable paragraphs,  

• 00:08:32 complete speaking sections and deliberate  pauses where the listener needs them.  

• 00:09:44 Some projects use names and technical words that  need special attention. Open Pronunciation check  

• 00:09:50 and dictionary below the preview. The app can flag  unfamiliar words and suggest readings for them.  

• 00:10:00 Here are the suggested readings. Dictionary-backed  suggestions have an existing pronunciation to  

• 00:10:05 build on. Inspect the proposed words, then add the  suggestions you want to keep and save them. The  

• 00:10:12 dictionary is reusable. You can edit an entry and  choose whether it applies everywhere or only when  

• 00:10:18 the voice has not learned that word from training.  Keep the apply-dictionary option enabled to use  

• 00:10:22 your saved entries. Now the preview reflects the  saved pronunciation. Generate the sentence and  

• 00:10:30 listen to those words in context. The same  dictionary also works when you make a batch  

• 00:10:34 of scripts. Voice identity and expression are  separate choices. Open Emotion Control. Same as  

• 00:10:57 speaker voice starts from the speaker reference,  which is a simple place to begin for everyday  

• 00:11:03 narration. Emotion reference audio lets another  sample guide the delivery while the speaker  

• 00:11:22 reference keeps the identity. Load an expressive  recording here, then choose how strongly that  

• 00:11:28 expression should influence the result. Emotion  vector gives you direct controls such as joy,  

• 00:11:49 surprise and calm. I will add a little joy to  this line. Change 1 thing at a time so you can  

• 00:11:54 hear what the control contributes. Emotion text  lets you describe the delivery in ordinary words.  

• 00:12:14 I will ask for a warm, excited announcement. The  app turns that direction into emotion controls  

• 00:12:20 for the voice. These are 4 ways to direct the same  speaker. Keep the script and reference fixed while  

• 00:12:42 comparing them, and use the emotion weight to make  the effect more restrained or more expressive.  

• 00:12:52 Speaking rate controls how quickly the generated  voice delivers the words. Open Segmentation and  

• 00:12:58 Timing. A value of 1 uses the normal pace, and a  slightly higher value asks for a faster delivery.  

• 00:13:19 Listen to the 2 versions with the same text.  Use a small change when you want a livelier  

• 00:13:24 introduction or a more relaxed explanation.  Judge the words and the pauses together.  

• 00:13:44 Target duration is for a whole output that needs  to fit a time slot. Natural mode asks the speech  

• 00:13:50 model to aim for that duration. I will use  a short line and a realistic target here.  

• 00:14:18 Pad can add extra silence to reach a longer  slot. Trim cuts at the target length. Choose  

• 00:14:25 the mode that fits your project, and check the  ending when the timing is strict. Return the  

• 00:14:30 mode to off for ordinary narration. The preset  is a good starting point. In Sampling, a fixed  

• 00:14:39 seed helps you compare settings because you keep  the random starting point consistent. I will use  

• 00:14:45 the same seed for these examples. Candidates  asks for several versions of the same text.  

• 00:14:53 This is useful when you want to choose a delivery  for an important line. Generate a small set,  

• 00:14:58 then compare the actual results. Temperature  controls variation in the choices the model  

• 00:15:27 makes. Beams changes how many candidate paths it  considers while decoding. Start from the preset,  

• 00:15:34 and compare a small change on the same sentence  before using it for a whole chapter. Length  

• 00:15:40 penalty influences how the decoder balances  sequence length. Diffusion steps belongs to  

• 00:15:46 the stage that turns the prediction into audio.  These are useful comparison controls, so keep your  

• 00:15:52 reference, text and seed fixed when testing them.  For day-to-day work, save a configuration that  

• 00:15:58 you like and keep generating. The app also has  automatic retries for incomplete speech, so that  

• 00:16:04 recovery can stay inside the generation workflow.  The Output section controls which files you keep,  

• 00:16:11 and a clear filename makes the result easier  to find. Leave Save MP3 off for a wave file,  

• 00:16:17 or enable it for a smaller MP3 export. Audio  tuning offers ready choices for finishing the  

• 00:16:25 sound. You can keep it on bypass, or compare a  preset such as voice clarity or clear narration.  

• 00:16:30 Use the result in your actual project to decide.  Generate the line with these output settings.  

• 00:16:39 The app will save the finished speech inside its  own numbered task folder. Open the output folder,  

• 00:16:48 then open the numbered task folder for this  generation. Each generation has its own folder,  

• 00:16:55 so related files stay together. Here are the saved  audio files and the reference used for this task.  

• 00:17:05 Clear names make the files much easier to reuse in  a larger project. Recent outputs gives you another  

• 00:17:11 way to find a completed task. Select a result and  load it into Reference Voice when you want to use  

• 00:17:18 that generated clip as a new reference. Now  let us create a narrated video. This subtitle  

• 00:17:34 file contains the words and the time slot for each  line. I have supplied a short example, so you can  

• 00:17:40 follow the same steps with your own captions.  Upload the subtitle file in Voice Generation.  

• 00:17:47 The app reads the text and shows its sections in  the preview. It supports common caption formats,  

• 00:17:53 including subtitle files exported by transcription  tools. Enable Use caption cue timing to keep the  

• 00:18:00 caption start times and fit speech into those  slots. Leave it off when you only want the  

• 00:18:07 words from the captions to flow as ordinary  narration. Add a still image when you want a  

• 00:18:14 video file as well as the audio. I will use this  observatory image. Keep the image large and clear,  

• 00:18:20 because it becomes the picture for the finished  video. Generate the result. Once it finishes,  

• 00:18:27 open the video and play it from the beginning.  The spoken lines follow the subtitle timeline,  

• 00:18:33 and the image fills the video frame. That gives  you a practical route from captions to a narrated  

• 00:19:07 video. For the next example, clear the image and  caption upload, turn cue timing off, and return to  

• 00:19:14 normal text generation. Batch Generation is for  a set of scripts you want to process together.  

• 00:19:20 Here are 2 text files for 2 short chapters.  Each file contains the words for 1 output,  

• 00:19:26 with a name that tells me what it is. You can  upload files, paste separate items, or enter  

• 00:19:31 a folder path. I will use the folder path. The  app can find text and caption files inside it,  

• 00:19:37 including supported files in subfolders. Choose  One reference for all when every item should use  

• 00:19:44 the same voice sample. Per-file reference looks  for a matching audio file beside each script,  

• 00:19:49 which is useful when different items need  different references. Set the output subfolder  

• 00:19:56 and the naming pattern. This pattern keeps an  item number and its name. Reuse loaded model  

• 00:20:02 between items keeps the generation model available  as the batch moves through the scripts. Click  

• 00:20:10 Generate batch. The table tracks each item and its  output. When the batch finishes, open its folder  

• 00:20:17 and check that each script has its own completed  audio file. Once a setup works for your project,  

• 00:20:35 save it as a preset. Enter a new name here and  press Save. User presets keep the registered  

• 00:20:41 settings across the application's tabs, so you  can return to the same workflow. I will change  

• 00:20:47 1 setting here. Watch the speaking rate as I  increase it slightly from the value we saved.  

• 00:20:57 Now select the saved preset and press Load.  This brings back the settings that we stored  

• 00:21:04 for the project. The saved value comes back. Use  clear preset names for different voices or jobs,  

• 00:21:11 so you know which setup you are restoring. Load  last values restores the last working values,  

• 00:21:16 and the theme button changes the appearance.  These are useful conveniences when you return to  

• 00:21:21 a project. Now let us build a trained voice from  the original recordings. For training, begin with  

• 00:21:36 recordings of the voice you want to learn. I am  using my own original videos and their captions.  

• 00:21:42 The source recordings stay untouched while we  build a separate training dataset. Copy the  

• 00:21:47 chosen recordings and their matching caption files  into a new folder. Keep each caption beside its  

• 00:21:54 recording with the matching filename. That lets  the app find the words that belong to each source.  

• 00:22:04 Open LoRA Dataset Preparation. In Input files  or folders, enter the full path to the separate  

• 00:22:11 folder that holds the copied recordings  and their captions. Give the dataset a new,  

• 00:22:20 descriptive name. This name identifies the  prepared clips, and we will use a different  

• 00:22:26 name for the audited training dataset. The  language is English for these recordings.  

• 00:22:32 Use 1 speaker name for the same person, so the  prepared clips carry a consistent speaker label.  

• 00:22:40 Press Scan inputs before preparing the dataset.  The table shows the discovered files, durations  

• 00:22:46 and sidecar captions. Check that these are the  sources you intended to use before starting the  

• 00:22:52 longer process. Good training starts with clear  speech from the same person and accurate words.  

• 00:22:58 Captions are especially useful here because  they preserve your names and technical terms.  

• 00:23:03 The next controls turn these long sources into  complete speaking clips. In Transcripts, Prefer  

• 00:23:11 sidecars uses the captions beside the recordings.  Whisper helps with the alignment when needed.  

• 00:23:17 Sentence-aligned segmentation places the words in  the audio so the app can cut at useful boundaries.  

• 00:23:24 Keep the caption cleanup options enabled for  these sources. They remove bracket notes and  

• 00:23:29 rolling-caption duplicates. That helps the  training text describe the speech instead of  

• 00:23:34 caption formatting. In Segmentation, the target is  14 seconds, with a range from 4 to 16 seconds. The  

• 00:23:41 app aims for complete sentences and also includes  shorter clips, so the voice learns useful lengths  

• 00:23:47 for different prompts. Silence snapping and edge  padding help keep the words intact at the cuts. In  

• 00:23:54 Cleanup and Loudness, leave the voice-training  defaults in place. They provide a consistent  

• 00:23:59 sample rate and screen the clips before training.  Leave Overwrite dataset off for this new dataset.  

• 00:24:07 Then press Prepare dataset and follow the  progress. The app extracts the audio, aligns  

• 00:24:12 the words and saves the accepted speech clips  with their text. Here is the prepared dataset.  

• 00:24:20 The summary reports what was accepted, and the  table shows each clip with its duration and  

• 00:24:26 transcript. Select a row to inspect the waveform  and read the words that will train the voice.  

• 00:24:46 Voice and transcript audit adds another check  before training. It compares the clips with a  

• 00:24:52 verified speaker reference and checks their  text. Give the audited dataset a new name,  

• 00:24:56 so the prepared source remains available.  

• 00:25:05 I will reserve 1 complete recording for  validation and another for final testing.  

• 00:25:10 Their filename stems go here. Training learns from  the other recordings, while these separate sources  

• 00:25:17 help us check how the voice handles unseen  speech. Keep the transcript and boundary-word  

• 00:25:23 checks enabled. The second-opinion option gives  rejected transcripts another recognition pass.  

• 00:25:29 Click Audit and create training dataset, then  let the checks finish. The audited dataset is  

• 00:25:36 now selected. Its report keeps the accepted clips  and the reasons for exclusions. Review that report  

• 00:25:43 and representative clips before you commit to a  training run. Now press Cache features now. This  

• 00:25:49 runs the expensive base-model preparation once  and saves the features that training will reuse.  

• 00:25:56 Wait for the cache to complete before opening the  training tab. The cached status is visible here,  

• 00:26:03 and the dataset carries over to Training.  We have gone from original recordings to an  

• 00:26:08 audited dataset that the trainer can use directly.  This is the training workspace. Check the dataset  

• 00:26:18 and graphics card preset first. Enter a new  adapter name so the new voice has its own folder  

• 00:26:25 and its own saved history. LoRA and DoRA train  small adapters that add a learned voice to the  

• 00:26:33 existing speech model. I will use the default  DoRA setup here. Rank and alpha start at 128,  

• 00:26:40 with the recommended training modules enabled.  Training data fluency shows how much material  

• 00:26:47 remains under different speech filters. Press  Analyze dataset. The table helps you choose  

• 00:26:52 between all curated clips and stricter limits on  hesitations. Look at the amount of speech each  

• 00:26:59 filter keeps before choosing. The selected filter  affects the training clips, while validation stays  

• 00:27:04 the same. That makes different runs easier to  compare fairly. In Optimization, epochs is the  

• 00:27:12 maximum number of passes through the dataset.  Learning rate controls the size of the training  

• 00:27:18 updates. I will begin with the app's quality  defaults for this new run. Batch size is 1, with  

• 00:27:25 1 gradient accumulation step. Keep these defaults  for this example. The graphics card preset also  

• 00:27:31 prepares the precision and memory settings, so  you have a coordinated starting point. Validation  

• 00:27:39 measures clips that are separate from training.  We already reserved complete source recordings  

• 00:27:44 during the audit. The trainer can use another  clip of the same speaker as the reference, which  

• 00:27:48 matches how we will generate new speech. Use the  separate final test dataset here. The app checks  

• 00:27:55 it only after the voice and generation settings  are chosen, so it stays separate from training  

• 00:28:01 and checkpoint selection. Automatically stop when  progress stalls lets the app stop a run that has  

• 00:28:08 stopped improving. The epoch probe also generates  held-out speech, so the stopping decision can use  

• 00:28:15 more than the training loss alone. Open Saving and  Resume. Start fresh begins this new voice from the  

• 00:28:22 base model. Save best keeps a useful checkpoint as  training progresses, and saving each epoch gives  

• 00:28:28 us versions to compare later. Training samples let  us inspect the voice during the run. Keep a short  

• 00:28:35 sample sentence here. Automatic speech comparison  will later compare generated speech using held-out  

• 00:28:41 prompts and consistent settings. Now click  Start training. The live dashboard shows steps,  

• 00:28:49 epochs and losses as they arrive. Keep the process  running and watch the saved checkpoint and sample  

• 00:28:55 information as the run progresses. Training loss  shows how the model fits the training material.  

• 00:29:02 Validation shows how it handles the held-out  material. The app also keeps generated samples,  

• 00:29:08 so we can connect the numbers with actual speech.  The Stop button requests a controlled stop if you  

• 00:29:22 need to end a run. For this example I will let the  configured training and automatic checks finish,  

• 00:29:28 then inspect the saved results together. The  run has finished its training and follow-up  

• 00:29:35 checks. Open the output folder and keep these  results together. The adapter, checkpoints,  

• 00:29:41 training state and analysis describe exactly  what this run produced. The automatic comparison  

• 00:29:49 includes the base voice and candidates from  this run. It checks generated words and  

• 00:29:54 speaker similarity on held-out speech. Use its  recommendation to decide which versions deserve  

• 00:30:00 a closer comparison. Voice decoder adaptation  is another optional stage. It adapts the part  

• 00:30:06 that turns the voice representation into sound.  The app checks that result before automatically  

• 00:30:12 using the decoder with the voice. The saved  analysis also records the selected reference,  

• 00:30:19 speaking pace and decoding settings. Those choices  help the generation screen reopen a practical  

• 00:30:26 setup for the trained voice. If you return to  training later, Weights only starts a fresh  

• 00:30:32 training schedule from saved weights. Continue run  restores the saved training state. Choose the mode  

• 00:30:38 that matches whether you want a new experiment  or to continue the same run. Now let us answer  

• 00:30:46 the question from the opening: which saved voice  should you use? Click Compare in grid to open  

• 00:30:52 this run in Checkpoint Grid. The checkpoint  table keeps the saved versions together.  

• 00:31:01 The analysis and evaluation help you narrow the  choices. For the listening grid, compare the base  

• 00:31:07 model and a small set of saved checkpoints using  the same text and the same reference. Enter 1 text  

• 00:31:15 per line and 1 reference path per line. Strength  controls how much of the adapter is applied.  

• 00:31:20 I will keep 1 reference, 1 short sentence  and the same strength for this comparison.  

• 00:31:28 Keep the sampling settings consistent too,  including the seed. Then click Generate grid. The  

• 00:31:33 app renders each selected combination and records  which checkpoint and settings created every  

• 00:31:39 result. Here are the completed versions. Listen  to the same sentence across the rows. Compare the  

• 00:31:45 words, speaker identity, pauses and delivery, then  choose the checkpoint that works for your use.  

• 00:32:14 Use best checkpoint sends the recommendation  back to Voice Generation. You can also choose  

• 00:32:20 a different saved checkpoint there  after your comparison. Keep the grid  

• 00:32:25 folder when you want to revisit that  decision. Back in Voice Generation,  

• 00:32:30 the adapter selector shows the saved voice. The  recommended reference can load automatically,  

• 00:32:35 and the voice decoder has its own selector when a  compatible decoder is available. Strength controls  

• 00:32:42 the adapter's influence. Start at 1 for this  example. The automatic options can apply the  

• 00:32:49 saved speaking rate and decoding settings, so  you do not need to re-enter the voice setup each  

• 00:32:55 time. The voice profile also shows typical line  lengths and pauses from the training recordings.  

• 00:33:02 Smart sentences can use that information to group  your script into sections suited to this voice. An  

• 00:33:10 expressive clip can guide the emotion while  the recommended reference keeps the speaker  

• 00:33:16 identity. You can compare that option  with the ordinary speaker reference,  

• 00:33:20 using the same words and emotion weight. Write  new text and generate it with the trained voice.  

• 00:33:26 This is the complete route: original recordings,  prepared clips, a checked dataset, training,  

• 00:33:32 comparison and a new piece of speech. If the  voice needs a different pace, adjust the rate  

• 00:33:49 and save it for this adapter. Checkpoint Grid  also offers speaking-rate calibration from  

• 00:33:54 its generated results. Use representative speech  when making that decision. Models and Performance  

• 00:34:01 shows the detected graphics card and the runtime  controls. The memory tier gives you a coordinated  

• 00:34:06 configuration for the card, including where the  different model parts will live. The preset starts  

• 00:34:14 with the standard model weights. INT8 ConvRot  reduces the memory used by the language-model  

• 00:34:20 weights. Choose the variant that fits your setup  and apply the runtime settings when you change  

• 00:34:25 them. Block swapping moves some model blocks  between system memory and graphics memory as  

• 00:34:36 needed. It gives smaller cards another way to fit  the workload. The preset prepares these choices,  

• 00:34:42 so start from the tier that matches your  card. Auxiliary Model Residency controls  

• 00:34:48 the other model parts. GPU keeps a part ready in  graphics memory. On demand moves it in when it  

• 00:34:55 is needed. These settings help balance repeated  generation and memory use. The isolated memory  

• 00:35:03 benchmark checks a runtime configuration without  replacing your working generation session. Run it  

• 00:35:08 for the selected setup, then read the reported  memory use and fit result for this machine.  

• 00:35:17 Model Files and Downloads shows the files the  application needs. Use verification when you want  

• 00:35:23 to check the installation, and use the download  or repair controls when model files need to be  

• 00:35:29 fetched again. Unload model frees the loaded  generation models when you are finished. The  

• 00:35:35 next generation loads what it needs again. This  is useful when you want to use the graphics card  

• 00:35:41 for another application. The Help tab keeps  the main workflows inside the application,  

• 00:35:46 and Changelog explains what has  been added in each release. Use  

• 00:35:51 these when you return to a feature you have  not used recently. To open the app next time,  

• 00:35:57 use Windows Start App again. When you choose to  update, the install or update file is in the same  

• 00:36:02 extracted package. Keep your generated work and  saved voice files organized before maintaining  

• 00:36:08 the installation. You now have the complete  Windows workflow: install the application,  

• 00:36:13 clone a voice, control its delivery, generate  timed video and batches, and train a voice from  

• 00:36:19 your own recordings. The files are ready for your  next project. The package also supports RunPod,  

• 00:36:26 SimplePod and Massed Compute. Those options  are useful if you want cloud graphics cards  

• 00:36:30 or more capacity. This video covered the  Windows route, and the post explains the  

• 00:36:36 other supported platforms. You will find  the post in the description of the video  

• 00:36:42 and the pinned comment. Read its examples,  download the application, and start with 1  

• 00:36:48 clean voice sample and 1 short script. Then  build the workflow that fits your project.