Best open-source text-to-speech models in 2026
Open text-to-speech now covers hundreds of languages and can clone a voice from a few seconds of audio, and a new group of tiny models, from Kokoro to Pocket TTS and Phonon, runs on a laptop CPU or a phone. The catch is the license and the maintenance status. Several of the best-known models are non-commercial only, two popular projects have been retired or archived, and one has had its original code withdrawn. Read the license line before you ship anything.
-
Qwen3-TTS
Alibaba's Apache 2.0 family in 1.7B and 0.6B sizes: 10 languages, 3-second voice cloning and streaming at about 97 ms. Its 1.7B Base model had about 3.8 million downloads in the 30 days to 11 November 2026.
-
VoxCPM2
OpenBMB's 2B Apache 2.0 model: 30 languages, voice design from a text description, cloning and 48 kHz output on about 8 GB of VRAM.
-
Kokoro Most downloaded
hexgrad's Apache 2.0 82M model with 54 preset voices in 8 languages. It had about 12.5 million downloads in the 30 days to 11 November 2026, more than any other speech model here, but it cannot clone voices and has had no release since February 2025.
-
Fun-CosyVoice 3
Alibaba Tongyi's Apache 2.0 0.5B zero-shot model: 9 languages, 18+ Chinese dialects, voice cloning and streaming from 150 ms.
-
MOSS-TTS
OpenMOSS's Apache 2.0 flagship in 8B and 4B sizes: 31 languages, zero-shot cloning, Pinyin and IPA input, duration control and exact pauses. The v1.5 cards publish no new benchmark, and the 4B model writes 48 kHz stereo.
-
IndexTTS-2.5
Bilibili's 0.8B zero-shot cloner for Chinese, English, Japanese, Spanish and Arabic, with separate emotion, pronunciation and speed controls on about 6 GB of VRAM. Its own license has no general commercial ban but sets 100 million-user and RMB 1 billion thresholds.
-
ZONOS3-Mini Newest
Zyphra's 800M Apache 2.0 mixture-of-experts model, released on 9 November 2026: 43 languages, voice cloning and up to 10 minutes of 44.1 kHz audio in a 9 GB checkpoint. Zyphra says it runs in real time on laptops, phones and in a browser, and its first place is on Zyphra's own index.
-
Maya 2 Native
Maya Research's 2.5B Apache 2.0 streaming model for 11 Indian languages and English, with 20 built-in voices, cross-language cloning and a reported 224 concurrent streams on one H100 at a 111 ms median time to first audio. It needs an NVIDIA GPU, and Maya says it has no audio watermark.
-
Pocket TTS
Kyutai's 100M voice-cloning model that runs faster than real time on a laptop CPU, in 6 languages. The weights are CC BY 4.0 behind a prohibited-use gate, and each preset voice has its own license.
-
MOSS-TTS-Nano
OpenMOSS's 100M model for CPUs, with cloning, around 20 languages, 48 kHz stereo output, an ONNX route without PyTorch and a browser extension. It publishes no quality benchmark for the speech model, so listen to the demo.
-
Phonon 31M
Gradium's 31M English model in a 64.5 MB file under Apache 2.0, with 37 built-in voices. The multilingual 40M models are non-commercial, and cloning is not in the open weights.
-
Phonon 40M
Gradium's 40M models in 74.6 MB files: one each for English, French, German, Spanish and Portuguese, plus a multilingual one with 102 built-in voices. Gradium's own tables put them ahead of Kokoro, Magpie and NeuTTS, but the license is CC BY-NC and the weights are gated.
-
Supertonic 3
Supertone's 99M ONNX model for 31 languages, with ten preset voices and OpenRAIL-M weights that ban a list of misuses. The project was archived in September 2026, so expect no updates.
-
KittenTTS 2
KittenML's 1.7B ternary model for CPUs: 115 voices, cloning from 5 to 30 seconds of audio and emotion tags. The default file is 0.95 GB, and the README asks for about 6 GB of RAM or a GPU with roughly 8 GB. Commercial use needs registration and the grant ends at US$1 million in revenue or funding.
-
VieNeu-TTS v3
A single developer's 131M Vietnamese and English model, with 48 kHz output, 23 regional voices, cloning from 3 to 8 seconds and torch-free CPU inference. Everything, including the preset voices, is Apache 2.0, and the card publishes no quality benchmark.
-
Drama 3
Fish Audio's 36B mixture of experts with 3B active parameters and 83 languages, released on 13 October 2026, with directions such as [whispering] written into the text. Its license is free below US$500,000 a year in revenue, never covers a hosted voice service, and bans using the model or its output to train other speech models.
-
S2 Pro
Fish Audio's 4.56B previous-generation model with free-form tags such as [whisper in small voice], multi-speaker dialogue and cloning from 10 to 30 seconds. Fish Audio asks for a 24 GB GPU. Research and non-commercial only.
-
Higgs TTS 3
Boson AI's 4.65B model for voice chat: single-digit error rates in 102 languages according to Boson, zero-shot cloning and tags in the text for 21 emotions, nine sound effects, speed and pauses. The license is research and non-commercial, with a free grant for monetized creator content that credits Boson, and 114,000 downloads in the 30 days to 11 November 2026.
-
OmniVoice
Xiaomi's Next-gen Kaldi team covers 600+ languages with cloning and voice design, and had about 3.8 million downloads in the 30 days to 11 November 2026. The code is Apache 2.0 but the weights are CC BY-NC, so it is for non-commercial use.
-
Voxtral TTS
Mistral's 4B open-weights model: 9 languages and voice adaptation from a short clip. The CC BY-NC license makes it non-commercial only.
-
Chatterbox
Resemble AI's MIT-licensed 0.5B voice-cloning model, Multilingual V3 with 6 language packs. The project was retired in August 2026 and archived, so expect no updates.
-
Breeze TTS 2
BreezeBlue's 3.5B real-time English and Chinese model with voice cloning, voice design and delivery directions, and a reported time to first audio under 40 ms on an H100. The weights are non-commercial, and it needs Linux and a CUDA GPU, with 12 GB recommended.
-
AuK
Tencent's MIT-licensed 1.5B model that treats speech as instructions: it generates a voice from a description, rewrites words in a recording, changes emotion or accent, and cleans up or separates speakers. README peak memory is about 25 GiB in bf16, or about 17 GiB with CPU offload.
-
VibeVoice-TTS
Microsoft's MIT-licensed 1.5B model for up to 90 minutes and 4 speakers, plus a 0.5B realtime model. Microsoft limits the intended use to research and removed the original TTS code in September 2025.
-
MOSS-TTSD
OpenMOSS's Apache 2.0 8B model for scripted dialogue: 1 to 5 speakers, up to 60 minutes in one pass and cloning from 3 to 10 second references. It beats other open models on its own similarity tests but has higher word error rates than the proprietary systems it is compared with.
-
LFM3-Audio-2B
Liquid AI's 2B on-device model takes speech in and gives speech or text out: transcription, synthesis and translation in 18 languages. The LFM Open License allows commercial use only below US$10 million in yearly revenue.
How we chose
- Open weights you can download today from the developer's official repository.
- The license of the weights, not only the code, because voice models often carry non-commercial or revenue limits that differ from the code license.
- Voice quality is a matter of taste, so we describe features, languages and limits instead of ranking by sound. Listen to the developer's samples before you choose.
- Maintenance status is stated, since a retired or research-only project is a poor base for a product.