Open-source text-to-speech models compared: licence, VRAM and what each is for

The open text-to-speech model I would start with is Kokoro: 82M parameters, Apache-2.0, 54 preset voices in 8 languages, and small enough to run on a CPU. For cloning a voice under a licence that allows commercial use, I would pick Qwen3-TTS or VoxCPM2, both Apache-2.0. The licence is where most comparisons go wrong. F5-TTS, OmniVoice and Coqui's XTTS-v2 are non-commercial, Fish Audio's Drama 3 is free only below USD 500,000 in annual revenue, Microsoft's model cards limit VibeVoice to research although its licence is MIT, and Chatterbox is retired. None of these needs a 48 GB card to speak. A big card earns its place for batch synthesis, for the largest models and for serving many requests at once.

The open TTS models, compared#

I checked every row on the model card, licence file or repository on 2026-11-12, and chose the models from what people search for and what Hugging Face lists as most downloaded.

F5-TTS's parameter count is ours, read from its checkpoint, and Higgs TTS 3's comes from Hugging Face's count of its weights. Everything else is the maker's own figure. "Coqui TTS" usually means the library rather than a model. The original repository has had no release since December 2023 and no commit since February 2024, and its maintained fork, idiap's coqui-tts, version 0.28.0 of October 2026 under MPL-2.0, describes it as unmaintained. The library is free to use commercially, while the XTTS-v2 weights it is best known for are not.

What each licence allows#

The licence decides more than the voice quality, because a voice that is free for a hobby project can be off-limits for your company.

Fish Audio's older open models carry stricter terms than Drama 3: S2 Pro, from March 2026, is under the Fish Audio Research License, which grants no commercial rights, and the earlier OpenAudio S1-mini and Fish Speech 1.5 are CC-BY-NC-SA-4.0. VibeVoice's history is the other caution. Microsoft removed the VibeVoice-TTS inference code from its repository in September 2025, citing use inconsistent with its stated intent, and disabled the Large model, so the 1.5B weights still on Hugging Face no longer come with Microsoft's own code to run them. Treat this table as a summary and read each licence before you build a product on a voice.

Which model for which job#

The rule I follow is to settle the licence first, then pick on features, and only then on quality, which you can only judge on your own text.

For speech going the other way, transcribing audio with Whisper covers speech-to-text on the same GPUs.

When a 48 GB card earns its place#

The honest answer is that most of these models do not need a big GPU, and several run on a CPU. Kokoro's weights are 0.33 GB, VoxCPM2's maker asks for about 8 GB of VRAM, and Microsoft reports real-time speech from VibeVoice-Realtime on an NVIDIA T4. Every QuantaCloud GPU has at least 48 GB, so a single voice on a single request uses a small slice of the card. Three jobs use the rest.

The RTX A6000 is the card I would rent for everything except serving Drama 3 at scale, which belongs on the RTX PRO 6000 Blackwell.

Run Kokoro on a GPU VM#

The shortest recipe uses QuantaCloud's Bare Metal template, an Ubuntu 22.04 VM with the NVIDIA driver and Docker that you reach over SSH as ubuntu. Deploying a GPU walks through the launch and connecting over SSH the login.

1. Check the driver. The PyTorch build PyPI installs today uses CUDA 13.2, which needs driver 580 or newer:

nvidia-smi --query-gpu=name,driver_version --format=csv

2. Install espeak-ng, which Kokoro uses for words its dictionary does not know, and create a Python 3.12 environment with uv, because kokoro 0.9.4 supports Python 3.10 to 3.12:

sudo apt-get update && sudo apt-get install -y espeak-ng
curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env
mkdir -p ~/tts && cd ~/tts
uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate

3. Install Kokoro. On a driver older than 580, first install PyTorch's CUDA 12.8 build, which runs on older drivers, with uv pip install torch --index-url https://download.pytorch.org/whl/cu128:

uv pip install "kokoro>=0.9.4" soundfile

4. Save this as speak.py. It reads a text file, speaks it in the af_heart American English voice and writes one WAV file:

import sys

import soundfile as sf
import torch
from kokoro import KPipeline

pipeline = KPipeline(lang_code="a")  # "a" is American English, "b" British English
text = open(sys.argv[1], encoding="utf-8").read()
audio = torch.cat([chunk for _, _, chunk in pipeline(text, voice="af_heart")])
sf.write(sys.argv[2], audio.numpy(), 24000)
print(f"wrote {sys.argv[2]}: {len(audio) / 24000:.1f} seconds of audio")

5. Run it, then copy the file to your own machine before you stop the instance, because Stop deletes the disk:

python speak.py chapter1.txt chapter1.wav

Kokoro downloads its weights from Hugging Face on the first run. For cloning, give Qwen3-TTS a fresh environment of its own, as Qwen recommends, because qwen-tts pins its own transformers version: create a second venv the same way, run uv pip install -U qwen-tts in it, and take the code for each of its three model types from its model card. Checking the driver and CUDA version explains the nvidia-smi numbers, and moving files to and from a GPU server covers getting the audio off.

The official repositories#

Every model above, at its maker's own address:

Questions about open-source TTS#

What is the best open-source TTS model?

There is no single answer for every voice and language, so I choose on licence first. For commercial work that means Kokoro for fixed voices and Qwen3-TTS or VoxCPM2 for cloning, all Apache-2.0. Drama 3 handles long scripts with up to 16 speakers and Higgs TTS 3 is free for creators' own content, but read both licences before you rely on them.

Which open-source TTS models allow commercial use?

Kokoro, Qwen3-TTS, VoxCPM2 and Fun-CosyVoice 3 under Apache-2.0, Chatterbox under MIT, and Orpheus under Apache-2.0 with Meta's Llama 3.2 terms. Drama 3 allows it only below USD 500,000 in annual revenue. F5-TTS, OmniVoice and XTTS-v2 do not, and Microsoft's model cards limit VibeVoice to research, though its licence is MIT.

Can I clone a voice with open-source TTS?

Yes, with Qwen3-TTS, VoxCPM2, Fun-CosyVoice 3, Chatterbox, Drama 3, Higgs TTS 3, F5-TTS, OmniVoice and XTTS-v2, usually from a few seconds of reference audio. Fish Audio's licence and the Higgs and VibeVoice model cards all forbid imitating a real person without consent.

Does text-to-speech need a GPU?

Not always. Kokoro and Chatterbox Nano run on a CPU, and most of the others fit a small GPU. A large GPU pays for itself on batch jobs, on Drama 3, and when you serve many streams at once.

My decision rule for text to speech: Kokoro for fixed voices, Qwen3-TTS or VoxCPM2 when you need a cloned voice in a product, and Drama 3 on a 96 GB card only once its revenue limit and hosted-service clause fit your business. Run them on the smallest GPU on offer, copy the audio off before you stop, and the guides have the rest of the stack for a voice pipeline.