Release of Fish Audio Drama 3 Model

Fish Audio has released Drama 3, a new open-weights text-to-speech (TTS) model utilizing a Mixture-of-Experts (MoE) architecture with 35 billion parameters and 3 experts (35B3A). According to reports on Reddit, the model is designed to prioritize dramatic vocal expression, though some early user testing suggests that its voice cloning fidelity may be less precise than competitors, with results described as sounding like a “cousin” of the reference voice rather than an exact replica.

Technical Specifications and Hardware Requirements

The Drama 3 model is characterized by its significant computational scale. Users have noted that the full model requires approximately 48GB of RAM to operate effectively. However, 4-bit quantized versions are available, which reduce the requirement to approximately 24GB of RAM.

In comparative discussions regarding hardware limitations, users have noted that running high-fidelity TTS models alongside large language models (LLMs) and retrieval-augmented generation (RAG) systems can lead to significant VRAM bottlenecks. For instance, while smaller models like Fish Speech 1.4 may fit within 8GB of VRAM, newer iterations such as Fish Audio S2 or the Drama 3 series require significantly higher memory overhead, often exceeding the capacity of consumer-grade hardware like the RTX 3060 Ti.

Functional Capabilities and Features

The Drama 3 model includes several advanced linguistic and emotive features:

  • Multi-speaker Support: The model can handle up to 16 different speakers within a single long transcript.
  • Emotive Tagging: It supports standard utterance tags such as [laughs] and [sad].
  • Incremental Revision: The model allows for revision instructions, such as “a bit sadder,” to adjust the emotional tone of a voice line incrementally.
  • Multilingual Support: The model supports multilingual voice cloning.

Market Context and Competitive Landscape

The release of Drama 3 enters a highly competitive TTS market. Unverified user comparisons on Reddit suggest that while Drama 3 offers high-quality dramatic expression, other models currently hold different advantages:

  • Breezeblue-tts-2: Reported by some users to be a high-performing 2-billion parameter model, though it is limited to English and Chinese.
  • Moss TTS 1.5: An 8-billion parameter model noted for its variety and support for 31 languages.
  • Higgs Audio V3: Cited for superior multilingual pronunciation, though users noted its speaker similarity and naturalness were less consistent.
  • Omnivoice: Described as a smaller, versatile all-rounder with high voice-cloning accuracy.

While the Drama 3 model offers advanced emotive control, some users have expressed concerns regarding its licensing terms and the substantial hardware resources required for local deployment.