Open-weight TTS

The segment of text-to-speech models whose weights are downloadable and self-hostable — the speech analog of llm-providers-wiki's open-weight-models. Defined by two facts in 2026:

1. It trails the closed frontier, and by more than this page used to say

No open-weight model cracks the tts-arena-leaderboard top tier (all closed/API-only). The open leader is fish-audio-s2-pro (Elo 1123, 5B), and the 2026-08-03 board puts a number on "below" that this page previously lacked: it ranks 20th overall, with kokoro 48th, against a closed leader at 1229.

Updated 2026-10-07. One open model now nearly cracks the top tier: Breeze TTS 2 (BreezeBlue) ranks 11th at 1216. Under it the tail kept sinking: fish-audio-s2-pro is 33rd and kokoro 54th of 95 (tts-arena-leaderboard). The first sentence of this section held until this snapshot and is now true only by one place.

This section used to read "but not by much." That was wrong, and it got wronger. Over the summer the gap to the leader widened from 99 to 106 Elo — the top of the board rose while both open models drifted down (fish-audio-s2-pro 1128 → 1123, kokoro 1064 → 1055). Reporting only the Elo delta and never the rank made a twentieth-place finish sound like a near miss. The proprietary-premium-vs-open-wedge dynamic llm-providers-wiki tracks for text does hold for voice, but the wedge is thinner here than the earlier phrasing implied.

2. License is a first-class axis, not a footnote

The open field splits on terms:

• Permissive (Apache-2.0 / MIT): kokoro, orpheus, sesame-csm, Chatterbox, Dia, Higgs Audio V2 — freely commercial-usable open-source-tts-models. Coqui's XTTS belongs here too: a peer-reviewed, publicly released zero-shot cloner across 16 languages (xtts-paper) — the open field's evidence that even voice-cloning ships without a research-license gate.
• Research-only / paid-commercial: fish-audio-s2-pro (the highest-Elo open model is not freely commercial) and misotts's "modified MIT."
• Revenue-capped (added 2026-08-25): audio8-tts's Audio8 Community License v1.0 — free for non-commercial use and free commercially below US$2,000,000 in consolidated annual revenue, with a separate written licence required above it. So the practical "best open model" depends on whether you can use it commercially, not just its score.

How lopsided the split actually is (2026-08-03). A 36-model census of the TTS field (free-voice-clone-list, April 2026 snapshot) counts 27 Apache-2.0, 4 MIT, 1 MIT/Apache-2.0, 2 under Liquid's LFM licence, 1 Research License (fish-audio-s2-pro) and 1 OpenRAIL-M (supertonic-2). The restrictive corner is not a wing of the field, it is two models — but one of them is the quality leader, which is precisely why counting licences cannot settle the ceiling question in synthesis. Permissive availability is near-total; permissive at the top is still unproven on this branch.

The two-way split is now three (2026-08-25). audio8's revenue cap does not sit on either pole. It is not a research licence — a startup can ship a paid product on it — and it is not Apache-2.0 either, because the permission expires when the company grows. What it prices is not the use but the user, which is a different axis from the one this page was counting on, and it means a licence census sorted into "permissive / restrictive" now loses information. The shape is borrowed from source-available software licensing rather than from anything else in this field, and so far it is one instance.

A second thing that licence does, which none of the permissive ones do: it carries behavioural terms — consent before impersonation, disclosure of synthetic audio. synthesis records that the free-voice-clone-list census of 53 models contained no such sentence anywhere, and that the lone OpenRAIL-M entry belonged to supertonic-2, which cannot clone a voice. Audio8 is the second instance, and it does clone.

The shape of the field

Efficiency-first (kittentts 15M and supertonic-2 66M, then kokoro 82M — the last two without voice-cloning, then audio8-tts at ~170M with zero-shot cloning at 44.1 kHz and, by its own timings, above real time on a CPU) → all-rounder (Chatterbox 0.5B) → expressive/streaming (orpheus) → conversational (sesame-csm) → emotive (misotts) → top-quality but restricted (fish-audio-s2-pro) → and, added 2026-09-05, a fourth axis nobody else on this list competes on: language count, where omnivoice claims 646 at 0.8B. Read that axis differently from the others. Size, expressiveness and licence are properties a user can verify in an afternoon; coverage is a claim about 646 populations, and OmniVoice publishes benchmark numbers for at most 102 of them. Many are Llama-based — a bridge to the text-LLM market (gemini cross-wiki).

Related

text-to-speech · tts-benchmarks · tts-arena-leaderboard · kokoro · kittentts · supertonic-2 · fish-audio-s2-pro · open-source-tts-models · free-voice-clone-list · xtts-paper · voice-cloning · open-weight-models · audio8-tts · audio8