How to clone your voice for free with AI in 2026: complete tutorial (GPT-SoVITS, OpenVoice, Fish Speech)

August 15, 202620 minAI, voice, TTS, voice cloning, GPT-SoVITS, OpenVoice, Fish Speech, self-hosting

Ethical disclaimer (read this first): this tutorial is for cloning your own voice or the voice of someone who gave you written, signed consent. Cloning someone else's voice without permission is illegal in most jurisdictions and ethically unacceptable. The ELVIS Act in Tennessee (2024), the EU AI Act, and the FTC guidance all penalize impersonation and non-consensual distribution of AI-generated voices. A whole section of this post is dedicated to that.

Quick answer (60 seconds): cloning your voice for free in 2026 is viable with three open-source tools that run locally: GPT-SoVITS (best quality with fine-tune, needs an NVIDIA GPU ≥6 GB VRAM), OpenVoice V2 (instant cloning with 10-30 s of sample, MIT license), and Fish Speech 1.5/S2 (~100 ms latency, Apache 2.0, commercial-friendly). All free, all local. With 5-30 minutes of clean audio and a little patience, the quality is comparable to ElevenLabs Pro. Do not use these tools to clone someone else's voice without their written and signed consent: that is illegal and ethically unacceptable.

Voice cloning with AI stopped being science fiction in 2023-2024 and became open-source commodity in 2025-2026. You no longer need to pay ElevenLabs (USD 5-22/month), PlayHT (USD 31-99/month), or LMNT (USD 8-30/month) to have your own voice clone: with your laptop, a decent microphone, and one of the three tools in this post, you can train and run your model locally without sending a single byte to the cloud.

This post is the tutorial I wish I had read when I started: what works, what doesn't, how much audio you need, what hardware, what ethics to respect, and what regulation can come down on you if you cross the line. If you just want to try it fast, skip to OpenVoice V2. If you want max quality and have a GPU, start with GPT-SoVITS. If you want a voice agent in production today, Fish Speech is your option.

Disclosure: I ran the tests on an M3 Max 128 GB and an RTX 4090. The VRAM, latency, and inference duration numbers are mine or come from issues / official README files on GitHub (RVC-Boss/GPT-SoVITS, myshell-ai/OpenVoice, fishaudio/fish-speech). Where I cite a benchmark without an exact source, I flag it in the text.

---

Why clone your voice (valid use cases)

Before getting into the tutorial, let's enumerate what legitimately cloning your own voice is good for:

• Audiobooks you wrote. If you're an author and want to narrate your book without spending 40 hours at the microphone, train your voice with 30 min of sample and generate the audiobook in an afternoon.
• Podcasting and video. Multilingual content production: clone your voice and use it to narrate in languages you don't speak, keeping your timbre.
• Personal branding / AI agent. For a SaaS, a voice agent with your personal timbre sounds more professional than a generic TTS voice.
• Accessibility. People who lost their voice to a disease (ALS, laryngectomy) can clone their voice from old recordings and "speak" with their family again. This is the most humane use case of the technology.
• Dubbing and localization. For indie game devs, documentaries and educational content, cloning an actor's voice with their permission costs USD 5K-50K; cloning it yourself reduces the cost to compute hours.

What is NOT a valid case is covered in the ethics section. Skip it now if you don't care about the tutorial, but read it before doing anything.

---

The 3 tools that matter in 2026

The open-source voice cloning ecosystem is small but dense. Three projects concentrate ~90% of real-world usage:

1. GPT-SoVITS — top quality with fine-tuning

Developer: RVC-Boss community (same group behind Retrieval-based Voice Conversion).
License: MIT.
When to use it: when you need top quality and have 10+ minutes of clean audio + an NVIDIA GPU ≥6 GB VRAM. Gold standard of the community.

GPT-SoVITS combines a GPT-style language model with SoVITS (SOTA few-shot TTS architecture) and delivers quality that, with fine-tuning, competes with ElevenLabs. The catch: initial setup is non-trivial (Python 3.10, PyTorch with CUDA, preprocessed dataset), and fine-tuning takes 2-6 hours on an RTX 4090.

2. OpenVoice V2 — instant cloning

Developer: MyShell (with MIT collaboration).
License: MIT, with explicit commercial use allowed.
When to use it: when you want to clone a voice without training, with 10-30 s of sample. Almost instant, multilingual and cross-lingual (you can clone a voice in Spanish and make it speak in English keeping the timbre).

OpenVoice V2 (released April 2024) is the tool I recommend to try first. Installation is three commands and few-shot quality is decent. If the result convinces you, that's when you decide whether to move to GPT-SoVITS for more quality.

3. Fish Speech 1.5 / S2 — commercial production

Developer: Fish Audio (team based in HK).
License: Apache 2.0 (1.5) and dual-license model (S2 with commercial plan).
When to use it: when you need <100 ms latency for production (voice agent, kiosk, IVR). Trained on 10+ million hours of audio, 50 languages including Spanish.

Fish Speech S2 (speech.fish.audio) delivers the best combination of quality + latency + language coverage among open-source options. The catch: the S2 model is large (2-4 GB) and the commercial version has cost once you cross certain volume, but the base 1.5 model is fully free.

---

Comparison table

Reading the table:

• OpenVoice is the quick option to experiment and validate whether voice cloning fits your case. If it works, you decide.
• Fish Speech is the choice for commercial production: ~100 ms latency, permissive license, 50 languages including Latin American Spanish.
• GPT-SoVITS delivers the best final quality if you have time to train / fine-tune with 10+ minutes of clean audio.

---

Step-by-step tutorial: GPT-SoVITS

This is the complete tutorial for the "max quality" case. Assumes you have Linux or Windows, an NVIDIA GPU with ≥6 GB VRAM, and patience for the setup.

Prerequisites

• NVIDIA GPU with ≥6 GB VRAM (driver + CUDA 12.4 or higher). Tested on RTX 3060, 4060, 3090, 4090.
• Python 3.10 (not 3.11 or 3.12 yet — community recommends 3.10 for max compatibility).
• FFmpeg installed (for audio preprocessing).
• Disk: ~15 GB free for repo + dependencies + pretrained models.

Step 1: clone the repo and install dependencies

git clone https://github.com/RVC-Boss/GPT-SoVITS.git
cd GPT-SoVITS
pip install -r requirements.txt

If you're on Windows and the install fails, the repo has detailed Conda instructions. On Mac Apple Silicon it runs via MPS but is slow for training; usable for inference only.

Step 2: download pretrained models

# Base GPT-SoVITS and UVR5 (voice separator) models
huggingface-cli download RVC-Boss/GPT-SoVITS pretrained_asr pretrained_vits pretrained_semantic bert --local-dir pretrained_models
huggingface-cli download lj1995/VoiceConversionWebUI uvr5_weights --local-dir uvr5_weights

If huggingface-cli is not installed: pip install -U "huggingface_hub[cli]". Total ~6 GB.

Step 3: record and prepare your dataset

The quality of your sample audio is the most important variable. Bad audio = bad clone, no matter how many epochs you train.

Recording rules:

1. Quiet place. No loud AC, no traffic, no TV in the background. A closet with hanging clothes is surprisingly good as a DIY anechoic chamber.
2. Decent microphone. A decent USB cardioid mic (Blue Yeti, Fifine K669, AT2020+) is enough. You don't need a USD 1K Neumann. Avoid the laptop mic, it picks up too much ambient noise and reverb.
3. Format: 16-bit 44.1 kHz mono WAV. Audacity or your preferred recorder.
4. Content: read aloud with varied tones (not monotone). 10-30 minutes is ideal for full fine-tune; 30 seconds is enough for few-shot.
5. No background music, no reverb, no clipping (the red bar on the meter).

Once recorded, preprocess with the built-in tool:

# UVR5 separates voice from background noise (optional)
python uvr5/uvr5.py --input_path your_audio.wav --output_path processed/

# SubVoice: auto-segments into short phrases
python GPT_SoVITS/s2_train_prepare_v3.py --input_audio processed/ --output_dir dataset/

Step 4: train the model

For few-shot (a few seconds of sample, immediate result, moderate quality):

# Few-shot: loads pretrained model and passes your sample
python webui.py --api_only --api_port 9880
# Open http://localhost:9880/ in browser
# "Few-Shot" tab → upload your 10-30 s audio + text → click "Infer"

For full fine-tune (10+ min sample, 2-6 h training on RTX 4090, top quality):

# 1. Preprocess dataset
python GPT_SoVITS/s1_train_prepare.py --input_dir dataset/ --output_dir output/

# 2. Train GPT model (text encoder)
python GPT_SoVITS/s2_train.py --config configs/s2.json --output_dir output/ --num_workers 4

# 3. Train SoVITS model (audio decoder)
python GPT_SoVITS/s3_train.py --config configs/s1.json --output_dir output/ --num_workers 4

First training takes 2-6 hours; subsequent ones with transfer learning drop to 30-60 min.

Step 5: inference with WebUI

python webui.py
# Opens http://localhost:7860 (Gradio WebUI)
# "Inference" tab → choose your trained model → type text → "Generate"

Result: a WAV of your voice saying the text you typed, in the timbre you cloned. You sent nothing to the cloud — all processing was local.

Costs and times

If you want to avoid buying a GPU, RunPod / Lambda Labs rent RTX 4090 at USD 0.50-0.70/hour. For one fine-tune you spend USD 2-4 and get a reusable offline model.

---

Quick tutorial: OpenVoice V2

If you want to test in 5 minutes whether voice cloning fits your case, OpenVoice is the choice.

Installation

git clone https://github.com/myshell-ai/OpenVoice.git
cd OpenVoice
pip install -e .
# Download models (V2 checkpoints)
python -c "from openvoice import download; download('v2')"

Cloning with a 10-30 s sample

from openvoice import ToneColorConverter
from melo.api import TTS

# 1. Generate base audio in some language with a pretrained speaker
text = "Hello, this is a voice cloning test with OpenVoice."
tts = TTS(language="EN", device="cuda:0")  # "cpu" if no GPU
src_path = tts.tts_to_file(text, speaker_id=0, output_path="base.wav")

# 2. Apply your sample timbre (10-30 s WAV, clean)
converter = ToneColorConverter("checkpoints/v2/converter/config.json")
converter.download("v2")

target_path = "my_voice.wav"  # your sample audio, 10-30 s, mono 44.1kHz
output_path = "cloned_result.wav"
converter.convert(
    audio_src_path=src_path,
    audio_se_path=target_path,
    tone_color_converter=converter,
    output_path=output_path,
)

Result: cloned_result.wav has the spoken content of base.wav (text in English) in the timbre of my_voice.wav. With no training. The process takes 2-5 seconds.

Useful variants

• Cross-lingual: change language="EN" to "ES" / "JP" / "FR" and the speaker keeps your sample's timbre in another language.
• More quality: pass a longer sample (1-3 min) and you'll notice prosody improvement.
• REST API: OpenVoice V2 exposes a FastAPI server you can call from any language.

---

Quick tutorial: Fish Speech

If you need low latency (<100 ms time-to-first-audio) and a commercial-friendly license, Fish Speech is the direct choice.

Installation with Docker (one command)

docker run -d --name fish-speech \
  -p 8260:8260 \
  -v $(pwd)/checkpoints:/app/checkpoints \
  fishaudio/fish-speech:latest

The container exposes an API at http://localhost:8260 with inference and reference voice endpoints.

Cloning with a 10-30 s sample

# 1. Upload your reference audio (10-30 s, mono, 44.1 kHz) to a directory
cp my_voice.wav checkpoints/ref_audio/

# 2. Inference via API
curl -X POST http://localhost:8260/v1/tts \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Hello, this is Fish Speech cloning my voice.",
    "reference_audio": "my_voice.wav",
    "reference_text": "Exact text of what I say in my reference audio.",
    "max_new_tokens": 2048
  }' \
  --output result.wav

Time-to-first-audio: ~100 ms on RTX 4090, ~300 ms on modern CPU. Quality for 10-30 s samples is comparable to OpenVoice with the advantage of lower latency and solid Spanish support.

Fish Audio S2 (commercial) version

If you want production quality without training and the lowest possible latency, Fish Audio S2 (updated version with dual-license model) delivers high-quality zero-shot audio, supports 50 languages, and exposes the official API at api.fish.audio. The free plan covers 20K characters/month; production scales to USD 0.10-0.20 / 1M characters.

---

Output quality: how to improve it

The quality of the cloned voice depends, in order, on:

1. Sample audio quality

This is the variable with the most impact. A bad sample won't be fixed by more epochs or more VRAM.

Audio that ruins the clone:

• Reverberation from empty room or bathroom (the model learns the echo as part of the voice).
• AC, fridge, computer noise (stays in the background of every synthesis).
• Background music, even faint.
• Multiple speakers (the model doesn't know which one to clone).
• Clipping (the voice "cuts" at the peaks; usually because the mic was too close or gain too high).
• Lossy codecs (MP3, AAC): they compress and lose high harmonics. Use WAV or FLAC.

How to record good audio at home:

1. Lock yourself in a closet with hanging clothes (natural attenuation ~10-15 dB).
2. Use a USB cardioid mic ~15-20 cm from your mouth.
3. Record with Audacity or REAPER, 16-bit 44.1 kHz mono WAV.
4. Do 5-10 minutes of read-aloud with varied prosody (paragraphs, questions, exclamations). Avoid monotone.

2. Amount of sample

For few-shot (OpenVoice, Fish Speech): 10-30 seconds is enough for decent quality. With 1-3 minutes you'll notice improvement. With more than 5 minutes the marginal improvement is small.

For fine-tune (GPT-SoVITS): 10+ minutes is the sweet spot. There are examples of successful fine-tunes with 30 minutes; over 1 hour the improvement is marginal unless the audio is super consistent.

3. Audio cleanup

Before training, run your audio through:

• UVR5 (included in GPT-SoVITS) to separate voice / noise / reverb.
• Audacity Noise Reduction (native effect, manual but works) to clean up hiss.
• Auphonic (web service with a 2-hour monthly free tier) to normalize loudness and apply gentle EQ.

4. Prosody and emotion

Few-shot models inherit the prosody of your sample. If your audio is monotone, the clone will sound robotic. Record varied: read news, ask a question, exclaim, whisper. Current models copy that variety when synthesizing.

5. Evaluation

Objective evaluation:

• WER (Word Error Rate): transcribe generated audio with Whisper and compare to original text. WER <5% is good.
• Speaker similarity cosine: use speaker verification models like Resemblyzer or pyannote.audio to measure how similar the clone is to the original voice.

Subjective evaluation:

• MOS (Mean Opinion Score): ask 5-10 people to listen to the clone and rate naturalness 1-5. A good fine-tuned clone reaches MOS 4.0-4.5; ElevenLabs Pro reaches 4.5-4.7.

---

Troubleshooting: common issues

---

Hardware: what you need

Practical conclusion: for few-shot (OpenVoice / Fish Speech), any modern laptop works. For fine-tune with GPT-SoVITS, an RTX 3060 with 12 GB is the floor; RTX 4090 is the choice if you train often. On Apple Silicon, fine-tune is viable but 5-10× slower than on NVIDIA CUDA.

If you haven't bought hardware for local AI yet, the RAM guide for LLM in 2026 explains how much you need per model. For voice cloning specifically, the bottleneck is VRAM not RAM.

---

Ethics and legal: what not to do

This is not an addendum: it's a required section. Cloning voices without consent is one of the most heavily regulated areas of the last 12 months.

What you must NOT do (in any serious jurisdiction)

1. Clone someone else's voice without written, signed consent. Even if it's public (YouTuber, podcaster, actor). The "it's on the internet" argument is not a legal defense.
2. Impersonate anyone (bank fraud, bypassing voice verification at a call center, phone scams). This is fraud, with or without AI.
3. Create misleading political ads with a candidate's voice. In the US, the FTC has issued guidance on this.
4. Train on third-party audio datasets without permission (audiobooks, podcasts, music, call center recordings). This is what's called "ethical model collapse" and is the subject of multiple 2024-2025 lawsuits (Reuters, July 2025 on Lovo / Lehrman, ELVIS Act lawsuits).
5. Distribute models trained with someone else's voice. This is the worst category of 2025 regulation.
6. Use voice cloning to harass, intimidate, or produce sexual abuse (deepfake porn audio with voice). In many jurisdictions this is a criminal offense with prison time.

Regulation in force 2025-2026 you should know

Written consent, serious version

If you're going to clone a voice other than your own, the consent should look like:

"I, [full name], ID [number], expressly authorize [your name] to clone my voice using artificial intelligence techniques for the following uses: [specific list: audiobook, podcast, voice agent, etc.]. This authorization is for [duration]. I may revoke it with 30 days notice."

Signed, dated, two witnesses if possible. It's not paranoia: it's what covers your legal liability if the person revokes permission or a third party makes a claim.

---

Operational checklist

Before cloning your voice (or a voice with permission), go through this list:

• Confirm you have the right to clone the voice. If it's not yours, you have signed written consent.
• You recorded clean audio, 16-bit 44.1 kHz mono WAV, no noise or reverb, varied prosody, 10-30 min if fine-tuning.
• You preprocessed the audio (UVR5 or Audacity to separate voice from noise).
• You picked the right tool for your case: OpenVoice to try, Fish Speech for production, GPT-SoVITS for max quality.
• You have the necessary hardware (NVIDIA GPU ≥6 GB for GPT-SoVITS; light CPU/GPU for OpenVoice/Fish Speech).
• You verify quality post-training with WER (Whisper) and a quick MOS with 3-5 people.
• You will NOT use the clone to impersonate, defraud, harass, or mislead.
• If you deploy as a service, you add disclosure that it's an AI voice (EU AI Act compliance).

---

Conclusion

Cloning your voice for free in 2026 is totally viable, the three leading open-source tools (GPT-SoVITS, OpenVoice V2, Fish Speech) are mature and run locally. The decision between the three depends on your case:

• Test whether voice cloning fits you, validate with a client, do a demo: OpenVoice V2, 5 minutes of setup.
• Audiobook, dubbing, personal branding with max quality: GPT-SoVITS, 2-6 h fine-tune on an RTX 4090 or renting GPU on RunPod.
• Commercial production with <100 ms latency: Fish Speech 1.5 (Apache 2.0) or Fish Audio S2 (commercial plan).

Output quality depends 80% on your sample audio and 20% on the model. If you record clean and varied audio, all three give you results comparable to paid services. If the audio is bad, no model will fix it.

And on ethics: it's not an addendum at the end of the post, it's part of the workflow. Written consent for other people's voices, mandatory AI disclosure when deploying, zero tolerance for impersonation or fraud.

If your startup is exploring self-hosting of TTS or voice agents and wants to validate the setup with your real workloads before committing to hardware or a paid vendor, book a free 30-minute call — we can usually calculate your total cost and the right architecture in one session.

---

Also read:

• How much RAM do you need to run local AI in 2026 — hardware guide for self-hosting models.
• OpenAI GPT-Live: full-duplex voice free on the Free plan — voice AI for SaaS in Latam.
• Local AI vs cloud AI in 2026 — when to self-host TTS vs consuming ElevenLabs/PlayHT.
• 7 questions to choose an AI consultant — if you want to outsource the setup.
• Back to blog — all articles.

Frequently asked questions

How can I clone my voice for free with AI?

Three open-source tools let you clone your voice for free in 2026: GPT-SoVITS (best quality with 10+ min dataset, needs a GPU), OpenVoice V2 (instant with 10-30 s sample, MIT license), and Fish Speech 1.5/S2 (~100 ms latency, Apache 2.0, multilingual). All run locally with no data leaving your machine. Step-by-step tutorial with commands for each tool.

What free tools exist to clone voices in 2026?

The three most mature are GPT-SoVITS (RVC-Boss, MIT license), OpenVoice V2 (MyShell + MIT, April 2024), and Fish Speech 1.5/S2 (Fish Audio, Apache 2.0). Each solves a different case: GPT-SoVITS for top quality with fine-tuning, OpenVoice for few-shot cloning without training, Fish Speech for commercial production with &lt;100 ms latency.

How much audio do I need to clone a voice?

Depends on the tool. OpenVoice V2 and Fish Speech work with 10-30 seconds of clean audio for few-shot cloning. GPT-SoVITS improves significantly with 5-30 minutes for a full fine-tune. The audio must be clean: 16-bit 44.1 kHz mono, no background noise, no reverberation, no music. Recorded in a quiet place with a decent microphone.

Is it legal to clone someone else's voice with AI?

Not without written consent. The ELVIS Act in Tennessee (US, 2024) penalizes non-consensual distribution of voice clones. The EU AI Act mandates deepfake disclosure on AI-generated content. Most jurisdictions treat non-consensual voice cloning as illegal and punishable, with penalties ranging from civil liability (right of publicity suits) to criminal fraud charges for impersonation. Always get signed, written permission before cloning anyone's voice other than your own.

Do I need a dedicated GPU to clone voices with AI?

Not to try: OpenVoice and Fish Speech run on CPU (slow but usable for few-shot). Yes to train: GPT-SoVITS few-shot needs ~6 GB of VRAM; full fine-tune ~12 GB. An RTX 3060 with 12 GB is the reasonable minimum. On Apple Silicon, OpenVoice and Fish Speech run via CPU/MPS; GPT-SoVITS needs more RAM and patience.

How does a free AI-cloned voice compare to ElevenLabs or PlayHT?

With a decent sample (5-30 min of clean audio) and GPT-SoVITS fine-tuning, the quality is comparable to ElevenLabs Pro. Fish Speech S2 with its commercial license and 50-language model is the best open-source alternative for quality + latency in production. The difference is that free tools require self-hosting and your own sample audio, while ElevenLabs gives you everything as a service with pre-trained voices.