Best Local TTS Voice Cloning in 2026 With IndexTTS 2.5, OmniVoice and AuK
By Furkan Gözükara · Updated Nov 21, 2026
The best local TTS voice cloning in 2026 comes from three models that run on your own GPU: IndexTTS 2.5 for emotion control and long narration, OmniVoice for multilingual cloning and the strongest trained voices, and AuK for designing a voice from one sentence and editing recordings. SECourses packs all three into one app, Ultimate TTS, with GPU presets from 6 GB to 32 GB. To turn an existing recording or your live microphone into another voice, RVC is the separate tool.
Which model for which job
| Model | Clones from a clip | New voice without a recording | Languages | Standout feature |
|---|---|---|---|---|
| IndexTTS 2.5 | Yes | — | Five | Emotion control separate from the voice |
| OmniVoice | Yes | Yes, from voice tags | Many; full fine-tuning in 600+ languages | Best trained voices with DoRA |
| AuK | Yes | Yes, from one plain sentence | English and Chinese | 23 audio editing tasks |
| RVC | Converts one voice into another | — | Speech and song vocals | Real-time microphone voice changing |
You can train your own voice for IndexTTS 2.5, OmniVoice and AuK inside the same app, and each model keeps its own settings when you switch between them in the header.
IndexTTS 2.5: emotion and long narration
IndexTTS 2.5 separates the speaker's identity from the emotion. You can keep the speaker's own tone, copy the delivery of an emotion clip, mix eight emotion sliders (joy, anger, sadness, fear, disgust, depression, surprise and calm) or describe the emotion in words, and an emotion weight sets how strong it is.
It is also fast: on an RTX 5090 it made 31 seconds of speech in 18 seconds from a 14.6-second reference cut from a video. The IndexTTS 2.5 Premium guide documents the quality preset: temperature 0.8, top-p 0.8, four beams, CFM temperature 0.9 and 40 diffusion steps. The system default uses three beams and 25 steps, and 12 to 16 steps is the fast draft range. For audiobooks, the guide notes that Prevent VRAM accumulation, which clears caches between sections, is slower but useful for long books.
OmniVoice: languages and trained voices
OmniVoice clones a reference voice, designs a new voice from tags, or picks a voice by itself. A voice built from four tags (female, young adult, moderate pitch and British accent) spoke 6.1 seconds of speech in 1.2 seconds, with no reference recording at all. The Ultimate TTS tutorial plays Arabic, Chinese, French, German, Italian, Japanese, Korean, Polish and Spanish examples and shows cross-language cloning.
OmniVoice is also where training shines. SECourses reports that OmniVoice DoRA trainings (DoRA is an improved form of LoRA), combined with the app's take selection, give results better than ElevenLabs in its tests.
AuK: describe a voice, edit a recording
AuK made 6.6 seconds of speech in 1.5 seconds from a single description: "A calm middle-aged man with a deep, warm voice, speaking slowly and clearly, like a documentary narrator." Its audio editing tab has 23 tasks: it replaces, inserts or removes words, changes speed, pitch, volume or emotion, cleans up noise and reverberation, and keeps one speaker or the voices from music. Replacing two words in a 9.9-second recording took 1.7 seconds.
RVC: voice conversion instead of text to speech
RVC does not read text. It turns one voice into another, for speech or song vocals, and it can change your microphone voice in real time. The RVC WebUI Premium app uses under 4 GB of VRAM, downloads more than 30 demo voices, separates vocals from music for full-song conversion, merges several voices into a custom one and trains new voices. It is available to members on Patreon, and the public RVC WebUI guide walks through installation, conversion, batch inference and training. The tutorial is meant for educational and responsible use.
Hardware requirements
Ultimate TTS ships seven read-only GPU presets from 6 GB to 32 GB and selects the one that matches your card on the first start. A smaller card pays with speed first: helper models move to the CPU, and the 6 GB preset streams 22 of the 24 GPT blocks from system RAM.
| Task | VRAM from the sources |
|---|---|
| Voice generation | 6 GB preset and up, with BF16 or INT8 ConvRot models and block swap |
| LoRA or DoRA voice training | An IndexTTS 2.5 voice trained with 2.3 GB of VRAM |
| Full fine-tuning | 16 GB cards and larger (32 GB for AuK) |
| RVC conversion and training | Under 4 GB |
The Models & Performance tab has a fit check that shows the expected peak memory before anything loads, plus a built-in VRAM benchmark.
How to get a clean clone
- Use a clean, single-speaker reference with little music, reverb or room noise. A few seconds is enough, and 8 to 15 seconds is a strong starting point.
- Drop in audio or video, or record from your microphone. Time ranges such as 0.2:14.8 keep only the clean part of a long recording.
- Set the language to match the script, not the reference speaker.
- Leave Take quality on. The app renders several takes of every section, the built-in Whisper checks every word, and the take that sounds most like the voice without a word error is kept. In a 60-line test, word errors fell from 0.65% to 0.20% with OmniVoice and from 0.91% to 0.35% with IndexTTS 2.5.
- Insert exact silences with pause tags such as [pause:800ms], and save the readings of names and technical words in the pronunciation dictionary.
For longer work, Smart sentences packs whole sentences into each section, and subtitle files in 10 formats, including SRT, VTT and ASS, can drive speech that starts on every cue. Add an image and the same run also makes an MP4. In batch mode, three stories became 116 seconds of speech in 65 seconds with OmniVoice, with every section picked from 10 takes.
Training your own voice
A trained voice follows the speaker's pace, pauses and pronunciation much more closely than a single reference clip. In the documented run, four recordings totaling 46.8 minutes became 205 clean clips of 4 to 16 seconds in 50 seconds. The voice and transcript audit checked them in 1 minute 14 seconds and kept 194. After training, likeness to the speaker's real recordings rose from 0.810 with the base model to 0.867 for the chosen checkpoint, and the app saved a ready-to-use preset for the new voice.
Install on Windows or in the cloud
Ultimate TTS is available to members on Patreon.
- Install Python 3.12, Git, FFmpeg, CUDA 13, cuDNN 9.17 and Visual Studio with the C++ tools, following the Windows requirements video or the requirements installer guide.
- Extract the zip into a short folder path without spaces and run Windows_Install_or_Update.bat. It builds its own Python 3.12 environment with PyTorch 2.14.1 and CUDA 13 using uv; a fresh install took 5 minutes 39 seconds, models included.
- Start the app with Windows_Start_App.bat. It was ready in 9.5 seconds on the first start.
Cloud installers cover RunPod, SimplePod, Massed Compute and Linux, and a Gradio share link lets you use the app from any device. See RunPod vs Massed Compute vs SimplePod to pick a platform. Older SECourses installers for Tortoise TTS and Coqui TTS remain only as legacy posts.
FAQ
Is local voice cloning as good as ElevenLabs?
SECourses reports that OmniVoice DoRA voices trained with the app's pipeline and take selection beat ElevenLabs in its tests. The Ultimate TTS tutorial video shows a voice trained on a PC that users rate above ElevenLabs.
How long does the reference clip need to be?
A clean clip of a few seconds is enough to clone a voice, and the IndexTTS guide suggests 8 to 15 seconds as a strong starting point. For the closest match to a speaker, train a LoRA or DoRA voice from longer recordings with matching subtitles.
Which local TTS model supports the most languages?
OmniVoice. The Ultimate TTS tutorial covers OmniVoice full fine-tuning in 600+ languages, while IndexTTS 2.5 reads five languages and AuK speaks English and Chinese.
Can I clone a voice on a 6 GB GPU?
Yes. The 6 GB preset streams most GPT blocks from system RAM, so it is slower, but the seven presets fit generation, dataset preparation and training to your card.