Fish Audio Releases Drama 3 With Open Voice Weights
The 35B speech model supports plain-language performance direction, but its commercial license, temporary API pricing and invite-only editing deserve a closer look.
Listen
AI narration
11:21
0:00 / 11:21
AI SummaryGenerated from this article
Fish Audio released Drama 3, a 35-billion-parameter speech model that accepts plain-language performance directions embedded in scripts, allowing users to specify emotions and delivery changes within dialogue. The model is available through web, mobile apps, and API, supporting 83 languages and up to 16 speakers.
The open-weight release carries commercial restrictions: users earning over $500,000 annually or offering it as a service need separate licenses. Fish Audio's API launches at $10 per million characters through November 12, 2026, then rises to $50 per million. Interactive editing remains invite-only. An independent pronunciation benchmark shows Drama 3 scoring 93.2% versus 88.1% for the prior version, though this measures technical accuracy rather than emotional performance.
Fish Audio’s Drama 3 accepts acting directions inside the script it speaks. A line can start with restrained anger, fall to a whisper and change delivery again halfway through, according to the company. Downloadable weights and a hosted API give developers a choice between running the model themselves and buying access.
In its Drama 3 announcement, dated October 13, 2026, Fish Audio says the model is available through its web text-to-speech product, iOS and Android apps, and API. It supports 83 languages, while its downloadable model card lists multi-speaker generation for up to 16 speakers.
Access comes with qualifications. Commercial self-hosting depends on revenue and deployment type, the API’s launch discount expires, and interactive editing remains in small invite-only testing. The independent pronunciation result also answers a narrower question than whether Drama 3 consistently delivers a convincing performance.
Performance Direction Goes Inside the Script
Drama 3’s control interface uses natural-language instructions in square brackets. Users can choose a broad emotion for a passage or insert new directions wherever the delivery should change.
Fish Audio’s examples include instructions such as “barely holding it together, forcing a smile,” followed later by a shaky breath, a small laugh and whispering. These company-selected demonstrations show the intended workflow: write the dialogue and direct its delivery in the same text. They do not establish that every combination of instructions will work reliably.
Paired tags target particular words. An emphasis tag, for example, applies to the text between its opening and closing markers. In the web product, users can select words and choose a tag through the interface.
Fish Audio says directions need not match the language of the spoken line. Its examples use Chinese, Spanish and Japanese instructions to direct English dialogue. That could make multilingual production easier, although support for 83 languages should not be read as proof of equal pronunciation quality or emotional control across all of them.
The Drama 3 model card describes support for up to 16 speakers, making character transitions another part of the generation task for scripted dialogue. It does not establish how consistently the model preserves every character’s voice over a long production.
API users can put the same scripts and tags in the request’s text field, with the model header set to drama-3, according to Fish Audio. Performance direction is part of the generation input and is available beyond the consumer app.
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
A 35B Backbone Does Not Mean 35B Parameters Run Every Step
Drama 3 uses a 35-billion-parameter mixture-of-experts backbone, with approximately 3 billion parameters active per token. Fish Audio says it initializes that backbone from the language model of Qwen3.5-35B-A3B.
In a mixture-of-experts model, routing selects a subset of specialized components for each processing step. Drama 3 routes to eight of 256 experts alongside one shared expert, according to the company. Active parameter count helps describe computation, but it does not replace the full model’s storage and memory requirements when planning deployment.
The architecture mixes two kinds of sequence processing. Fish Audio describes 40 layers arranged in repeating blocks of three Gated DeltaNet layers and one gated full-attention layer. The DeltaNet layers carry history in a fixed-size state, while only one layer in four maintains a key-value cache.
Drama 3 also replaces S2’s per-frame decoder with a codebook-MTP decoder. It attends over the preceding sequence, including detailed information from reference audio, while predicting the codes that become the waveform.
These architectural disclosures do not independently demonstrate better speech. Fish Audio also reports a training change: a learned human-preference reward replaces separate rule-based rewards. The company’s explanation is that requested laughter or whispering should not be penalized merely because it makes speech harder to transcribe.
For throughput, Fish Audio reports 271 audio frames per second for a single stream on one NVIDIA H200, compared with 177 for S2. This vendor-reported result uses a particular hardware setup; it does not establish consumer-GPU requirements, end-to-end API latency or performance under concurrent production traffic.
The Independent Benchmark Tests Pronunciation, Not Acting
Fish Audio cites a 93.2% score for Drama 3 on Artificial Analysis’s Pronunciation Robustness benchmark, compared with 88.1% for S2.1 Pro. Its announcement describes Drama 3 as the benchmark leader.
The benchmark addresses pronunciation, separately from the company’s claims about stronger emotion and more precise performance direction. Artificial Analysis’s Pronunciation Robustness methodology tests difficult spans within sentences. Those spans include context-dependent readings, numbers and notation, exact sequences such as codes and email addresses, and standalone brand, place and technical names.
Every model reads the same sentences with a consistent voice. Reviewers hear clips in randomized order without model names and judge whether highlighted spans were pronounced naturally and correctly against predefined accepted readings. Published results require at least three reviewers for at least 95% of samples.
The score is the share of eligible judgments answered “Yes,” excluding “I could not tell” responses. It is not the percentage of complete scripts delivered without errors.
On Fish Audio’s published figures, Drama 3 improves by 5.1 percentage points over S2.1 Pro. The company also reports an increase from 79.4% to 86.3% for exact sequences such as paths and email addresses. That is useful evidence for narration involving technical material, though mistakes remain in a category where one wrong character can change the meaning.
How convincingly Drama 3 cries, whether it places every requested laugh correctly, and whether emotional delivery remains consistent across 83 languages all need evaluation on representative scripts. The independent test does not establish those capabilities. Its strong pronunciation result supports one part of production readiness.
Open Weights Have Revenue and Service Restrictions
Drama 3 is an open-weight release, not an unrestricted open-source release.
Fish Audio describes the governing terms as the Fish Audio Community License v1.0. Research and non-commercial use are permitted, along with commercial use where the user and its affiliates have less than US$500,000 in annual revenue.
Two restrictions materially affect deployment:
Commercial users at or above the revenue threshold need a separate license.
Offering Drama 3 to others as a text-to-speech, voice-cloning or inference service or API also requires a separate license.
The service restriction applies independently of the revenue threshold. A small company should not assume that being below the threshold allows it to launch a hosted Drama 3 inference service.
Because revenue includes affiliates, the threshold is broader than the income of an individual voice project. A low-revenue application within a larger business should not treat its own sales as the only relevant figure.
The license additionally prohibits cloning a real person’s voice without consent, according to Fish Audio. That condition still applies to downloadable weights.
Developers should inspect the current license distributed through the Hugging Face repository, instead of relying on a repository metadata label or the shorthand “free commercial use.” Fish Audio also says S2 remains under its earlier Research License, so Drama 3’s permissions should not be assumed to apply to older releases.
The $10 API Rate Is a Launch Offer
Fish Audio lists Drama 3’s standard API price at $50 per million characters, discounted to $10 per million characters through November 12, 2026. That is the advertised 80% launch discount.
The company also advertises 50% off new creator plans during the launch period. This offer is separate from API character pricing and should not be treated as a permanent reduction.
The temporary rate makes the hosted API a cheaper place to evaluate scripts before committing to deployment. Long-term budgets need to account for the listed standard rate, which is five times the promotional rate, as well as repeated generations needed to obtain an acceptable take. Character pricing alone does not determine the cost of a finished production.
Self-hosting has a different cost calculation. Eligible users may avoid Fish Audio’s API charges, but they still need infrastructure and an inference setup. Above the license threshold, or for a hosted service, a separate commercial agreement becomes part of that calculation.
Editing Is Still Invite-Only
Interactive editing and audio inpainting remained in small invite-only testing at launch. They should not be presented as generally available simply because Drama 3 itself has shipped.
For long-form work, generating a directed passage and replacing a flawed moment while preserving the surrounding recording are distinct capabilities. A production plan that depends on fixing individual words without rebuilding a larger passage should verify access before committing to that workflow.
Drama 3’s consequential contribution is the combination of script-level direction, downloadable weights and a production API. The purchasing decision still requires more than the most expressive demo: teams need to test pronunciation and delivery on their own material, budget at the standard API rate, and check whether their intended deployment qualifies under the license. Until editing access is confirmed, the released generation model is the product to evaluate and the workflow to plan around.
