Fish Audio Releases Drama 3 35B TTS Model
Fish Audio launches Drama 3, a 35B mixture-of-experts text-to-speech model supporting 83 languages with open weights release.
Stock photo for illustration only, not from the actual event
- Fish Audio introduces Drama 3, a 35B text-to-speech model
- Built on Mixture-of-Experts architecture with 3B active parameters
- Supports an extensive range of 83 languages globally
- Released as open weights under a revenue-capped community license
The artificial intelligence and speech synthesis landscape witnesses a new milestone as Fish Audio officially announces the release of Drama 3, an advanced text-to-speech model designed to understand plain-language direction and deliver highly natural vocal outputs.
Drama 3 is engineered with a total parameter scale of 35 billion (35B), utilizing a Mixture-of-Experts (MoE) architecture. This sophisticated design allows the model to maintain exceptional intelligence while keeping computational efficiency high, operating with approximately 3 billion active parameters (3B) at any given time.
Stock photo for illustration only, not from the actual event
A standout feature of Drama 3 is its capability to interpret plain-language directions seamlessly, enabling creators and developers to guide tone, emotion, and style effortlessly. Furthermore, the model demonstrates remarkable versatility by expanding its language coverage to 83 distinct languages worldwide.
Implementing a Mixture-of-Experts (MoE) architecture in text-to-speech models allows Fish Audio to achieve sophisticated voice modulation without incurring massive computational costs, as only a fraction of the parameters are activated during inference. Releasing the model weights under an open-weights approach with a revenue-capped community license strikes a balance between open-source collaboration and commercial oversight.
Regarding distribution, Fish Audio has made Drama 3 available as open weights, though governed strictly by a revenue-capped community license to manage large-scale commercial deployments effectively.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment