Drama 3 with open weights beats ElevenLabs on pronunciation, according to benchmark
A 35-billion-parameter text-to-speech model with open weights is now ranked ahead of ElevenLabs on pronunciation robustness, according to figures from the benchmarking firm Artificial Analysis as reported by secondary sources. With a free license below half a million dollars in revenue and directing features aimed at production, the release shifts the comparison baseline for developers choosing a voice model — but large parts of the feature and business story still rest on the company's own claims.
The news: open weights at the top of a pronunciation benchmark
Fish Audio launched Drama 3 Preview on October 13 with full model weights published on Hugging Face. According to reporting in Startup Fortune, Artificial Analysis — the independent benchmarking firm behind the Speech Arena leaderboard — places Drama 3 Preview as number one among open-weights models on its Text to Speech leaderboard.
More striking is the second ranking: on the separate Pronunciation Robustness benchmark, Drama 3 is said to lead all models, open and closed, at 93.2 percent — ahead of both Eleven v4 and Eleven v4 Turbo from market leader ElevenLabs. That is unusual, because closed commercial voice models normally top such leaderboards. The figures are attributed to Artificial Analysis via Startup Fortune, which alone does not constitute primary documentation: neither the leaderboard itself, the methodology, nor the model card appears in the coverage. ElevenLabs has not commented.
Pronunciation robustness is a measurable dimension that matters in practice — names, foreign words and numbers pronounced correctly on the first try — and is thus one of the points where an open model can beat a closed alternative on something concrete, not just on price.
How the model is built
Drama 3 is a mixture-of-experts model with 35 billion parameters in total, but activates only 3 billion per token. It is this setup that, according to the reporting, allows the model to run fast enough for real-time use despite its size: only a fraction of the network is used per generation step, while total capacity is available across the inputs.
The model covers 83 languages and is, according to Startup Fortune, built specifically for multi-character scenes and long-form audio — work that previously required stitching together separate clips each with a single voice. Generating an entire scene with multiple speakers in one pass is therefore positioned as a core feature, not an add-on.
It should be emphasized that the detailed architecture specifications for Drama 3 have not been published. Unite.AI notes in its coverage that benchmark results from Fish Audio's earlier S2 model cannot be transferred to Drama 3, since the architecture is not sufficiently documented in the launch material.
The license: free below half a million dollars
The weights are released under the Fish Audio Community License v1.0. The license costs nothing until a company passes $500,000 in annual revenue. Above the threshold, and for hosted resale of the service, a paid license is required.
It is this combination — a top ranking on a reported benchmark and a revenue threshold instead of per-API pricing — that makes Drama 3 relevant to small players: a podcast studio, a game developer or an audiobook producer below the threshold can run the model themselves without license costs, and retain control of the inference stack.
The features: directing rather than generating
In the launch announcement, as reproduced by Unite.AI, Fish Audio co-founder and CEO Rissa Cao frames the release this way: "Voice AI has become very good at generating speech. The next step is directing a performance."
The features behind the rhetoric, as described from the preview page via Unite.AI:
- Natural-language directing: creators can control the performance with plain language instead of manipulating waveforms.
- Controls in the text: users can place direction cues, sounds and pauses in square brackets, or use paired tags around selected words — concrete mechanisms described from the preview page itself.
- Multi-speaker scenes in one pass: entire dialogue scenes are generated in a single run instead of clip by clip.
- Audio inpainting: a selected segment of synthesized speech can be regenerated without regenerating the performance around it.
The inpainting feature, however, is where the sources begin to diverge. Unite.AI treats single-word editing as a headline capability at launch. Startup Fortune, by contrast, states that single-word inpainting is still in invitation-based testing, with a public editing API planned. Its status is therefore unresolved in the coverage, and there is no independent evaluation of how reliably such single-word repairs work across voices, accents or difficult transitions — a point Unite.AI itself raises in its editorial assessment of the launch material.
The company behind it: seed capital and its own headline numbers
Fish Audio raised $52 million in a seed round in July, led by Coreline Ventures and Capital Today. The company grew out of Fish Speech, one of the most-starred open-source voice projects on GitHub, about a year ago.
The company itself reports $21 million in annual recurring revenue and more than 8 million users. These figures are Fish Audio's own claims and are not independently verified.
What is missing
Several parts of the story rest on secondary sources and the company's own numbers:
- Benchmark methodology. The methodology of the Pronunciation Robustness benchmark is not described in the coverage. Readers therefore do not know what the 93.2 percent is measured over, or how. The comparison with ElevenLabs rests on Artificial Analysis' figures via a single reporting chain, not on primary documentation.
- ElevenLabs' response. None of the coverage sources quote any reaction from ElevenLabs.
- Inpainting status. The disagreement between the sources over whether the feature is launched or in invitation testing is unresolved.
- Architecture. Model card and architecture details for Drama 3 have not been published in the coverage, limiting how much earlier results reporting from Fish Audio can be transferred.
These gaps are also the basis for follow-up: a primary-source check of the leaderboard and license terms, and independent evaluation of inpainting quality, would determine how much of the promise holds.
Why it matters
The release is interesting regardless of how the unresolved details fall out. If Artificial Analysis' rankings are accurate, it shows that open weights can lead on a measurable quality dimension against a closed market-leading product — not merely compete on price and control. The license model with a flat revenue threshold lowers the barrier for small players to try it, and the production-oriented features — directing, scenes in one pass, inpainting — point toward competition in voice models shifting from raw generation quality to the workflow around it. For now, however, it is largely the company's own claims that describe that workflow.