How to choose an open-source text-to-speech model

Compare open text-to-speech models by languages, latency, voice cloning and license: Qwen3-TTS, VoxCPM2, Fun-CosyVoice3, OmniVoice, Voxtral TTS, Chatterbox and tiny CPU models like Kokoro.

  • Beginner
  • 8 min read
  • Published
  • Updated

Short answer. For a product, start with the Apache 2.0 models: Qwen3-TTS for streaming voice agents in 10 languages, VoxCPM2 for 30 languages and 48 kHz audio, Fun-CosyVoice3 for Chinese and its dialects, and the new ZONOS3-Mini for 43 languages with fine control over delivery. For 11 Indian languages, Maya 2 Native streams under Apache 2.0 on an NVIDIA GPU. Four well-known models are non-commercial, Chatterbox was retired in August 2026, and VibeVoice is limited to research. With no GPU, the small Kokoro and Phonon 31M run on a laptop CPU under Apache 2.0. Whatever you choose, clone a real person’s voice only with their consent.

What to look for

A text-to-speech model reads text aloud. Four things separate the open ones:

  1. Languages. Supported languages range from 9 or 10 to 600+, and quality differs inside each list.
  2. Latency. A voice assistant needs audio to start within a fraction of a second, so look for streaming.
  3. Voice control. Some clone a voice from a short clip, some design a new voice from a text description, and some do both. See voice cloning.
  4. License. This is the one that surprises people. Several popular models are non-commercial.

Our text-to-speech models list covers every model we track, and the speech models list adds recognition and voice-to-voice models.

The shortlist

Model Size License Languages What stands out
Qwen3-TTS 0.6B, 1.7B Apache 2.0 10 3-second cloning, voice design, streaming at about 97 ms
VoxCPM2 2B Apache 2.0 30 Voice design from a description, cloning, 48 kHz output
Fun-CosyVoice3 0.5B Apache 2.0 9 plus 18+ Chinese dialects Zero-shot cloning, streaming from 150 ms
ZONOS3-Mini 0.8B, 0.1B active Apache 2.0 43 Cloning, controls for emotion and single words, runs in a browser
Maya 2 Native 2.5B Apache 2.0 12 (11 Indian plus English) Streaming, 20 voices, cloning, NVIDIA GPU needed
VieNeu-TTS v3 131M, 48M preview Apache 2.0 2 (Vietnamese and English) 23 regional voices, cloning, 48 kHz, ONNX on a CPU
MOSS-TTS 8B, 4B Apache 2.0 31 Pinyin and IPA input, duration control, exact pauses, cloning
MOSS-TTSD 8B Apache 2.0 20 Dialogue with up to 5 speakers and 60 minutes in one pass
Drama 3 36B, 3B active Fish Audio Community License 83 Acted scenes with up to 16 speakers; free below US$500K revenue
S2 Pro 4.56B Fish Audio Research License 80+ Free-form inline tags, cloning from 10 to 30 seconds; non-commercial only
Higgs TTS 3 4.65B Boson research and non-commercial, creator grant 100+ Tags for emotion, sound effects and pauses; products need a license
OmniVoice 0.6B Weights CC BY-NC, code Apache 2.0 600+ Cloning and voice design in low-resource languages
Voxtral TTS 4B CC BY-NC 9 Voice adaptation from a short clip
VibeVoice TTS 1.5B, 0.5B realtime MIT, research use Realtime model: English Up to 90 minutes and 4 speakers; realtime model starts in about 300 ms
Chatterbox 0.5B MIT 21 to 23, depending on the source Retired in August 2026
LFM3-Audio 2B LFM Open License 18 Speech in and out, on-device
Kokoro 82M Apache 2.0 8 54 fixed voices, runs on a CPU, no cloning
Pocket TTS 100M CC BY 4.0, gated 6 Cloning from about 5 seconds, runs on a CPU
MOSS-TTS-Nano 100M Apache 2.0 Around 20 Cloning, 48 kHz stereo, ONNX route on a CPU
Phonon 31M 31M Apache 2.0 English 64.5 MB file, 37 voices
Phonon 40M 40M CC BY-NC, gated 5 One model per language or one multilingual model, 74.6 MB files
Supertonic 3 99M OpenRAIL-M 31 ONNX on a CPU, archived in September 2026

Check the language list on the model card before you rely on it: for example, the Chatterbox card and its announcement give slightly different language counts.

Pick by the job

  • A voice assistant or agent. Choose streaming. Qwen3-TTS reports about 97 ms in its best case, Fun-CosyVoice3 starts streaming from 150 ms, and Mistral’s Voxtral TTS is designed to pair with Voxtral Realtime for speech in and speech out. Latency depends on your hardware and serving stack, so measure it yourself.
  • Narration, audiobooks and video voice-overs. Qwen3-TTS and VoxCPM2 cover many languages under Apache 2.0. For very long, multi-speaker audio, VibeVoice supports up to 90 minutes and 4 speakers, but only for research use, while Apache 2.0 MOSS-TTSD handles up to 5 speakers and 60 minutes in one pass. For acted dialogue, Drama 3 takes directions such as [whispering] and up to 16 speaker tags in one script, but its license caps free commercial use at US$500,000 a year.
  • Chinese, including dialects. Fun-CosyVoice3 and VoxCPM2.
  • Indian languages. Maya 2 Native covers Hindi, Tamil, Telugu, Bengali and seven more Indian languages plus English under Apache 2.0. Its error rates come from Maya’s own automated judges, and Assamese is the weakest, so test the languages you need. It runs on an NVIDIA GPU with CUDA.
  • Vietnamese. VieNeu-TTS v3 is a single developer’s 131M model with 23 voices from the North, Central and South regions, cloning from a 3 to 8 second clip and Vietnamese-English code-switching, all under Apache 2.0. Its card gives speed figures but no quality benchmark, so listen before you commit.
  • Exact pronunciation and timing. MOSS-TTS takes Pinyin or IPA input, sets duration by token count and accepts pauses such as [pause 3.2s] in 31 languages, under Apache 2.0. Its v1.5 cards publish no new benchmark, so test the languages you need.
  • Laughs, whispers and pauses written into the text. Higgs TTS 3 takes tags for 21 emotions, nine sound effects, speed and pauses, and Boson reports single-digit error rates in 102 languages. It is free only for research, personal use and creators who credit Boson. Fish Audio’s S2 Pro takes free-form tags such as [whisper in small voice] and runs on a 24 GB GPU, but it is research and non-commercial only.
  • A voice you design rather than clone. VoxCPM2 and Qwen3-TTS can create a new voice from a text description, without recordings.
  • A rare language. OmniVoice lists 600+ languages. Its weights are non-commercial, and results can be unstable in low-resource languages, so treat it as a research tool.
  • Private, on-device voice. LFM3-Audio is a 2B model for speech in and speech out on laptops and edge devices, free to use commercially below US$10 million of revenue.
  • Cloning with fine control on a laptop. ZONOS3-Mini clones voices in 43 languages, with controls for accent, emotion and single-word emphasis. Zyphra says it runs in real time on laptops, phones and in a browser, but it launched on 9 November 2026, so it has little track record.
  • No GPU at all. Kokoro, Pocket TTS, Phonon 31M and Supertonic 3 are small enough for a laptop CPU. Phonon 31M is the smallest at 64.5 MB, Pocket TTS is the one that clones voices, and Kokoro is the most widely used. MOSS-TTS-Nano and VieNeu-TTS v3 are two newer CPU options that also clone voices. Tiny text-to-speech models that run on a CPU compares them, and Kokoro vs Pocket TTS covers the first decision.

OmniVoice vs VoxCPM2 shows the trade-off between language reach and license in one place.

Read the license before you fall for the demo

The license decides whether a promising voice model can ship. Five patterns to know:

  • Code and weights differ. OmniVoice’s code is Apache 2.0, but its weights are CC BY-NC, so commercial use is not allowed.
  • Non-commercial by inheritance. Voxtral TTS carries CC BY-NC, which Mistral’s card ties to the reference voices, so you need Mistral’s hosted service or a different license for a product.
  • A creator exception inside a non-commercial license. Higgs TTS 3’s Boson license lets creators monetize podcasts, videos and audiobooks on channels they own if they credit Boson AI’s Higgs Audio. Serving the model, building it into a product and training other speech models still need a separate agreement.
  • Research-only intent. Microsoft’s VibeVoice cards limit use to research, and Microsoft removed the original TTS code from GitHub in September 2025 after misuse.
  • Use restrictions that travel with the weights. Supertonic 3 uses OpenRAIL-M, which allows commercial use but requires every copy and derivative to carry a list of banned uses. The glossary entry OpenRAIL explains the pattern.

Revenue caps matter too: the LFM Open License ends at US$10 million of yearly revenue, and the Fish Audio Community License behind Drama 3 ends at US$500,000 and never covers a hosted text-to-speech or voice-cloning service. Its older sibling S2 Pro has no commercial tier at all, because the Fish Audio Research License counts even a company’s internal use as commercial. AI licenses explained and the glossary entry non-commercial license cover the rules in general.

Use voice cloning responsibly

Cloning a real person’s voice without permission can be illegal or harmful, and an Apache 2.0 or MIT license does not change that. The developers say so themselves: OpenBMB forbids impersonation, fraud and disinformation with VoxCPM2 and asks you to label AI-generated audio, the OmniVoice authors prohibit unauthorized cloning, impersonation and fraud, Fish Audio’s license bans imitation without consent and removing its watermark, and Boson’s license for Higgs TTS 3 requires explicit, verifiable consent and records for high-risk uses. Chatterbox outputs carry a PerTh watermark by default, while Maya says Maya 2 Native adds none, so labeling falls to you. Get consent from every speaker, label synthetic audio, and keep your own safeguards in the product.

Test it like a listener

Pick 20 sentences from your real use: names, numbers, acronyms, a long paragraph, and some text in each language you serve. Generate each with two or three candidates and listen blind. Record the time to the first audio on your hardware, because the published latency figures are best cases. Note that some models give different results between runs: OpenBMB says voice design and style control in VoxCPM2 vary, so you may need to generate a few times. How to evaluate AI models yourself explains how to turn this into a repeatable test.

For a full voice pipeline, add a recognizer in front, and how to choose a speech-to-text model covers that half.

Frequently asked questions

Which open text-to-speech model sounds the most natural?
We cannot give an independent ranking. The quality and latency comparisons on the model cards come from the developers, such as Qwen's own report and Mistral's human evaluations against ElevenLabs. Generate the same twenty sentences with two or three models and listen to them blind.
Can I clone a voice with an open model?
Yes. Qwen3-TTS, Fun-CosyVoice3, VoxCPM2, ZONOS3-Mini, Maya 2 Native, VieNeu-TTS v3, MOSS-TTS, MOSS-TTSD, MOSS-TTS-Nano, Drama 3, S2 Pro, Higgs TTS 3, OmniVoice, Voxtral TTS, Chatterbox and Pocket TTS all clone a voice from a short reference clip. Do it only with the speaker's consent, because an open license does not give you rights to anyone's voice.
Which open text-to-speech models can I use in a commercial product?
Qwen3-TTS, VoxCPM2, Fun-CosyVoice3, ZONOS3-Mini, Maya 2 Native, VieNeu-TTS v3, MOSS-TTS, MOSS-TTSD, MOSS-TTS-Nano, Kokoro and Phonon 31M are Apache 2.0. The weights of OmniVoice and Voxtral TTS are CC BY-NC, so they are non-commercial only. Higgs TTS 3 is research and non-commercial too, with one exception: creators may monetize podcasts, videos and audiobooks made on their own channels if they credit Boson, but a product or hosted service needs a paid license. Fish Audio's S2 Pro is research and non-commercial as well, and any commercial use needs a separate license from Fish Audio. VibeVoice is MIT but limited to research use. LFM3-Audio is free to use commercially only below US$10 million of yearly revenue, and Drama 3 only below US$500,000, with no hosted voice service at any size.
How much GPU memory does a text-to-speech model need?
It is modest compared with language models. OpenBMB says VoxCPM2 runs on about 8 GB of VRAM. By our estimate, the 4B Voxtral TTS needs about 9 GB in BF16, and the 0.5B models need about 1 to 2 GB. The 8B MOSS-TTS has 17.0 GB of BF16 weights, so it needs a GPU with more than that. Fish Audio's 4.6B S2 Pro asks for at least 24 GB of GPU memory, and Drama 3 is the heavy exception: its FP8 weights are 41.5 GB, so it needs a GPU with about 48 GB.
Which open text-to-speech models run on a CPU?
Kokoro (82M), Pocket TTS (100M), Phonon 31M, Supertonic 3 (99M), MOSS-TTS-Nano (100M) and VieNeu-TTS v3 (131M, Vietnamese and English) are small enough for a laptop CPU, according to their developers. Kokoro, Phonon 31M, MOSS-TTS-Nano and VieNeu-TTS v3 are Apache 2.0, Pocket TTS clones voices but its weights are gated, and Supertonic 3 is archived.
Is there an open model for Chinese dialects?
Fun-CosyVoice3 supports 9 languages and more than 18 Chinese dialects, and VoxCPM2 lists Chinese dialects among its 30 languages. Test your own dialect, because quality differs by language.

Key terms in this guide

  • Benchmark A benchmark is a standardized test used to compare models, such as a set of maths, coding or knowledge questions; scores are useful but easy to over-read.
  • Non-commercial license A non-commercial license lets you use a model for research, study and personal projects, but not to make money from it.
  • On-device AI On-device AI means the model runs on your own phone, laptop or small board instead of in the cloud, so data stays local and it works offline.
  • OpenRAIL (Responsible AI License) OpenRAIL is a family of licenses that let you use, modify and share a model, including commercially, but require every copy and derivative to carry a list of banned uses.
  • Text-to-speech (TTS) Text-to-speech models turn written text into spoken audio, and many can copy a voice from a short sample.
  • Voice cloning Voice cloning makes a text-to-speech model speak in a particular person's voice, often from only a few seconds of recording.

Models mentioned in this guide

Model Released Parameters Context License Smallest variant needs
Qwen3-TTS Alibaba Qwen 910M – 1.92B – Apache 2.0 (Permissive) 3.68 GB Laptop or phone
VoxCPM2 OpenBMB 2.29B – Apache 2.0 (Permissive) 8 GB Laptop or phone
Fun-CosyVoice 3 Tongyi Lab Speech Team 500M – Apache 2.0 (Permissive) 6.21 GB Laptop or phone
ZONOS3-Mini Zyphra 800M 100M active – Apache 2.0 (Permissive) 10 GB Laptop or phone
Maya 2 Native Maya Research 2.5B – Apache 2.0 (Permissive) 5.75 GB Laptop or phone
VieNeu-TTS v3 Phạm Nguyễn Ngọc Bảo (VieNeu-TTS) 48M – 131M – Apache 2.0 (Permissive) 0.322 GB Laptop or phone
MOSS-TTS OpenMOSS 4.55B – 8.49B – Apache 2.0 (Permissive) 10 GB Laptop or phone
MOSS-TTSD OpenMOSS 8.36B – Apache 2.0 (Permissive) 19 GB Gaming GPU
MOSS-TTS-Nano OpenMOSS 100M – Apache 2.0 (Permissive) 0.276 GB Laptop or phone
Drama 3 Fish Audio 35.9B 3B active – Fish Audio Community 1.0 (Conditional) 25 GB Gaming GPU
S2 Pro Fish Audio 4.56B – Fish Audio Research (Non-commercial) 24 GB Gaming GPU
Higgs TTS 3 Boson AI 4.65B – Boson Higgs Non-Commercial (Non-commercial) 11 GB Laptop or phone
OmniVoice Xiaomi 613M – CC BY-NC (Non-commercial) 3.75 GB Laptop or phone
Voxtral TTS Mistral AI 4.09B – CC BY-NC (Non-commercial) 9.2 GB Laptop or phone
VibeVoice-TTS Microsoft 1.02B – 2.7B 64K MIT (Permissive) 2.35 GB Laptop or phone
Kokoro hexgrad 82M – Apache 2.0 (Permissive) 0.38 GB Laptop or phone
Pocket TTS Kyutai 18M – 100M – CC BY 4.0 (Permissive) 0.127 GB Laptop or phone