Fish Audio Drama 3: Directable AI Voice Acting Explained
Fish Audio's Drama 3 lets you direct AI voices like an actor. Here's how it works, what hardware it needs, and how it stacks up to ElevenLabs.
What is Fish Audio Drama 3?
Drama 3 is a text-to-speech model from Fish Audio that lets you direct a voice performance in plain English instead of picking from a dropdown of preset emotions. You write stage directions in square brackets right inside the script (things like “barely holding it together, forcing a smile”), and the model adjusts tone, pacing, and delivery to match. It also supports word-level control, such as whispering a single phrase, inline sound effects like a shaky breath or a small laugh, and multi-speaker scripts with distinct voices for each character.
TL;DR
- Directable performance is the core idea: instead of choosing “sad” or “angry” from a menu, you write actor-style direction in brackets inside the text, and Drama 3 performs it.
- Word and phrase level control lets you target specific words for effects like whispering, rather than applying one emotion to an entire line.
- Inline sound effects and multi-speaker dialogue are supported natively, so a single script can produce a short scene with several distinct voices and nonverbal cues like breaths or laughs.
- It tops pronunciation benchmarks on Artificial Analysis, scoring higher than ElevenLabs and other competitors on reading difficult words correctly, though it ranks lower on blind human preference tests.
- It’s open weight, released on Hugging Face as a 36 billion parameter model (fishaudio/drama-3), which means you can self-host it instead of relying only on the API.
- Running it locally takes real hardware, with reports pointing to roughly 46GB of VRAM for the full model or around 32GB using a community 4-bit quantized version, putting it well outside laptop territory.
- API pricing starts at a launch discount of about $10 per million characters, set to rise to $50 per million characters after November 12, so costs will look very different depending on when you start using it.
How does directable text-to-speech work?
Most TTS tools treat emotion as a setting. You pick “happy” or “sad” from a dropdown, and the whole line gets generated with that single flavor. Drama 3 flips that by letting you write direction the way you’d brief a voice actor. You embed instructions in brackets directly in the script text, and the model reads both the words and the direction together to decide how to perform the line.
That granularity is the interesting part. You’re not just tagging an entire sentence as “sad.” You can mark a single word or phrase for a specific treatment, like whispering just the last few words of a line while the rest plays at normal volume. You can also drop in nonverbal cues, such as a shaky breath or a quiet laugh, and place them exactly where you want them in the audio rather than hoping the model inserts them somewhere reasonable.
The model also handles multiple speakers in one generation. A short back-and-forth exchange between two characters, each with a different voice, can come out of a single script rather than requiring separate generations stitched together afterward. Fish Audio’s demo page even includes a licensed voice modeled on the YouTube personality Uncle Roger, which gives a sense of how far the voice cloning and character consistency goes.
How does Drama 3 compare to ElevenLabs and other TTS models?
On Artificial Analysis’s benchmark tracking, Drama 3 currently scores the highest on pronunciation accuracy among the models tested, beating out ElevenLabs and SpeechX AI’s text-to-speech model. That means it’s particularly good at correctly voicing difficult or uncommon words, which matters a lot for use cases like audiobooks or technical narration where mispronunciations break immersion.
It does not lead on every metric, though. On the Arena-style leaderboard, which relies on blind listening tests where people pick their preferred voice without knowing which model generated it, Drama 3 doesn’t come out on top. In other words, it reads hard words more accurately than most competitors, but when people are judging overall voice quality and naturalness without that context, other models still win more often. If your use case is heavy on technical vocabulary, names, or scripted dialogue with precise pronunciation needs, that tradeoff favors Drama 3. If you’re optimizing purely for how pleasant or natural a voice sounds in a blind comparison, it’s worth testing both and listening yourself.
What hardware do you need to run Drama 3 yourself?
Drama 3 is released as an open weight model on Hugging Face under the repository fishaudio/drama-3, built on a 36 billion parameter architecture (the “fish_omni” model type, using safetensors format split across eight shards). That size puts real hardware requirements on self-hosting.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Reports from people running the model suggest you need roughly 46GB of VRAM to run it at full precision. A community-produced 4-bit quantized version lowers that requirement to around 32GB. Either number rules out running this on a typical laptop or consumer GPU with 8 to 16GB of VRAM. You’re looking at workstation-class or data center GPUs, or quantization plus a higher-end consumer card, to get it running locally.
This is the core tradeoff with open weight release: you get the ability to self-host, fine-tune, and avoid per-character API costs, but only if you have access to serious compute. For most individual creators, the API will be the more practical path, with self-hosting making more sense for teams or companies with existing GPU infrastructure and high enough volume to justify it.
Is Drama 3 worth using, and what does it cost?
Whether Drama 3 is worth adopting depends heavily on timing and use case. The API currently runs about $10 per million characters, but that’s described as a launch discount that jumps to $50 per million characters after November 12. That’s a five-fold price increase, so anyone evaluating this for a production pipeline needs to factor in which pricing tier they’ll actually be paying under.
A few practical notes matter if you’re testing it. Output quality varies between takes, so plan on generating several versions of a line and keeping the best one rather than expecting one-shot perfection every time. Also double check that your API calls are explicitly set to the Drama 3 model, since it can quietly fall back to older Fish Audio models, and you’ll still get billed even though you’re not getting Drama 3’s directable features.
On licensing, Fish Audio’s terms allow free commercial use as long as your business makes under $500,000 a year in revenue, which covers most individual creators and small studios but not larger commercial operations.
For audiobook producers, game developers needing character dialogue, or anyone building narrated content with multiple voices and emotional range, Drama 3’s bracket-based direction system offers a level of control that preset-based tools don’t match. The combination of word-level targeting, inline sound effects, and multi-speaker support in a single script is aimed squarely at that kind of narrative, performance-driven use case rather than simple text readout.
Frequently Asked Questions
What is Fish Audio Drama 3 used for?
It’s designed for narrative and performance-driven voice generation, such as audiobooks, game dialogue, animated scenes, and any content where a character needs to deliver a line with specific emotion, pacing, or nonverbal cues like a whisper or a breath.
How much VRAM does Drama 3 need to run locally?
Running the full model takes roughly 46GB of VRAM based on user reports. A community 4-bit quantized version reduces that to around 32GB, but both figures are beyond what most consumer laptops or entry-level GPUs can handle.
Is Fish Audio Drama 3 free to use?
The API has a launch price of about $10 per million characters, rising to $50 per million characters after November 12. The open weight model is free to self-host and license for commercial use if your business earns under $500,000 a year in revenue.
How is Drama 3 different from ElevenLabs?
Drama 3 uses plain-English bracketed direction embedded in the script to control emotion, pacing, and word-level effects like whispering, rather than relying on preset emotion menus. It also scores higher than ElevenLabs on pronunciation benchmarks, though ElevenLabs and other models still rank higher in blind human preference listening tests.
Can I fine-tune or self-host Drama 3?
Yes. It’s released as an open weight 36 billion parameter model on Hugging Face, so it can be downloaded and run on your own hardware, provided you have enough VRAM, instead of relying solely on Fish Audio’s API.
