Fish Audio Drama 3

Fish Audio Drama 3 shifts the benchmark for open-weights speech generation by approaching text-to-speech as a directed theatrical performance rather than a flat transcription readout. Developed by Fish Audio, Drama 3 is an open text-to-speech architecture designed specifically for complex dialogue, multi-character scenes, and long-form narrative production. While traditional open-source voice generators struggle with emotional range and multi-voice consistency, Drama 3 uses a 35B-parameter mixture-of-experts (MoE) text backbone paired with an advanced codebook audio decoder to deliver nuanced 44.1 kHz broadcast-grade speech.

The defining technical advantage of the model is its operational efficiency. By activating only 3B parameters per token out of its total 35B parameter pool, Drama 3 combines the broad linguistic modeling and contextual control of a large model with the rapid generation speeds expected of a lightweight 3B engine. This design allows it to follow conversational subtext, execute natural-language direction tags, and synthesize multi-speaker dialogue scripts without accumulating audio drift.

Distributed under the Fish Audio Community License v1.0, the weights are freely accessible for commercial deployment by organizations with annual revenues under US$500,000. For engineers building automated audio production pipelines, interactive game agents, or local voice workflows, Drama 3 offers a compelling self-hosted alternative to proprietary speech APIs like ElevenLabs.

Drama 3 abandons standard dense TTS designs in favor of an asymmetric architecture that separates broad contextual comprehension from high-speed acoustic generation.

• Sparse Mixture-of-Experts Backbone: The core text model uses a 35B MoE framework that routes each token through specialized sub-networks, activating only 3B parameters per forward pass. This Fish Audio Drama 3 MoE efficiency yields high inference throughput while preserving the 35B backbone's ability to interpret nuanced script context.
• Hybrid Attention: The text transformer combines full multi-head attention with linear attention mechanisms. This hybrid approach keeps memory consumption stable during long-form synthesis and enables prompt-based multi-turn performance revisions.
• Codebook-MTP Audio Decoder: Rather than using a standalone diffusion pipeline, Drama 3 features a codebook multi-token-prediction (MTP) decoder. It predicts 10 acoustic codebooks causally across the entire sequence in dedicated transformer layers, directly producing high-fidelity 44.1 kHz waveforms.
• Mixed-Precision Weight Distribution: The model distributes weights natively using FP8 precision for the MoE expert modules, while attention layers and shared trunk modules remain in BF16 to maintain numerical stability during voice synthesis.

Because the model routes text inputs through an MoE backbone before acoustic decoding, it processes complex prompt structures, such as stage directions and speaker assignment tags, as native conditioning signals rather than external metadata.

Drama 3 is built for end-to-end production environments that require fine control over expressive pacing and vocal performance.

The model interprets plain-language direction tags placed directly inside dialogue text. Markers such as [whispering], [sighs], [forcing a smile], or [chuckling] dynamically alter pitch, breath support, and prosody right where they appear. In addition, Drama 3 supports speech infilling, allowing practitioners to edit specific phrases in an existing audio file by re-generating localized sections without re-rendering the entire take.

Drama 3 supports up to 16 distinct voices inside a single prompt using speaker tokens ranging from <|speaker:0|> to <|speaker:15|>. This makes it ideal for scripted scene generation, radio plays, and video game dialogue trees where multiple characters interact seamlessly without requiring separate audio stitching passes.

Using reference audio clips as short as three to five seconds, Drama 3 mirrors speaker timbre, cadence, and room acoustics. The model handles voice prompts across dozens of languages, including English, Chinese, Japanese, Korean, Spanish, French, and German, maintaining voice identity across multilingual transitions.

• Audiobook and Audio Drama Narration: Chapter-level rendering with automatic text chunking and dynamic pacing for distinctive character voices.
• Dynamic Video Game NPCs: Low-latency conversational agents that adjust emotional tone based on in-game context tags.
• Automated Localization and Dubbing: Infilling and scene dubbing pipelines that match the pacing of original video tracks.

Deploying Drama 3 on local workstations requires planning around parameter capacity and memory bandwidth. Even though active compute is low at 3B parameters, all 35B parameters must reside in memory during generation.

The native release uses FP8 experts alongside BF16 core layers, keeping base storage compact for a model of its class. The following table details the Fish Audio Drama 3 VRAM requirements across different quantization setups:

The best quantization for Fish Audio Drama 3 when running on a single prosumer card is Q4_K_M. This configuration preserves vocal timbre, acoustic clarity, and emotion tag responsiveness while reducing the overall memory footprint to fit within accessible hardware tiers.

To run Fish Audio Drama 3 locally at interactive speeds, consider these hardware configurations:

• Best GPU for Fish Audio Drama 3 Workstations: A dual-GPU setup featuring two NVIDIA RTX 3090 (24GB) or two RTX 4090 (24GB) cards linked via high-speed PCIe offers the best price-to-performance ratio. This setup provides 48GB of aggregate VRAM, running the native FP8/BF16 weights without offloading.
• How to Run a 35B Model on Consumer GPU Setups: If you only have a single 24GB graphics card, such as an RTX 4090 or RTX 3090, use a 4-bit quantized build (such as Q4_K_M or an AWQ/EXL2 equivalent). This fits the base weights and leaves sufficient VRAM for the codebook-MTP decoder layers and KV cache during short to medium generation runs.
• Unified Memory Systems: Apple Silicon machines (such as an M3 Max or M4 Max equipped with 64GB or 128GB of unified memory) run the model comfortably, making them practical development platforms for audio generation.

For maximum throughput, Drama 3 runs natively on top of the SGLang-Omni engine, which provides optimized kernels for multi-token codebook prediction and sparse MoE routing.

To launch a local API instance via SGLang-Omni:

1git clone https://github.com/sgl-project/sglang-omni.git && cd sglang-omni
2uv pip install --prerelease=allow -v -e ".[fishaudio]"
3sgl-omni serve --model-path fishaudio/drama-3 --port 8000

On enterprise hardware like an NVIDIA H200, Drama 3 reaches inference speeds up to 271 audio frames per second. On dual RTX 4090 setups using optimized kernels, expect real-time audio generation factors between 3x and 7x depending on prompt length and audio context size. In standard token metrics, Fish Audio Drama 3 tokens per second performance scales well ahead of dense 35B models due to the low 3B active compute footprint.

While Ollama is widely used for text LLMs, multi-modal TTS models requiring causal acoustic decoders and 44.1 kHz vocoders are best deployed using SGLang-Omni, vLLM-Omni, or the official fish-speech runtime.

Evaluating Fish Audio Drama 3 vs Fish Audio S2-Pro highlights the architectural evolution between generations. S2-Pro uses a compact 4B dense backbone. While S2-Pro requires less memory (fitting comfortably into a 16GB VRAM envelope) and offers rapid baseline synthesis, Drama 3 provides significantly higher emotional range, superior pronunciation accuracy, and native support for multi-turn direction. S2-Pro reads scripts well, but Drama 3 delivers an expressive performance.

Comparing Fish Audio Drama 3 vs models like CosyVoice 2 or F5-TTS shows a clear architectural divergence. F5-TTS relies on diffusion transformers, which produce clear audio but struggle with fast real-time streaming and long multi-character scenes due to high iterative compute demands. Drama 3 uses autoregressive codebook prediction alongside sparse MoE routing, making it faster for streaming audio and capable of handling complex scripts with up to 16 distinct speakers.

For engineers seeking a powerful local AI model with 35B parameters in 2026, Fish Audio Drama 3 provides an optimal combination of expressive direction, multi-voice management, and computational efficiency for dedicated production setups.