Fish Audio open-sources Drama 3 for 16-speaker voice acting

Fish Audio has officially open-sourced Drama 3, a next-generation text-to-speech model built specifically for multi-character scenes and long-form audio storytelling.

While most open TTS models cap out at a few billion parameters and struggle beyond solo narration, Drama 3 lets you juggle up to 16 distinct speaker voices within a single passage of text using native speaker tags. It is tailor-made for radio dramas, podcasts, and fiction narration.

What it does

Beyond multi-speaker dialogue, Drama 3 supports fine-grained emotional control using inline text tags.

You can drop action cues like sighs or laughs directly into the sentence, and the voice will naturally weave in the emotion instead of sounding like a flat robotic reader. It also supports instant voice cloning from a brief reference audio clip.

The real game-changer for production workflows is continuation-style audio editing. If the model mispronounces a single word, you do not have to regenerate the whole take from scratch. You simply send the previous acoustic tokens and the edited text; Drama 3 will stitch the new word seamlessly into the existing flow and cadence.

For longer manuscripts, the system automatically chunks text and supports per-chunk streaming so listeners do not have to wait for the whole chapter to finish rendering.

Specs and performance

Under the hood, Drama 3 uses a 35B MoE text backbone with hybrid attention, but only 3B parameters are active per token. The experts run in FP8 while the remaining weights use bf16, paired with a causal 10-codebook audio decoder.

Because only 3B parameters activate at any time, users have reported running it locally with as little as 8GB of VRAM, despite the massive 35B total size.

For production serving, it launched with native day-one support in SGLang-Omni and vLLM-Omni. On a single NVIDIA H200 GPU, single-stream time-to-first-audio lands under 100 milliseconds. At 16 concurrent streams, it still generates audio 2.7 times faster than playback speed, maintaining a mean real-time factor around 0.8 even at 64 concurrency.

When scaling to 256 concurrent streams, offloading the vocoder to a second GPU yields about 72% higher throughput and cuts median TTFA by 37%.

Drama 3 is currently ranked number one among open models on Artificial Analysis, and the weights are freely available on Hugging Face under the Fish Audio Community License.

Why it matters

For podcasters, indie game devs, and audiobook creators, Drama 3 eliminates the two biggest headaches in synthetic speech: managing multiple voices in one script and re-rendering whole files just to patch a typo.

Opting for a 35B-A3B MoE architecture is a brilliant move. It leverages the scale needed for nuanced acting and seamless continuation without shutting out local hobbyists on consumer hardware. If you are building automated audio drama pipelines, this is definitely worth deploying right now.