Fish Audio
Expressive Text-to-Speech and Voice Cloning
4.7•15 reviews•2.1K followers
Expressive Text-to-Speech and Voice Cloning
4.7•15 reviews•2.1K followers
Fish Audio is the most expressive and emotionally rich text-to-speech model. It generates lifelike voices that capture emotion, rhythm, and nuance with remarkable realism. Fish Audio Voice Clone recreates a natural voice from just 10 seconds of audio—preserving accent, tone, and speaking habits. Proudly built by the open-source team behind So-VITS-SVC and Bert-VITS2, giving a soul to every voice.
This is the 7th launch from Fish Audio. View more
Fish Audio Live 3
Launching today
Fish Audio Live 3 is a full-duplex voice model that keeps listening while it speaks, so pauses, “mm-hmm” acknowledgments and changes of mind shape what it does next. Per Fish Audio, timing is learned from real human conversations inside the speech model, not a pipeline of separate parts, and it can decide to respond in as little as 186 ms. English only at launch. On the Fish Agent Platform and the API, with on-prem for enterprises.
Launch Team
Render Sandboxes: a computer for every agentRun untrusted code safely. Idle sandboxes cost nothing
Promoted
Fish Audio Live 3 is a full-duplex voice model, which means it keeps listening while it speaks. That’s the difference between a voice agent that feels like a phone menu and one that feels like a conversation.
Most voice agents work as a pipeline of separate parts for speech recognition, reasoning, speech generation and turn-taking. Fish Audio built conversational timing into the speech model itself. Per Fish Audio, it trained on real human conversations, so the model learned when to acknowledge, wait, respond or stop, not from hand-written rules. As CEO Rissa Cao puts it, “We think a model should learn this from real conversations, the same way language models learn grammar.”
What that looks like in a call:
Short acknowledgments like “mm-hmm” and “okay”, including reading digits back in chunks
Not stopping when the caller says “yeah” during an explanation
Stopping and making room when the caller has a new request
Waiting through a self-correction, like changing an order, instead of answering early
Leaving space when someone is still mid-phone-number
Speed: Fish Audio says Live 3 can decide to respond in as little as 186 milliseconds. That’s the only number in the announcement, and it’s the time to decide to respond, not end-to-end latency. There are no comparisons with other voice models.
A separate reasoning layer handles your agent’s instructions and tools. You build that, and Live 3 handles the conversational rhythm.
It’s available on the Fish Agent Platform and through the API, and enterprises can deploy it on-premises. Live 3 speaks English only at launch. The post doesn’t mention pricing or voice options, so check the pricing page and docs.
Good fit if you’re building phone or voice agents for things like orders, bookings or support, where timing and interruptions matter as much as the words.