Fish Audio Releases Open Source Drama 3 Text-to-Speech Model
Fish Audio has launched Drama 3, claiming it to be the world's most emotionally expressive and controllable text-to-speech model, with weights now open-sourced. The company previously completed a $52 million seed round financing, with a user base exceeding 15 million.
Drama 3 is a 35B parameter mixture model, activating about 3B parameters per token. Users can describe single-line performances in natural language, including emotions, speech rate, and characters. It supports switching tones or voices mid-sentence, applying effects to individual words, and adding sounds or pauses. The model supports multi-character scenes, voice cloning, local editing, and long audio streaming generation. It supports 83 languages.
The company showcased Uncle Roger using its voice cloning in the launch video. Weights are released on Hugging Face under the Fish Audio Community License v1.0. Research and non-commercial use are free; commercial use with annual revenue under $500,000 is free; higher revenue or API services require separate licensing. Cloning real human voices without consent is prohibited. API pricing is $50 per million characters, with discounts available until November 12.
Fish Audio was initiated by former NVIDIA researcher Shijia Liao, initially training and open-sourcing on a single GPU. Within a year, the company grew from a project to a platform with an annual recurring revenue of $21 million and over 8 million users, later increasing to over 15 million. Clients include HeyGen, LiveKit, Retell, and others. The model replaces traditional labels with natural language control.
Funding has shifted from traditional voice actors and post-recording to model calls and open-source weights. Creators and businesses reduce long audio modification costs through controllable voice. Fish Audio generates revenue through hosted APIs and enterprise licensing. Open-source weights attract developers, while hosted services and low latency maintain paid subscriptions. Competitors relying on fixed emotional labels face relative pressure on controllability.
Source: Public Information
ABAB AI Insight
Fish Audio was founded in 2025 by former NVIDIA researcher Shijia Liao, initially open-sourcing Fish Speech. In July 2026, it completed a $52 million seed round, with around 8 million users and an annual recurring revenue of $21 million at that time. The user base has since grown to over 15 million. The company releases Drama 3 as open-source weights while providing a hosted API.
The company shifts voice control from SSML tags to natural language commands, supporting tone adjustments for words and mid-sentence switches. Resources are moving from closed API models to open-source weights plus hosted services. Creators can embed performance instructions in scripts, reducing the need for complete re-recording. Businesses obtain controllable voice for customer service and content through the API.
Similar open-source voice models have been released by other labs. The current text-to-speech technology is transitioning from fixed emotional labels to natural language directorial control.
Essentially, this represents a technological substitution. Natural language commands replace some traditional labels and post-production adjustments in voice generation. The mechanism involves directly embedding performance descriptions into text, allowing the model to handle emotional changes, word corrections, and multi-character dialogues in a single generation, while lowering the entry barrier for developers through open-source weights, shifting revenue from pure model calls to hosting and licensing.
ABAB News · Cognitive Laws
Voice control shifts from fixed parameters to performance descriptions after natural language commands replace labels
Open-source weights attract developers, while hosted services maintain paid revenue
Word corrections reduce long audio re-recording costs, shifting modifications from entire segments to local adjustments