Fish Audio Drama 3: Directable AI Voiceover in 83 Languages

The AI voice generation race is shifting from "can it talk" to "can it perform." Fish Audio's Drama 3 model, launched on Product Hunt, focuses on a capability most voice tools overlook: directability. It doesn't just turn text into speech—it lets you control emotion, pacing, and pauses like a director, and even switch from a whisper to a laugh within the same sentence.

From "generation" to "direction": what's the difference

Most TTS tools in the past stayed at the level of "reading it out." You enter text, it outputs a chunk of audio, and emotion is basically hard-switched via tone tags. If you want to change emotion within a sentence, you have to split it into multiple segments, generate them separately, and stitch them together.

Drama 3 takes a different approach. It supports using natural-language instructions to control performance details—emotion, speaking rate, whispering, laughter, pauses—and these controls can happen within a single sentence. This means creators don't need to break the script into fragments; a character's shift from suppression to outburst in one line can be achieved with a single instruction.

This difference looks small, but its practical impact is large. For audiobooks, game dialogue, and multi-character radio dramas, the biggest fear is obvious stitching artifacts and emotional discontinuity. A directable model moves the work of post-production editing earlier into the generation stage.

TTS (Text-to-Speech) is the core of speech synthesis technology, and its goal is to convert written text into natural speech. Traditional TTS systems are usually based on concatenative synthesis or parametric synthesis, while in recent years end-to-end neural networks (such as Tacotron, FastSpeech, etc.) have become common. These models have made significant progress in audio quality and naturalness, but their controllability remains limited, often relying on preset speaker IDs or simple prosody markers. The "directability" emphasized by Drama 3 targets exactly this limitation: it attempts to upgrade TTS from mere "reading aloud" to "performance," allowing users to guide emotion, pacing, and details in real time through natural-language instructions, thereby achieving complex expressive changes in a single generation.

Multi-character scenarios and coverage of 83 languages

Drama 3 is explicitly aimed at four types of content: multi-character scenes, audiobooks, game dialogue, and narration. Multi-character means the same model must be able to stably distinguish different characters' timbres and performance styles, which is much harder than generating a single voice.

Language coverage is another notable number: 83 languages. For teams distributing multilingual content, this means the same script can be extended directly to multiple languages without having to separately find voiceover resources for each language. Of course, whether actual performance is balanced across languages still needs more real-world testing to verify—the official launch did not provide quality comparison data for each language.

Multi-character speech synthesis requires the model to distinguish and stably perform different speakers' timbres, intonation, and styles within the same audio clip. Traditional methods require training separate models for each character or using multi-speaker models, but the latter often suffer from timbre confusion or style drift when switching characters. Drama 3's claim to handle multi-character scenarios suggests that its underlying approach may use conditional generation or prompt learning mechanisms, dynamically adjusting output through character descriptions or reference audio. This capability is crucial for audiobooks and radio dramas, because listeners can easily notice inconsistencies in character voices. The actual effect depends on how well the model disentangles character embeddings and its ability to maintain character consistency across different languages.

Available across platforms, with open weights

Drama 3's access methods cover mainstream scenarios:

• Web: suitable for quick previews and adjustments
• Mobile: convenient for creating anytime, anywhere
• API: for teams that need batch generation or integration into their own products
• Open weights: self-hosting supported under the Fish Audio community license

The open-weights point deserves separate mention. Self-hosting speech models is important for scenarios that are privacy-sensitive or have compliance requirements; creators can deploy the model in their own environment without relying on cloud services. However, note that this is a "community license" rather than a fully open-source license, so specific terms need to be confirmed before commercial use.

Open weights means that the trained parameter files of a model are publicly released, allowing anyone to download and use them for inference or fine-tuning. This differs from fully open source, which usually also includes training code, data, and permissive licenses. The "community license" adopted by Fish Audio may include commercial restrictions, attribution requirements, or constraints on derivative works. For enterprises hoping to self-host, open weights mean data does not leave local systems, satisfying privacy and compliance needs, but the license terms must be carefully read to avoid legal risks. In addition, self-hosting requires a certain level of technical capability and hardware resources, such as GPU servers, which may pose a barrier for small teams.

What this launch says

Drama 3 received 91 votes and ranked 8th on Product Hunt, with only 5 comments. This combination of data shows that it is currently still in the stage of "being seen but not fully discussed"—a vertical tool with some visibility, but not yet large-scale user feedback.

What is truly worth watching is the product direction: the competitive focus of AI voice is shifting from audio quality to controllability. Audio quality has already been raised to a relatively high baseline over the past two years, and what will create differentiation next is whether creators can precisely express the effect they want. Giving directorial control back to creators, rather than letting the model improvise randomly, is the position Drama 3 is trying to occupy.

For creators working in podcasts, audio content, and game localization, this type of tool is worth adding to the testing list. But it is advisable to first run it with your own real script, focusing on the stability of emotional switching within a single sentence and consistency across multiple characters—these two points are the real touchstones of directability.

One-line Summary

Fish Audio released the Drama 3 voice model, emphasizing directability, with support for emotional control within a single sentence and 83 languages.

Paragraph Summary

Fish Audio launched the Drama 3 speech generation model on Product Hunt. Its core innovation is "directability": users can use natural-language instructions to control emotion, speaking rate, whispering, laughter, and pauses, with support for emotional switching within a single sentence. The model targets multi-character scenes, audiobooks, game dialogue, and narration, covers 83 languages, offers web, mobile, and API access, and opens weights for self-hosting under a community license. At launch it received 91 votes and ranked 8th, but had only 5 comments, so user feedback is still insufficient. The article argues that competition in the AI voice space is shifting from audio quality to controllability, and Drama 3 tries to return directorial control to creators. Creators are advised to test single-sentence emotional switching stability and multi-character consistency with real scripts.