Fish Audio releases Drama 3 Preview, tops open-weights text-to-speech leaderboard

In its October 13, 2026 release, Artificial Analysis said Fish Audio's Drama 3 Preview is the top-ranked open-weight model on its three text-to-speech leaderboards. That sentence is easy to read as "leading across the board," but broken down, the three leaderboards give inconsistent positions: first in pronunciation robustness, 10th in Provider Voice, and 14th in Controlled Voice.

What truly sets it apart is pronunciation robustness. Drama 3 Preview scores 93.2%, above Eleven v4's 91.7% and Eleven v4 Turbo's 90.1%; on the same leaderboard, the next open-weight model, Mistral Voxtral TTS, scores only 78.1%, a 15-percentage-point gap. In other words, its lead "within the open-source camp" is far larger than its lead over the closed-source front-runners.

The subcategory distribution also supports this "specialist" judgment: it ranks first in Standalone Terms and Preserving Exact Sequences, and second in Expanding Shorthand. Artificial Analysis explains that the leaderboard is judged item by item by human reviewers across four categories of challenging text, against pre-agreed acceptable pronunciations—this is both the source of its score and the source of its subjectivity risk.

If you look only at pronunciation, Drama 3 Preview has already surpassed the Eleven v4 series; but switch to arenas closer to overall listening preference, and its position is much lower. In the Provider Voice Arena, its Elo is 1,243 (±16), based on 1,857 appearances, ranking 10th.

The significance of this score lies in the longitudinal comparison: the previous generation Fish Audio S2.1 Pro had an Elo of 1,142, and Drama 3 Preview improved by 101 points. Horizontally, it is slightly below Speechify Simba 3.2's 1,245 and above Breeze TTS 2's 1,221—the gaps with the names immediately before and after are both around the order of the margin of error, and do not constitute a decisive advantage.

Its performance on the Controlled Voice leaderboard is weaker still: Elo 1,058 (±15), 1,804 appearances, ranking 14th. From these data it can be inferred that Drama 3 Preview surpasses the Eleven v4 series in pronunciation robustness, but still trails some closed-source models in the two arena leaderboards; this is an inference drawn from the source data, not a direct conclusion of the source.

Drama 3 Preview is offered through the Fish Audio API at $15 per million characters, the same as S2.1 Pro. According to the prices given by Artificial Analysis, Eleven v4 is $80 per million characters and Eleven v4 Turbo is $40 per million characters.

From this it can be inferred that Drama 3 Preview's API unit price is about one-fifth of Eleven v4's and about one-third of Eleven v4 Turbo's. This price gap, combined with its lead in pronunciation robustness, forms the most direct competitive narrative of this release: not "cheaper, therefore worse," but more accurate on a quantifiable dimension while also cheaper.

What needs restraint is this: both the price and the leaderboard results come from the same source, and the price is an API-side quote, which does not equal self-deployment cost. The source does not give deployment-side metrics such as latency, throughput, or VRAM usage, so "cheap" can currently only rest at the unit-price level.

Drama 3 Preview's weights are offered on Hugging Face under the Fish Audio Community License; the model itself is a 35B-parameter mixture-of-experts architecture with 3B activated parameters, supporting 83 languages. The source says it targets multi-character scenarios and long-form content, and supports voice cloning and natural-language control—this is a source claim, not a verified capability boundary.

Several key pieces of information are still missing. The source does not explain the specific terms of the license, so the actual freedom of "open weights" for commercial use and redistribution cannot be judged; the source also does not disclose training data sources, data scale, or training method details; the actual performance and coverage quality of each of the 83 languages are likewise unexplained.

More noteworthy is the nature of the evaluation itself: all leaderboard data comes from Artificial Analysis's own leaderboards, and the source provides no third-party independent reproduction. Pronunciation robustness relies on human reviewers judging against pre-agreed acceptable pronunciations, leaving room for subjective judgment; the Elo of the two arenas also carries a margin of error of ±15 to ±16, which is insufficient to support strong conclusions when rankings are close.

Drama 3 Preview is a preview version. The source does not state when the official version will be released, nor whether the weights will change—this means the current leaderboard results, license, and price could all be reshuffled in the official version.

The source likewise does not provide direct comparison results with closed-source front-runners in long-form, multi-character scenarios. Considering that the model is positioned for multi-character and long-form use, this gap falls exactly in its claimed home turf, not in an edge case.

Therefore the more prudent reading is: Drama 3 Preview has delivered the best publicly available result in the open-weight camp on the quantifiable dimension of pronunciation robustness, and has entered the market at a clearly lower API unit price; but whether it has caught up with the closed-source front-runners in overall speech quality, the existing evidence is not yet sufficient to answer.