GPT-SoVITS: Clone Any Voice From 1 Minute of Audio, Locally

Short answer: GPT-SoVITS clones a voice from a 5-second sample instantly, or from ~1 minute of audio with fine-tuning — free, local, MIT-licensed, and at 60,487 GitHub stars it is the most-starred open text-to-speech project in existence. Use the v2ProPlus model, launch the UI in English with python webui.py en_US, and expect it to run on any NVIDIA GPU with 4GB+ of VRAM (the code falls back to CPU below that).

The catch — and the reason most English-speaking searchers bounce off it — is that GPT-SoVITS is a Chinese-first project. The official tutorials are in Chinese, the WebUI defaults to Chinese on many systems, and the version naming (v2, v3, v4, v2Pro, v2ProPlus — in that chronological order, with v2Pro newer than v4) makes no sense until someone explains it. This page is that explanation: the install commands verified against the current repo, the real VRAM numbers from the official changelog, and every UI quirk you will hit as an English user.

---

What GPT-SoVITS Is (and Why 60K Stars) {#what-it-is}

GPT-SoVITS is a local voice-cloning and TTS toolchain by RVC-Boss — the developer behind the RVC voice-conversion project — with 60,487 stars and 6,593 forks as of August 2026 — ~15K more than Coqui TTS (45,857) and 4x F5-TTS (15,070), per the GitHub API. It launched in January 2024 and the main branch was last pushed July 22, 2026 — this is an actively maintained project, not an abandoned viral repo.

Three capabilities explain the stars (all from the official README):

• Zero-shot TTS: "Input a 5-second vocal sample and experience instant text-to-speech conversion." No training. Upload a clip, type text, get that voice back.
• Few-shot fine-tuning: "Fine-tune the model with just 1 minute of training data for improved voice similarity and realism." This is where GPT-SoVITS pulls away from most rivals — a fine-tuned voice is noticeably closer to the source than any zero-shot result.
• Cross-lingual synthesis: English, Chinese, Japanese, Korean, and Cantonese, including generating in a language the reference speaker never spoke.

It is also not just a model — it is a full pipeline. The WebUI bundles vocal/accompaniment separation (UVR5), automatic dataset slicing, speech recognition for transcript generation (Faster-Whisper or the newer multilingual Fun-ASR-Nano for English, classic FunASR for Chinese/Cantonese — see our Faster-Whisper guide), a transcript proofreading tool, and the training UI. Everything you need to go from "a folder of podcast MP3s" to "a fine-tuned voice" ships in one repo.

---

Which Version to Use: v2ProPlus {#which-version}

Use v2ProPlus. The README states the v2Pro series has "slightly higher VRAM usage than v2, surpassing v4's performance, with v2's hardware cost and speed" — better output than v4 at a fraction of the training cost. The version history only makes sense once you know v3 was a misstep that v4 fixed, and the v2 lineage then leapfrogged both:

Source: RVC-Boss/GPT-SoVITS README, official changelog, and GitHub releases, checked August 2026.

The strongest signal of where the project stands: the current WebUI's version selector offers v1, v2, v4, v2Pro, and v2ProPlus — v3 has been removed entirely. The README's own guidance is that v1/v2/v2Pro handle average-quality training audio better, while v3/v4 chase maximum timbre similarity and want cleaner source audio. For most real-world recordings — podcasts, meeting audio, home mics — v2ProPlus is the right default; try v4 only if your source is studio-clean and 48kHz output matters to you.

---

Hardware: The Real VRAM Numbers {#hardware}

Inference runs on any NVIDIA GPU with 4GB+ of VRAM — that threshold is in the code, not marketing: config.py routes GPUs under 4GB (or compute capability below 5.3) to CPU mode. Fine-tuning's only official numbers are for v3: 14GB full, 12GB with gradient checkpointing, 8GB with LoRA. The v2Pro series is documented as needing only "slightly" more than v2, which the community runs on much smaller cards — but the repo publishes no hard v2-line minimum, so treat 8-12GB as the comfortable fine-tuning zone rather than a spec.

Speed is the part nobody expects. The README's official benchmark for v2ProPlus inference, measured as RTF (real-time factor — generation time divided by audio length, lower is better):

Source: RVC-Boss/GPT-SoVITS README performance figures.

That RTX 4060 Ti number is the story: a mid-range card generates an hour of audio in under two minutes. Notable platform quirks from the code: GTX 16-series cards are forced to fp32 (slower, works fine), AMD cards are supported on Linux via the ROCm install path, and Apple silicon plus pure CPU are officially supported. Supported stack: Python 3.9-3.11, PyTorch 2.2.2+, CUDA 12.4/12.6/12.8.

If you are choosing a GPU partly for this, a 12GB card comfortably covers every published fine-tuning path (LoRA and gradient-checkpointing included) — see what else 12GB unlocks in our best models for 12GB VRAM picks, or start from the GPU buying hub if you are building fresh.

---

Install on Windows (10 Minutes, No Python Required) {#install-windows}

Windows users should skip the manual install entirely: the project ships a prebuilt integrated package — download the 7z from the Releases page, extract, double-click go-webui.bat, done. The current package is the 20250606v2pro release, which includes the pretrained models, and there is a separate build for RTX 50-series cards — pick that one if you are on Blackwell.

1. Go to the GPT-SoVITS releases page and download the latest 7z under the newest release (standard NVIDIA build, or the 50-series build for RTX 50x0 cards).
2. Extract with 7-Zip to a path without spaces or non-ASCII characters — Gradio toolchains are notoriously fragile about paths.
3. Double-click go-webui.bat. The main WebUI opens on http://localhost:9874.

If you prefer a manual install (conda, matching the official docs):

conda create -n GPTSoVits python=3.10
conda activate GPTSoVits
pwsh -F install.ps1 --Device CU128 --Source HF

--Device accepts CU126, CU128, or CPU; --Source HF pulls the pretrained models from Hugging Face (the script also accepts HF-Mirror and ModelScope — those mirrors exist for users in China; you want HF).

---

Install on Linux / macOS {#install-linux}

Linux and Mac get one install script that handles dependencies and downloads all pretrained models:

git clone https://github.com/RVC-Boss/GPT-SoVITS.git
cd GPT-SoVITS
conda create -n GPTSoVits python=3.10
conda activate GPTSoVits
bash install.sh --device CU126 --source HF

Device options on Linux: CU126, CU128, ROCM (AMD), or CPU — Macs use CPU, and Apple silicon is officially supported. Then launch:

python webui.py en_US

That en_US argument matters — see the next section. A Dockerfile and compose setup also ship in the repo if you prefer containers. The pretrained model weights (including s2Gv2ProPlus.pth and friends) land in GPT_SoVITS/pretrained_models/ — if the WebUI complains about missing models, that folder is the first thing to check.

---

The English-User Survival Guide {#english-ui}

GPT-SoVITS is fully usable in English — but only if you know four things the Chinese-first docs never spell out: the en_US launch flag, the four-port architecture, which ASR model to pick, and where the real documentation lives.

1. Force the English UI. The WebUI picks its language from your system locale ("Auto") — on many machines that means Chinese. The last command-line argument overrides it, and the repo ships 13 locales:

python webui.py en_US

2. Learn the port map. GPT-SoVITS is not one app — the main WebUI spawns sub-tools, each on its own port. When you tick a checkbox and a new tab opens somewhere else, this is why:

3. Pick the right ASR for English — and know that the choice depends on your install path. The dataset pipeline transcribes your audio automatically. On the 20250606v2pro Windows package the dropdown offers classic FunASR (Chinese/Cantonese) and Faster-Whisper — pick Faster-Whisper for English. On a current manual install from main, the repo now ships four backends, and its own recommended default for English (and Chinese, Japanese, Korean) is the newer Fun-ASR-Nano (31 languages); Faster-Whisper remains available and works fine for English too. Just don't feed English audio to the classic Chinese-only FunASR backend. If you want to understand what the Whisper engine is doing under the hood, our Faster-Whisper guide covers it.

4. Know where the docs are. The tutorials linked from the releases are Chinese-language documents. The English README covers installation well but goes thin on workflow. Between this page and the README you have the full path; expect Chinese screenshots anywhere else.

---

Zero-Shot Cloning: A Voice in 5 Seconds {#zero-shot}

The fastest result: open the TTS inference UI (port 9872), upload a 3-10 second reference clip, paste its transcript, type what you want said, and generate. The reference must be 3-10 seconds — the UI errors above 10, by design.

The workflow, with the settings that actually matter:

1. In the main WebUI, open the inference section and launch the TTS Inference WebUI (it opens on port 9872). In its model dropdowns, pick the v2ProPlus pretrained weights (s2Gv2ProPlus.pth) — the inference UI reads the version from the weights you load, so this dropdown is the version choice.
2. Upload reference audio, 3-10 seconds. This is a hard limit, not a suggestion. Clean speech, no music, one speaker.
3. Type the reference transcript — what is said in the clip, in its original language. This materially improves similarity. There is a "no reference text" checkbox for true zero-shot, but providing the transcript wins.
4. Enter your target text and language (v2-line models support Chinese, English, Japanese, Korean, Cantonese, and mixed modes).
5. Generate. Defaults are sane; the knobs are top_k (default 15), temperature (default 1.0), a speed slider (0.6-1.65x), and sentence-pause length (0.1-0.5s).

Zero-shot quality is genuinely good for prosody and accent, and it captures tone better than most rivals at this speed. But it flattens the speaker's identity somewhat — a zero-shot clone sounds like the person; a fine-tuned clone sounds like them. Which brings us to the feature that earns the GitHub stars.

---

Fine-Tuning on 1 Minute of Audio {#fine-tuning}

The official pipeline is five stages, all in the main WebUI (port 9874): slice → denoise → transcribe (ASR) → proofread → fine-tune. One minute of clean speech is the documented minimum; a few minutes is the practical sweet spot.

What each stage does, in order, on the main WebUI's dataset tab:

1. (Optional) UVR5 separation — if your source has background music, strip it first (port 9873).
2. Slice — chops long recordings into training-sized clips automatically.
3. Denoise — optional cleanup pass; skip for already-clean audio.
4. ASR — auto-transcribes every slice. For English audio, select Faster-Whisper (Windows package) or the recommended Fun-ASR-Nano (current manual installs) (see above).
5. Proofread — the labeling tool (port 9871) shows each clip with its transcript for correction. Ten minutes here is the highest-leverage quality work in the whole pipeline: bad transcripts teach the model bad mappings.
6. Fine-tune — the training tab trains both stages (the SoVITS model and the GPT module) with sensible defaults. On the v2 line this is a short job, minutes-scale on a mid-range card, not an overnight run.

Then load your fine-tuned model in the inference UI and A/B it against zero-shot with the same reference clip. The gap is obvious — pacing, breath sounds, and the small vocal habits that make a voice recognizable all snap into place. This zero-shot vs fine-tuned gap is exactly the trade we map across engines in best local TTS models.

---

Honest Limitations {#limitations}

GPT-SoVITS is the best open fine-tuning cloner, not the best everything. Five real limitations before you commit:

• Five languages, full stop. English, Chinese, Japanese, Korean, Cantonese. No Spanish, French, German, Hindi, or Portuguese — Coqui's XTTS v2 covers 17 languages per its docs, and if you need any of them, read our XTTS v2 cloning guide instead.
• Chinese-first everything. The UI translates cleanly with en_US, but official tutorials, release notes, and most community troubleshooting are in Chinese. Budget extra time for edge-case debugging.
• It is a toolchain, not an app. Four ports, a multi-stage pipeline, and a Gradio UI with a lot of surface area. F5-TTS or Kokoro get you to first audio with far less ceremony — at the cost of no comparable fine-tuning path.
• Release cadence has slowed. The last packaged release is 20250606v2pro (June 2025), while the main branch keeps moving (last push July 22, 2026). The Windows 7z lags the git code; manual installs get fixes first.
• The legal weight is on you. MIT covers the code, not someone's voice. Cloning a real person without consent can breach personality-rights and biometric-privacy laws regardless of license. Use your own voice or licensed/contracted voices.

---

GPT-SoVITS vs XTTS vs F5-TTS {#vs-alternatives}

Pick GPT-SoVITS for maximum-fidelity cloning of one voice (fine-tuning is the moat), XTTS v2 for language coverage, F5-TTS for the lightest credible zero-shot setup. The thirty-second version:

• vs XTTS v2: XTTS wins on languages (17 vs 5) and ecosystem maturity; GPT-SoVITS wins on cloned-voice fidelity once fine-tuned, on inference speed (RTF 0.028 on a 4060 Ti), and on license — GPT-SoVITS is plain MIT while XTTS weights carry the Coqui Public Model License's non-commercial restriction, a distinction our XTTS commercial-license breakdown covers in detail.
• vs F5-TTS: F5 is simpler and zero-shot-only; GPT-SoVITS matches it zero-shot and then beats it with a fine-tune. Star counts tell the adoption story: 60,487 vs 15,070.
• vs the fast-TTS crowd: if you need a good generic voice rather than a specific one, a lightweight engine is less work — the Kokoro vs XTTS vs Chatterbox comparison maps that territory.

---

Verdict {#verdict}

GPT-SoVITS earns its 60K stars: nothing open-source clones a specific voice better on consumer hardware, and nothing else ships the entire dataset-to-fine-tune pipeline in one repo. The honest trade is friction — a Chinese-first toolchain with a four-port UI — against fidelity no zero-shot rival matches.

The playbook this page boils down to:

1. Windows: grab the 20250606v2pro 7z from Releases, extract, go-webui.bat. Linux/Mac: conda + install.sh, then python webui.py en_US.
2. Model: v2ProPlus. Ignore the confusing version numbers — the README itself says it beats v4 at v2's cost.
3. Hardware: any 4GB+ NVIDIA card infers; 8-12GB fine-tunes via the published LoRA/checkpointing paths. A 4060 Ti generates ~36x faster than realtime.
4. Start zero-shot (3-10s reference + transcript) to validate the voice, then fine-tune on a minute+ of clean audio for the result you keep.

Free, MIT-licensed, running on hardware you already own — this is the local-first argument in one tool. If it clicks for you, the rest of the local voice stack is mapped in our local TTS model rankings.

---

Sources {#sources}

• RVC-Boss/GPT-SoVITS — README (features, version notes, RTF benchmarks, install commands), official changelog (v3 VRAM figures), releases page (package versions/dates), and repo code (config.py GPU threshold, webui.py locale handling, ASR backends) — all checked August 2026
• GitHub API — star/fork counts for GPT-SoVITS, Coqui TTS, and F5-TTS, retrieved August 2026
• Coqui TTS docs — XTTS v2 language count and model license terms

---

FAQ {#faq}