by Fish Audio
Fish Audio's directable text-to-speech model controlled by plain-language bracketed directions, sound tags and multi-speaker scripts.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Drama 3 is a text-to-speech model from Fish Audio built around performance direction. Alongside the script, users describe how a line should be delivered in natural language inside square brackets (for example "[sigh]" or "[barely holding it together]"), insert sounds and pauses, and use paired tags such as <whisper> or <slow> to limit an effect to selected words. Directions can be written in English or other languages.
It supports multi-speaker dialogue: each speaker is assigned a voice and the model generates the whole scene in one pass. An "Auto Tag" feature adds directions to a plain script as a starting point, which users can keep or rewrite. According to YouTube creator coverage, it ranks highest among open models on Artificial Analysis for pronunciation, though it scored lower on arena preference. Reported specs are a roughly 35-36B parameter mixture-of-experts model with about 3B active parameters, reportedly running on around 8 GB VRAM; these figures are not confirmed on the official page.
Drama 3 is available in the Fish Audio Text to Speech editor and through the API (https://api.fish.audio/v1/tts with the model: drama-3 header). Creators report open weights with free commercial use under a revenue cap and API pricing of about $10 per million characters at launch; the exact license terms were not verified from the provided pages.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Text-to-speech model where performance is directed in plain English via bracketed instructions, with sound effects and multi-speaker support. It scores highest on pronunciation on Artificial Analysis but lower on arena preference. Open weights (36B params), free commercial use under a revenue cap, API about $10 per million characters at launch.
Open-source text-to-speech model supporting meta tags to control emotion and delivery. Ranked number one among open models on the Artificial Analysis leaderboard; 35B total parameters with 3B active, reportedly runs on 8 GB VRAM.
Closest neighbours by meaning, not just by category.
Model
Expressive text-to-speech model with zero-shot voice cloning, derived from the LTX 2.3 audio model.
Model · Resemble AI
Open-weights expressive text-to-speech model from Resemble AI with voice cloning, fine-tuned from LTX-2.3 audio.
Model · Ant Group
Open-weights text-to-speech models from Ant Group's Inclusion AI with voice design, cloning and tag-based control.
Model · IndexTeam
Open-weights zero-shot text-to-speech model from Bilibili's IndexTeam with voice cloning and emotion control in 5 languages.
Model
Compact 0.6B open text-to-speech model with zero-shot voice cloning and support for 23 languages.
Model · Google
Google's Gemini 3.8 text-to-speech model for expressive, controllable speech generation via the Gemini API and AI Studio.
