Open speech models now run from 80 million to 35 billion parameters

Open text-to-speech has no typical size. The cards on Shipped run from 80 million parameters to 35 billion, and size tracks the job.

The small end is built to run anywhere. Soprano-80M is English only, has no voice cloning and was trained on about 1,000 hours of audio 1. Its card points to a newer Soprano 1.1. Supertonic 3, at about 99M, ships as ONNX, runs locally with no GPU and covers 31 languages 2. Its open package has fixed preset voices only. Chatterbox-Turbo, at 350M, targets voice agents; its decoder was distilled from ten steps to one 3. MagpieTTS covers 12 languages with five preset voices, and NVIDIA removed zero-shot cloning "for security reasons" 4.

The middle is where cloning and voice design arrive. OmniVoice, at 0.6B, is tagged for 646 languages 5. VoxCPM2, at 2B, builds a voice from a text description and outputs 48 kHz audio 6. Dia2, also 2B, writes English dialogue and can start speaking before the full text arrives 7.

The large end takes on length and scale. MOSS-TTS, at 8B, describes long-form generation of up to an hour in one run 8. Voxtral 4B speaks nine languages with 20 preset voices 9. Drama 3 is a mixture-of-experts model with 35B parameters and 3B active. It takes speaker tags for up to 16 characters in one script 10. The card reports about 89 ms to first audio on an H200, and gives no memory requirement.

Size does not settle the licence. Chatterbox-Turbo is MIT, and VoxCPM2 and MOSS-TTS are Apache 2.0. Voxtral's weights are CC BY-NC 4.0. OmniVoice's code is Apache 2.0 but its weights are non-commercial, and at 0.6B it is far from the largest. Drama 3's card names two different licences, so check the file before use.

Most of the numbers here are the makers' own, and several cards give none. Match the model to the job, then read the licence.