Fish Audio has released Drama 3, a text-to-speech model aimed at creators and developers who need voice output that behaves like a performance rather than a readout. Developers write plain-language stage directions in square brackets alongside dialogue, including cues that change delivery within a line , and the model supports mid-sentence voice shifts and word-level audio inpainting so a single take can be edited without regenerating an entire clip. It supports mid-sentence tone changes, laughs, sighs and multi-speaker dialogue, with access through its app and API.
The release lands alongside Fish Audio's recent $52M seed round. The round was led by Coreline Ventures and Capital Today, with participation from 359 Capital, Parable, Play Time, Alphalist Partners, Bayhouse Ventures, Carya Venture Partners and others , giving the Palo Alto company runway to push both its open-source releases and its hosted platform. Co-founders Shijia Liao, chief scientist, and Rissa Cao, CEO , have been building toward production use cases like AI avatars, game characters, and voice agents, with customers such as HeyGen and LiveKit cited in press coverage.
What makes this launch notable is the open-weight distribution of a frontier-scale speech model. Drama 3 is a 35B mixture-of-experts system published on Hugging Face , and installation requires an NVIDIA GPU with at least 48 GB of memory in FP8 , putting it within reach of independent studios and infrastructure teams who want to self-host rather than depend on a closed API. For founders building voice-native products, it is one of the first directable TTS systems that can be inspected, fine-tuned, and shipped on owned infrastructure.
Comments (10)
Sign in to join the discussion.
N
Nnedi W.20h ago
The per-word tone control demo in the video actually made me rewind. Most TTS launches I skip at second three.
R
Rina Kowalczyk19h ago
Open-weight means nothing without a decent model card and a quickstart that doesn't require a PhD. Where are the docs living?
T
Tomás Berila20h ago
Open-weight is cute until someone has to pay the GPU bill. Who's eating the inference cost when hobbyists go wild this weekend?
H
hoshikawa19h ago
The tweet buried the lede. $52M and 15M users as throat-clearing before the actual announcement is a flex I respect.
Z
Zee Abdelrahman19h ago
"The most emotive and controllable TTS model" reads like a spec sheet. Try: "Every word, exactly how you'd say it." You're welcome.
A
Aminata Diouf19h ago
This might be the last voice model shipped without an agent wrapper on top. Enjoy the quiet before every startup bolts a persona onto it.
V
Vikram P.19h ago
Dumb question but if I can shift tone mid-word, what happens with languages that use pitch for meaning? Does the controllability break Mandarin?
L
Lin Oyelaran19h ago
Open-weight is lovely for the demo crowd. For my team I still need SOC2, DPA, and someone to pick up the phone when audio goes sideways in prod.
M
Moke Haugen19h ago
Curious how lean the team shipping Drama 3 actually is. The ratio of model quality to headcount is the real benchmark now.
G
Gunnar Molloy19h ago
Still not sure why anyone needs TTS beyond airport announcements, but sure, make the robot cry.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Fish Audio
The most emotive and controllable TTS model
The most emotive and controllable TTS model
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Fish Audio has released Drama 3, a text-to-speech model aimed at creators and developers who need voice output that behaves like a performance rather than a readout. Developers write plain-language stage directions in square brackets alongside dialogue, including cues that change delivery within a line , and the model supports mid-sentence voice shifts and word-level audio inpainting so a single take can be edited without regenerating an entire clip. It supports mid-sentence tone changes, laughs, sighs and multi-speaker dialogue, with access through its app and API. The release lands alongside Fish Audio's recent $52M seed round. The round was led by Coreline Ventures and Capital Today, with participation from 359 Capital, Parable, Play Time, Alphalist Partners, Bayhouse Ventures, Carya Venture Partners and others , giving the Palo Alto company runway to push both its open-source releases and its hosted platform. Co-founders Shijia Liao, chief scientist, and Rissa Cao, CEO , have been building toward production use cases like AI avatars, game characters, and voice agents, with customers such as HeyGen and LiveKit cited in press coverage. What makes this launch notable is the open-weight distribution of a frontier-scale speech model. Drama 3 is a 35B mixture-of-experts system published on Hugging Face , and installation requires an NVIDIA GPU with at least 48 GB of memory in FP8 , putting it within reach of independent studios and infrastructure teams who want to self-host rather than depend on a closed API. For founders building voice-native products, it is one of the first directable TTS systems that can be inspected, fine-tuned, and shipped on owned infrastructure.
Comments (10)
The per-word tone control demo in the video actually made me rewind. Most TTS launches I skip at second three.
Open-weight means nothing without a decent model card and a quickstart that doesn't require a PhD. Where are the docs living?
Open-weight is cute until someone has to pay the GPU bill. Who's eating the inference cost when hobbyists go wild this weekend?
The tweet buried the lede. $52M and 15M users as throat-clearing before the actual announcement is a flex I respect.
"The most emotive and controllable TTS model" reads like a spec sheet. Try: "Every word, exactly how you'd say it." You're welcome.
This might be the last voice model shipped without an agent wrapper on top. Enjoy the quiet before every startup bolts a persona onto it.
Dumb question but if I can shift tone mid-word, what happens with languages that use pitch for meaning? Does the controllability break Mandarin?
Open-weight is lovely for the demo crowd. For my team I still need SOC2, DPA, and someone to pick up the phone when audio goes sideways in prod.
Curious how lean the team shipping Drama 3 actually is. The ratio of model quality to headcount is the real benchmark now.
Still not sure why anyone needs TTS beyond airport announcements, but sure, make the robot cry.