Drama 3
An open-source large speech generation model released by Fish Audio. It uses a MoE architecture with 35B total parameters and about 3B activated parameters, and supports director-style natural language control, word-level inpainting, multi-speaker dialogue generation and language switching.
Timeline (last 90 days)
Fish officially only states that Drama 3 runs well on the DGX Spark, but does not provide hard metrics such as real-time factor (RTF) or concurrency numbers.
The Fish team released a brand-new text-to-speech model, Drama 3, using an MoE architecture with 35B total parameters and 3B activated parameters.
The Fish team provides an online experience space for Drama 3 on Hugging Face (fishaudio/Drama-3), and the model weights are also open for download, supporting local operation.
The 4bit quantized version of Drama 3 performs very close to the original model.
Drama 3 runs well on the DGX Spark.
Drama 3 enters the TTS track with an MoE architecture, representing a new approach of using sparsely activated large models to capture the complex patterns of speech while controlling inference costs.
Some community users built a demo application for Drama 3 on the same day.
Drama 3 covers 83 languages.
Drama 3 accepts natural language instructions (plain-language direction) to control speech, rather than requiring professional tuning of voice tags or parameter panels.
Drama 3 is released with open weights, accompanied by a revenue-capped community license, and commercial use has a revenue threshold.
23 more timeline events
All Facts (20)
The 4bit quantized version of Drama 3 performs very close to the original model.
50%UnverifiedDrama 3 runs well on the DGX Spark.
50%UnverifiedDrama 3 enters the TTS track with an MoE architecture, representing a new approach of using sparsely activated large models to capture the complex patterns of speech while controlling inference costs.
50%UnverifiedFish officially only states that Drama 3 runs well on the DGX Spark, but does not provide hard metrics such as real-time factor (RTF) or concurrency numbers.
50%UnverifiedDrama 3 is released with open weights, accompanied by a revenue-capped community license, and commercial use has a revenue threshold.
50%UnverifiedDrama 3 covers 83 languages.
50%UnverifiedMost open-source speech models have parameter counts stuck in the range of a few hundred million to just over a billion, and open-weight TTS models at the 35B level remain rare.
50%UnverifiedDrama 3 accepts natural language instructions (plain-language direction) to control speech, rather than requiring professional tuning of voice tags or parameter panels.
50%UnverifiedDrama 3's API price is $10 per million characters, rising to $50 after November 12, 2025.
50%UnverifiedDrama 3's interactive editing feature is currently invitation-only and not yet launched.
50%UnverifiedDrama 3 officially claims support for 83 languages, but the cited benchmark only tests pronunciation robustness, not emotional expressiveness authenticity.
50%UnverifiedIn Drama 3 API calls, if the model header is set incorrectly, the request does not fail but silently falls back to the S2.1 Pro model.
50%UnverifiedThe naming of Drama 3 is easily misunderstood; it is actually the third generation of another product line, not a continuation of the S2 series.
50%UnverifiedDrama 3 has practical significance for long-form audio dramas, game voiceovers, and audiobook production, but limited significance for simple short-video voiceover needs.
50%UnverifiedRunning Drama 3 locally requires about 39GB of VRAM (a dual RTX 3090 configuration).
50%UnverifiedDrama 3 is a mixture-of-experts (MoE) model with 35B total parameters, activating only about 3B parameters per token.
50%UnverifiedDrama 3 scored 93.2% on Artificial Analysis' pronunciation robustness benchmark, ranking first, up from 88.1% for the previous generation S2.1 Pro.
50%UnverifiedOn exact sequences such as file paths and email addresses, Drama 3's accuracy improved from 79.4% to 86.3%.
50%UnverifiedDrama 3 supports director-style control, where users can insert natural-language instructions such as [barely holding it together, forcing a smile], and the instruction language can differ from the dialogue language.
50%UnverifiedDrama 3 supports word-level inpainting, allowing a single word in the audio to be replaced while keeping the surrounding tone and rhythm seamlessly connected.
50%