Fish Audio says its new Live 3 model can decide whether to speak in 186 milliseconds, a figure the company announced on 2026-11-10 but has not yet submitted for independent verification. The model is built for full-duplex voice interaction, meaning it listens while it speaks, and its core claim is that it learned turn-taking from recorded human conversation instead of relying on silence thresholds.
How the Architecture Differs
Most voice agents treat any pause longer than a set threshold as the end of a turn. That works until someone reads a phone number in chunks — “415… 555… 0132” — and the agent hears three separate turns, or one long hesitation. Fish Audio’s worked example shows Live 3 responding with “mm-hmm” or “okay, 415” between the chunks, confirming receipt while leaving the floor open. The same logic applies when a caller says “yeah” during a long readout: the model keeps going. When the caller then says “Wait, actually, I just need Thursday,” it stops and yields. Corrections mid-utterance — “two pizzas… no, three” — are documented to wait through each revision rather than firing on the first plausible complete order.
What the 186-Millisecond Figure Represents
The 186-millisecond number is vendor-reported and unaudited. It measures the time the model takes to decide whether to speak, not the full round-trip latency a user would experience. Fish Audio has published no evaluation for Live 3 as of 2026-11-11, no per-byte pricing, and no model identifier or endpoint documentation naming the model. Artificial Analysis, which runs a duplex speech leaderboard, carries no Live 3 row; the board lists Fish Audio as a model family but scores none of its duplex work.
Context From Competing Open Models
Open-source alternatives are publishing self-reported benchmarks on the same day. Zyphra released ZONOS3-Mini, an 800-million-parameter mixture-of-experts text-to-speech model with 100 million active parameters, and reported 80 milliseconds median time to first audio on a single H100 at eight concurrent requests. Zyphra’s own ZTTS2-Eval benchmark places ZONOS3-Mini ahead of Fish Audio S2 Pro on a capability index, though Zyphra built and ran the benchmark itself. Reference-clip cloning scores for ZONOS3-Mini trail embedding mode significantly. Fish Audio’s Live 3 announcement does not reference these comparisons.
What This Means for Developers
One day after launch, a developer cannot call a documented endpoint, compare an independent score, or estimate cost. The worked examples suggest the turn-taking logic solves real problems in support calls, bookings, and orders where chunked speech and mid-utterance corrections are common. Whether the 186-millisecond decision window holds under load, across languages, or in noisy environments remains untested in public. Fish Audio’s chief executive framed the approach as learning what not every sound is an interruption without a developer enumerating every sound a human might make.
Sources
- Fish Audio Live 3: 186 ms Full-Duplex Voice, Still Unscored
- Soundtrack Beat Battle
- How this year's record El Niño could hit your wallet : Short Wave
- RuneScape: Dragonwilds Update 1.0.2 Patch Notes
- Nearly 2,000 dead seabirds wash ashore in California. Scientists point to the ocean
- Zyphra Releases ZONOS3-Mini: Open-Source TTS With 100M Active Parameters
Photo: Will Francis – AI & Marketing / Unsplash (Unsplash License)
