Fish Audio launches Drama 3, a 35B MoE TTS model with real-time emotional expression. Open-weight, multi-speaker, supports 100+ languages including Cantonese.
A new open-weight TTS model just dropped that brings genuine emotional range to AI-generated voices. Fish Audio Drama 3 is a 35-billion parameter mixture-of-experts text-to-speech model that doesn't just read words — it delivers them with [excited], [laughing], [sighing], and a full palette of emotional tags.
Launched in mid-October 2026 and already trending on HuggingFace (with 339 likes in its first two weeks), Drama 3 is Fish Audio's first MoE architecture for voice generation. It builds on the company's earlier 2.x generation with a completely new model type called FishOmniModel, quantized via BSGemm-MoE for efficient inference.
What Makes Drama 3 Different
The headline feature is real-time emotional control. Instead of flat, monotone delivery, you write a script and tag specific phrases with emotional markers:
- [excited] — for product launch announcements and hype content - [laughing], [chuckling] — for casual, conversational tone - [sighing], [groaning] — for storytelling or dramatic effect - [whisper] — for intimate or serious moments
This is a fundamental shift from previous TTS models that could only mimic emotion through prompt engineering. Drama 3 makes it a first-class parameter.
The model supports over 100 languages, including Chinese (Cantonese and Mandarin), Japanese, Korean, and most European languages. For Hong Kong creators producing bilingual content, this means one model handles both English voiceovers and Cantonese narration without switching providers.
Open Weight Means Cost Control
Drama 3 is fully open-weight on HuggingFace under a permissive license, complete with the arxiv paper (2610.18010) describing the architecture. This opens up two practical paths for creators:
- Self-hosted inference — run Drama 3 on your own hardware for voiceover production, avoiding per-character API fees - API usage — Fish Audio's existing API infrastructure already supports Drama 3 for those who prefer managed service
For Hong Kong creative agencies producing 20-50 voiceover clips per week for social media ads, product demos, or video content, the cost difference between API-based TTS and self-hosted open-weight models can be significant.
Practical Use Cases for HK Creators
1. Social media ads — emotionally varied voiceovers without recording studio time. A single 60-second ad can shift between [excited] product intro and [whisper] serious closing 2. E-commerce product videos — generate 20-30 product voiceovers in batch with consistent tone, then adjust emotional intensity per product category 3. Bilingual content production — one model for English + Chinese voiceover, using the same emotional tags across both languages 4. AI video voiceovers — pair Drama 3 with an AI video generation pipeline in Cooly Studio for end-to-end content creation: script → TTS → video → final cut
Frequently Asked Questions
Q: Is Fish Audio Drama 3 free to use? A: The model weights are open-source on HuggingFace. You can self-host it for free. Fish Audio also offers a cloud API with a free tier for testing.
Q: Does Drama 3 support Cantonese? A: Yes. The model supports Chinese dialects including Cantonese through its multilingual training. Tags like zh cover standard Chinese, and the model handles tonal languages naturally.
Q: How does the emotional tagging work? A: You add inline tags like [excited] or [sighing] directly in your script text. The model reads the tag and adjusts its delivery in real time. Tags can change mid-sentence for dynamic narration.
Q: Can I use Drama 3 with Cooly.ai? A: Yes — you can generate voiceovers with Drama 3 separately and import them into Cooly Studio for video production. The open-weight format means you can set up a local pipeline without API dependencies.
Q: How does it compare to ElevenLabs or Gradium? A: Drama 3 competes on emotional expressiveness and open availability. Unlike closed APIs, you can self-host it. Its MoE architecture gives it strong inference speed for a 35B model.