Best Open Source Text-to-Speech Models

Open source text to speech has never had more good options, or more ways to pick the wrong one.

The models are close enough on audio quality that the deciding factor is rarely how they sound. It is whether you can legally ship the voice, whether it runs on your hardware, and whether it does the specific job in front of you: a standalone voice track, an on-device assistant, a cloned voice, or speech that has to come out of a face on screen.

What counts as open source?

Open source in text to speech means three separate things, and a project can be one without being the others.

• Open code. F5-TTS publishes its inference and training code under MIT.
• Open weights. Downloadable is not commercially usable: F5-TTS's pre-trained models are CC-BY-NC, which the project attributes to its Emilia training data.
• Open licence on the voice. Piper's voices documentation states that each voice's MODEL_CARD carries important licensing information, and that some voices may have restrictive licences.

Knowing which layer a project fails on tells you whether the blocker is fixable by swapping a voice, or fatal. LTX-2.5 adds a fourth position to this map: open weights on a revenue-gated licence, covered next.

The best open source TTS models

#1 LTX-2.5

LTX-2.5 is the outlier on this list, and the reason to read it first is that it answers a different question. The other five generate a voice track as a standalone asset. LTX-2.5 generates speech as one stream of an audio-video model, synchronized to picture in the same pass.

It is a 22-billion-parameter asymmetric dual-stream diffusion transformer that models video and audio jointly rather than running text-to-video then video-to-audio. Its custom Gemma 4 12B text encoder produces separate video and audio context embeddings from the same prompt, which is what lets it synthesize speech that lip-syncs to on-screen mouth movement while staying natural in cadence, accent and emotional tone. Audio decodes through a HiFi-GAN-based vocoder to a 24 kHz stereo waveform.

For a genuinely audio-only use, the open-source pipelines include a text-to-audio path, T2AOneStagePipeline, that denoises the audio stream alone and writes a wave file with no video generated. That is the closest LTX-2.5 comes to a conventional TTS call, and it is also how you would drive an audio-style LoRA such as an accent or voice adapter.

On the article's own licence-first test, LTX-2.5 is open weights under the LTX-2.x Community License: free for commercial use under $10M in annual revenue measured across the whole entity, with a paid agreement required above that. That is more permissive than F5-TTS's non-commercial weights and less permissive than Kokoro's Apache 2.0, so it sits in its own category rather than at either end.

What it does not do is reference-voice cloning as a product feature the way F5-TTS and Coqui TTS do. If your need is a standalone voice track from arbitrary text, one of those is the better tool. If the words have to come out of a face on screen, this is the only model here that does it in one model. The LTX-2.5 model page covers setup, and the LoRA training path on the LTX Trainer covers custom voices.

#2 Kokoro

Kokoro is an open-weight TTS model with 82 million parameters, published on Hugging Face as Kokoro-82M under Apache 2.0, with StyleTTS2-LJSpeech as its base model. Those facts are the whole pitch: 82 million parameters is tiny next to multi-billion-parameter speech models, and Apache 2.0 on the weights themselves is the most permissive position here.

#3 Piper

Piper is a fast local neural text-to-speech engine that embeds eSpeak NG for phonemization, installed with pip install piper-tts and maintained by the Open Home Foundation under GPLv3. Its voices are trained with VITS and exported to ONNX Runtime, two files per voice: an .onnx model and an .onnx.json config. Its documentation lists voices across more than forty languages and locales, and names Home Assistant, the NVDA screen reader and the Open Voice Operating System as users. Two caveats: the Foundation is looking for maintainers, and the voices documentation states that Piper is intended for personal use and text to speech research only, while adding that the project imposes no additional restrictions of its own on voice models beyond each voice's licence.

#4 F5-TTS:

F5-TTS is a Diffusion Transformer with ConvNeXt V2 that uses flow matching, published by X-LANCE at Shanghai Jiao Tong University. Its distinctive inference feature is Sway Sampling, which the project says greatly improves performance. Reference-audio voice cloning is the default path: the CLI takes --ref_audio and --ref_text, and leaving --ref_text empty makes an ASR model transcribe the reference.

F5-TTS is also the only project here publishing a latency benchmark. Decoding on a single L20 GPU with 26 prompt-audio and target-text pairs at 16 NFE, F5-TTS Base with the Vocos vocoder records 253 ms average latency and a real-time factor of 0.0394 in client-server mode, 0.0402 offline with TensorRT-LLM, and 0.1467 offline in plain PyTorch. TensorRT-LLM is roughly 3.6x better on real-time factor, so the serving stack moves latency more than a model swap would. The constraint is the licence: MIT code, CC-BY-NC weights.

#5 Coqui TTS

Coqui TTS is a library rather than a single model, and it carries the widest catalogue here: pretrained models in over 1,100 languages, reached by combining its own with roughly 1,100 Fairseq MMS models addressed as tts_models/<lang-iso_code>/fairseq/vits. Its flagship XTTSv2 covers 17 languages and streams under 200 ms. Voice cloning works from a reference clip through speaker_wav, and from version 0.27.0 cloned voices can be cached under a custom speaker ID.

The maintenance story is where most outdated roundups go wrong. The original coqui-ai/TTS repository is described by its successor as unmaintained. Development continues in the idiap/coqui-ai-TTS fork, published on PyPI as coqui-tts rather than TTS, tested on Python 3.10 to 3.14, licensed MPL 2.0. If a tutorial says pip install TTS, it predates the fork.

#6 eSpeak NG

eSpeak NG is a compact open source formant synthesizer covering more than 100 languages and accents, released under GPLv3 or later, written in C, and available as a command line program, a shared library and a Windows SAPI5 voice. Its documentation is candid: the speech is clear and can be used at high speeds, but is not as natural or smooth as larger synthesizers based on human speech recordings. Program and data total a few megabytes. It is the right answer for accessibility tooling and languages neural projects skip.

Side-by-side comparison

Which licence can you ship?

Licence outcomes divide these six into groups, and the group decides your plan.

• Ship the weights, revenue-gated. LTX-2.5 is open weights under the LTX-2.x Community License, free to ship commercially under $10M in annual revenue and a paid agreement above it. More permissive than CC-BY-NC, less than Apache 2.0.
• Ship the published weights. Kokoro-82M is Apache 2.0 on the model itself, the cleanest position here for a commercial product.
• Ship, with obligations. Piper and eSpeak NG are GPLv3, so distributing either inside your own software brings copyleft obligations your legal team needs to see. Coqui TTS is MPL 2.0, file-level copyleft, but each pretrained model carries its own terms.
• Do not ship the published weights. F5-TTS's pre-trained models are CC-BY-NC. The code is MIT and fine-tuning is documented, so the commercial route is training on data you have rights to: a project, not a config change.

One further layer applies whatever the licence says: a cloned voice belongs to the person whose voice it is. Consent is separate from the model licence, and no licence grants it.

Where TTS fits a video pipeline

Text to speech produces a voice track as a separate asset, one of two routes to audio in a video. The other generates audio with the picture. LTX-2.5 unifies video generation with synchronized audio in a single model rather than running text-to-video then video-to-audio; the documentation states that this captures true joint dependencies and maps visual cues to auditory events including lip-sync, foley and environmental acoustics.

One documented detail explains why the routes differ in kind rather than degree: the text encoder produces separate video and audio context embeddings from the same prompt, which is what lets the model synthesize speech synchronized with visual lip movement while keeping it natural in cadence, accent and emotional tone. Audio decodes through a HiFi-GAN-based vocoder to 24 kHz stereo.

Pick the separate-asset route when you need exact control over the words, a specific voice, or a track you can re-cut. Pick generated audio when the sound has to match the picture frame for frame. The temporal consistency guide covers the visual side of the same problem, and the capabilities overview covers what else the model does in that pass.

Summary

The models on this list are close on audio quality, so the decision is made on the three questions quality does not answer: what you can legally ship, what runs on your hardware, and what the specific job needs.

Check the licence on the weights before you fall for the samples, because it is the constraint most likely to be fatal and the one most roundups skip.

If your voice track has to match the picture, compare the separate-asset route against generating both together with the open-source LTX-2 repository, and read the best open source video generation models for the video side.