Fish Audio Drama 3: What the Emotive Voice AI Model Costs and How to Use It
Fish Audio's Drama 3 lets you direct AI voices with plain-English emotion tags. Here's how the API, MCP server and node access it works.
What is Fish Audio’s Drama 3 model?
Drama 3 is a text-to-speech model from Fish Audio built around emotional direction rather than flat narration. Instead of just converting text to audio, it lets you insert plain-English stage directions, like “slowly realizing something’s wrong”, directly into the script. The model carries that instruction forward through the line, so a single sentence can start calm and end in panic, the same way a voice actor would shift delivery mid-take. It also supports multi-speaker conversations, where each character keeps a consistent voice across an entire scene, and it generates word-level timestamps for synced captions.
TL;DR
- Plain-English direction tags let you embed instructions like “slowly realizing something’s wrong” inside the script text, and the model applies that emotional shift to everything that follows the tag.
- Multi-speaker scenes are supported natively, with each character assigned its own voice that stays consistent across a full conversation rather than drifting between lines.
- Word-level timestamps come back through the streaming endpoint, which matters if you’re building captions or subtitles that need to line up exactly with generated speech.
- Direction tags work across languages, not just English scripts, according to Fish Audio’s demonstration of the model.
- Access runs through the Fish Audio API by setting the model parameter to drama 3, and the company also offers an MCP server and an official node integration for workflow builders.
- The MCP server connects to coding agents like Claude Code and Codex, which matters for developers who want to script voice generation directly inside an agentic coding workflow instead of hand-calling a separate API.
- Use cases span voice agents, game characters, and video narration, anywhere flat, emotionless TTS delivery has been a limitation.
How does emotional direction actually work in Drama 3?
The core idea is that you write direction the way you’d write a note to an actor, inline with the dialogue. A tag like “slowly realizing something’s wrong” isn’t a separate metadata field you configure in a settings panel. It sits in the text itself, and the model treats everything after that tag as governed by it until a new direction appears. That’s a different approach from TTS systems that rely on a single global tone setting or a fixed “style” parameter chosen before generation starts.
This matters for anything with narrative arc. A customer service line that starts upbeat (“Thanks for calling, your order is on its way”) and pivots to alarm (“Wait, that’s not the right address, who approved this workflow?”) needs the delivery to shift mid-sentence, not just read in one uniform register. Traditional TTS tends to flatten that kind of line, because it optimizes for correct pronunciation rather than dramatic pacing. Drama 3 is positioned specifically to close that gap, and Fish Audio describes it as their most emotive model to date.
How does the multi-speaker and conversation feature work?
Beyond single-line delivery, Drama 3 can generate full back-and-forth conversations in one pass. You assign each speaker a distinct voice and mark which lines belong to which character. The model is built to hold each voice steady across the whole exchange, so a scene with two or three speakers doesn’t have one character’s voice drifting or blending into another’s over the course of a longer script.
That’s a practical requirement for anything resembling a scripted scene: game dialogue between NPCs, a dramatized ad, a narrated story with multiple characters, or a customer support simulation with an agent and a caller. Generating that kind of audio previously often meant running separate TTS calls per character and stitching the output together, which makes consistent tone and pacing across the full scene harder to control.
How do you access Drama 3 through the API, MCP server, and node?
Fish Audio exposes Drama 3 through three integration paths:
- API: You set the model parameter to drama 3 in a standard API call to the Fish Audio platform. This is the most direct route if you’re already building a custom pipeline and want full control over inputs and outputs.
- MCP server: Fish Audio provides an MCP (Model Context Protocol) server, which lets you connect the model to coding agents like Claude Code or Codex. This is aimed at developers who are already working inside an agentic coding environment and want to call voice generation as a tool within that workflow, rather than writing separate API integration code.
- Official node: There’s also an official node for node-based workflow builders, which matters for teams working in visual automation or app-building tools rather than writing raw API calls.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
The streaming endpoint is a relevant detail for anyone pairing audio with video: it returns word-level timestamps alongside the generated audio, so captions can be aligned precisely to speech without a separate forced-alignment step.
Is Drama 3 worth using over standard TTS?
The honest answer depends on what you’re building. If your use case is reading flat, informational text (a document summary, a simple notification), the emotional direction layer in Drama 3 won’t add much, and a simpler, cheaper TTS model may be sufficient. Where Drama 3 is positioned to matter is content where delivery carries meaning: narration with emotional beats, voice agents that need to sound natural across a range of moods, game characters with personality, or ad reads where tone shift is part of the message.
The ability to write direction inline, in plain English, also lowers the production overhead compared to older approaches that required SSML tags, custom prosody markup, or manual post-processing to fake emotional inflection. For builders prototyping voice agents or narrated content quickly, that’s a meaningful difference in iteration speed, since you can adjust delivery by editing a sentence rather than reconfiguring a pipeline.
No specific pricing figures for Drama 3 were detailed in Fish Audio’s public demonstration referenced here, so anyone evaluating cost should check Fish Audio’s own API documentation and pricing page directly before committing to a production integration.
Frequently Asked Questions
What makes Drama 3 different from other text-to-speech models?
It accepts plain-English emotional direction embedded directly in the script text, so delivery can shift mid-line (for example, from calm to panicked) based on a tag like “slowly realizing something’s wrong,” rather than applying one fixed tone across the whole output.
Can Drama 3 handle multiple characters in one generation?
Yes. You can assign distinct voices to different speakers and generate a full conversation in one pass, with each character’s voice held consistent throughout the scene.
Does Drama 3 only work with English scripts?
No. The direction tags are designed to work across multiple languages, not just English, based on Fish Audio’s own demonstration of the model.
How do developers integrate Drama 3 into their own apps?
Through the Fish Audio API by specifying drama 3 as the model, through Fish Audio’s MCP server for connecting to coding agents like Claude Code or Codex, or through Fish Audio’s official node for node-based workflow tools.
What can word-level timestamps be used for?
They let generated audio stay precisely synced with on-screen captions or subtitles, which is useful for video content where speech and text need to match exactly rather than approximately.
