Fish Audio's Drama 3 Tops Open-Weight Voice Rankings at $15 per Million Characters

Fish Audio has released Drama 3, a 35-billion-parameter text-to-speech model. At launch, it ranked first among open-weight systems on both Artificial Analysis voice arenas and first overall on the Pronunciation Robustness benchmark. Its downloadable checkpoint supports 83 languages, voice cloning, emotion changes within a sentence, and scenes with as many as 16 speakers in one generation.

Fish Audio describes Drama 3 as the first downloadable checkpoint that interprets bracketed stage directions like instructions to a voice actor. For developers, the release combines expressive prompt controls with self-hosting, subject to a 48 GB GPU requirement and a commercial license tied to annual revenue.

Drama 3 leads pronunciation tests

Artificial Analysis derives its arena Elo scores from pairwise listening tests. The Provider Voice Arena uses each provider’s default voice, while the Controlled Voice Arena gives every model the same cloned voice to reduce the influence of voice selection.

On Provider Voice, Drama 3 finished 38 Elo points ahead of Drama 3 Preview and 50 points ahead of BreezeBlue’s Breeze TTS 2. On Controlled Voice, it narrowly led Ant Group’s Ming-3.0-tts, which scored 1,068.

The pronunciation benchmark tests whether models expand abbreviations and units, choose context-dependent pronunciations, preserve codes and serial numbers, and speak standalone terms correctly.

Standalone terms remain the weakest category for the release model. Drama 3 Preview scored 95.9% there, four percentage points higher. Overall generation speed reached 66.6 characters per second, between Murf Speech Gen 2 at 63.0 and Eleven v4 Conversational at 66.8.

Thirty-five billion parameters, three billion active

Drama 3 uses a sparse mixture-of-experts architecture with 35 billion total parameters and 3 billion active for each token. The model routes tokens through a subset of specialized parameter groups, reducing computation per step, although self-hosters must still store the full checkpoint in GPU memory.

The network combines full attention with linear attention. One of every four attention layers uses full attention, while the linear layers use Gated DeltaNet to carry information across the sequence more efficiently.

Audio generation runs through a 10-codebook Codebook-MTP decoder. Codec codebooks represent separate layers of audio information as discrete tokens, and multi-token prediction allows the model to emit several audio tokens per step. The decoder processes those codebooks causally across the sequence.

The mixture-of-experts weights ship in 8-bit FP8 format, with the remaining weights stored in 16-bit bfloat16. Speaker control is part of the token vocabulary, using tags from <|speaker:0|> through <|speaker:15|>. Long inputs are split automatically and can stream one chunk at a time.

Hosted pricing undercuts higher-ranked APIs

Fish Audio prices the hosted Drama 3 API at $15 per million characters. Each model ranked above it in the Provider Voice Arena costs more under the published prices cited for the comparison.

Self-hosting starts at 48 GB of VRAM

Local deployment requires an NVIDIA GPU with at least 48 GB of VRAM when using FP8 weights. The current release does not support CPU-only execution, Apple MPS, or AMD ROCm.

Fish Audio reports the following serving performance on a single NVIDIA H200:

Time to first audio, or TTFA, measures the delay before playback can begin. Real-time factor, or RTF, divides generation time by audio duration, so a value below 1.0 indicates generation faster than playback. The reported median TTFA is 88.9 milliseconds for one stream, and mean RTF remains below 1.0 at 64 concurrent streams.

Moving the vocoder to a second GPU increased throughput at 256 concurrent streams by roughly 72% and reduced median TTFA by 37% in Fish Audio’s tests.

Developers can serve the checkpoint through SGLang-Omni or vLLM-Omni. A minimal SGLang-Omni command is:

sgl-omni serve --model-path fishaudio/drama-3 --port 8000

Commercial rights stop at a revenue threshold

The checkpoint is licensed under the Fish Audio Community License. Its commercial restrictions place the release in the open-weight category, with permission to inspect and self-host the model under the published terms.

Teams building internal products or applications below the revenue threshold can use the checkpoint without a license fee. TTS vendors, API resellers, and companies outside that threshold need commercial terms from Fish Audio.

Stage directions live inside the prompt

Drama 3 interprets square-bracketed directions embedded directly in the text, allowing a prompt to change delivery, emotion, or nonverbal expression within a line. A request body for the /v1/tts endpoint can include both those directions and reference audio for voice cloning:

{
  "text": "[sighs] Fine. If we miss it, it's on you. [laughs] Well, maybe that's fair.",
  "references": [
    {
      "audio_path": "...reference.wav",
      "text": "..."
    }
  ]
}

The editing mode accepts a previous take as context and regenerates changed text while conditioning on the earlier audio tokens. That approach helps preserve prosody when replacing a word or short phrase, such as changing “Thursday” to “Friday” without regenerating an unrelated performance from scratch.

Where Drama 3 fits

Drama 3 targets scripted dialogue, long-form narration, dubbing, audiobooks, and game production where delivery needs to be controlled from text. Its strongest differentiators are pronunciation, multi-speaker generation, inline performance directions, and audio-conditioned edits.

Checks before production

• Quality: Test names, codes, abbreviations, and target languages against real application inputs.
• Capacity: Benchmark expected concurrency on the intended GPU rather than treating H200 vendor results as universal.
• Long-form output: Evaluate prosody and transitions across automatically generated chunks.
• Licensing: Confirm revenue, affiliate, and resale terms before deployment.
• Voice rights: Obtain permission for reference recordings and cloned voices.

The hosted API offers the lowest operational burden at $15 per million characters. Self-hosting provides control over data and serving infrastructure, with a 48 GB NVIDIA GPU and the community license setting the practical entry requirements.