Open speech models: transcription is open, text-to-speech is mixed

Ten of our eleven open speech-to-text models have no commercial limit, but only six of fourteen text-to-speech models are plain Apache 2.0. The rest are non-commercial, capped, restricted or retired.

  • Analysis
  • 4 min read
  • Published

Open speech models split along a license line. Across the 24 models in our speech recognition list and text-to-speech list, almost every transcription model is permissive, while the open voices come with more strings attached. For a voice product, check the license first.

Transcription: the license is rarely the problem

Ten of the eleven speech-to-text models in our list use permissive terms. Whisper large-v3 turbo and VibeVoice-ASR are MIT. Qwen3-ASR, Voxtral Realtime, Cohere Transcribe, Granite Speech 5.0 and Whistle are Apache 2.0. Nemotron 3.5 Transcribe and Nemotron 3.5 ASR use OpenMDW 1.1, and Parakeet TDT 0.6B v3 is CC BY 4.0, which asks only for credit. The eleventh, LFM3-Audio, is free for commercial use only below US$10 million in yearly revenue.

The limits that matter here are technical. Nemotron 3.5 Transcribe has 68 GB of BF16 weights and works on finished recordings only, while Whistle ships as a 16.9 MB file for seven languages. Voxtral Realtime lets you trade delay for accuracy, and its card shows an average error rate of 12.60% at 160 ms against 8.72% at 480 ms. Every accuracy figure is the developer’s own, so test on your audio.

Text-to-speech: several kinds of catch

The fourteen text-to-speech models fall into several groups, and only six are plain Apache 2.0:

Model Weights license Commercial use The catch
Qwen3-TTS Apache 2.0 Allowed Ten languages
VoxCPM2 Apache 2.0 Allowed Voice design varies between runs
Kokoro Apache 2.0 Allowed No voice cloning; one independent developer
Fun-CosyVoice3 Apache 2.0 Allowed Only the 0.5B model is released
Phonon 31M Apache 2.0 Allowed English only
ZONOS3-Mini Apache 2.0 Allowed Text normalization for nine languages only
Pocket TTS CC BY 4.0 on the card Allowed with credit The launch post says MIT; gated download
Supertonic 3 OpenRAIL-M Allowed, with banned uses Archived; the restrictions travel with every copy
OmniVoice CC BY-NC (code Apache 2.0) Not allowed Non-commercial weights
Voxtral TTS CC BY-NC Not allowed without another license Non-commercial weights
LFM3-Audio LFM Open License Only below US$10M yearly revenue Revenue cap
Drama 3 Fish Audio Community License Only below US$500K yearly revenue No hosted voice service at any size
Chatterbox MIT Allowed Retired in August 2026
VibeVoice TTS MIT Cards limit intended use to research Microsoft removed the TTS code in September 2025

A project can look permissive on GitHub and still be closed to products. OmniVoice’s code is Apache 2.0, but its weights are CC BY-NC, so about 3.8 million downloads in the 30 days to 11 November 2026 say nothing about whether you can ship it. Mistral’s card ties the CC BY-NC license of Voxtral TTS to its reference voices, and Phonon splits inside one family: the English 31M model is Apache 2.0, but the multilingual 40M models are CC BY-NC.

Revenue caps come in very different sizes. LFM3-Audio is free below US$10 million a year, but Drama 3 stops at US$500,000, and its license never covers a hosted text-to-speech or voice-cloning service, so a voice API built on it needs a separate deal at any size.

Maintenance is a second filter. Resemble AI retired Chatterbox and archived the repository in August 2026, and Supertone archived Supertonic 3, so neither will be updated. VibeVoice’s cards limit use to research, and Microsoft pulled the TTS code from GitHub in September 2025 after misuse.

A permissive license grants rights over the software, not over a person’s voice. 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 too: OpenBMB forbids impersonation, fraud and disinformation with VoxCPM2 and asks you to label AI-generated audio, the OmniVoice authors prohibit unauthorized cloning, and Fish Audio’s license bans imitation without consent. Chatterbox adds a PerTh watermark to its output by default.

What to do

  • Read the license of the weights, not only the code.
  • Check whether the project is still maintained.
  • For a commercial product, start with Qwen3-TTS, VoxCPM2, Kokoro, Fun-CosyVoice3 or the new ZONOS3-Mini, and test your language.
  • Get consent from every speaker you clone, and label synthetic audio.

Our guides how to choose a text-to-speech model and how to choose a speech-to-text model turn this into a shortlist, and OmniVoice vs VoxCPM2 shows the language-versus-license trade-off. For models that run without a GPU, read tiny text-to-speech models run on a CPU. The glossary explains a non-commercial license.

Sources

Models in this story

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
Kokoro hexgrad 82M – Apache 2.0 (Permissive) 0.38 GB Laptop or phone
Fun-CosyVoice 3 Tongyi Lab Speech Team 500M – Apache 2.0 (Permissive) 6.21 GB Laptop or phone
Phonon 31M Gradium 31M – Apache 2.0 (Permissive) 0.0747 GB Laptop or phone
ZONOS3-Mini Zyphra 800M 100M active – Apache 2.0 (Permissive) 10 GB Laptop or phone
Pocket TTS Kyutai 18M – 100M – CC BY 4.0 (Permissive) 0.127 GB Laptop or phone
Supertonic 3 Supertone 99M – OpenRAIL-M (Conditional) 0.46 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
Chatterbox Resemble AI 500M – MIT (Permissive) 3.69 GB Laptop or phone
LFM3-Audio-2B Liquid AI 2B 32K LFM Open 1.0 (Conditional) 4.6 GB Laptop or phone
Drama 3 Fish Audio 35.9B 3B active – Fish Audio Community 1.0 (Conditional) 25 GB Gaming GPU
Whisper large-v3-turbo OpenAI 810M – 1.54B – MIT (Permissive) 6 GB Laptop or phone
VibeVoice-ASR Microsoft 2.81B – 8.67B 64K MIT (Permissive) 6.47 GB Laptop or phone
Qwen3-ASR Alibaba Qwen 918M – 2.35B – Apache 2.0 (Permissive) 2.12 GB Laptop or phone
Voxtral Realtime Mistral AI 4.43B 128K Apache 2.0 (Permissive) 10 GB Laptop or phone
Cohere Transcribe Cohere 2.07B – Apache 2.0 (Permissive) 4.75 GB Laptop or phone
Granite Speech 5.0 IBM 470M – 2.31B – Apache 2.0 (Permissive) 1.09 GB Laptop or phone
Whistle Cactus Compute 55M – Apache 2.0 (Permissive) 0.253 GB Laptop or phone
Nemotron 3.5 Transcribe NVIDIA 34B 4.3B active 64K OpenMDW-1.1 (Permissive) 25 GB Gaming GPU
Nemotron 3.5 ASR NVIDIA 638M – OpenMDW-1.1 (Permissive) 2.93 GB Laptop or phone
Parakeet TDT 0.6B v3 NVIDIA 627M – CC BY 4.0 (Permissive) 2.89 GB Laptop or phone