On November 10 (02:02 on November 11 Singapore time) Fish Audio released Live 3, a full-duplex voice model that keeps listening while it speaks. It can acknowledge with short cues such as "mm-hmm" or "okay" while leaving the floor to the user, tell a backchannel apart from a real interruption, and wait while a caller corrects themselves. Fish Audio says these behaviors were learned from real human conversations rather than hand-written rules and thresholds, and that Live 3 can decide to respond in as little as 186 milliseconds. Live 3 speaks English only for now, is available on the Fish Agent Platform and via API, and can be deployed on-premises by enterprises; voice agents are priced at a limited-time $0.035 per minute (list $0.07, offer ends December 31, 2026). No standardized benchmark has been published.
Key Takeaways
- ✓Full duplex: Live 3 keeps listening while it speaks, moving turn-taking into the speech model itself instead of rules in an ASR, LLM and TTS pipeline
- ✓Latency (vendor figure): decides to respond in as little as 186 ms, using short replies like "mm-hmm" or "okay" while the full answer is prepared
- ✓Pricing: voice agents at $0.035/min (list $0.07, 50% off until 2026-12-31), with speech recognition, LLM and voice billed together; phone calls add $0.01/min and sign-up includes $5 credit
- ✓Limits: English only for now; no standardized benchmark published, only 3 excerpts from real test calls
- ✓Deployment: Fish Agent Platform, API, or enterprise on-premises, supporting both full-duplex and cascaded modes
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and Industry Pain Points
Most voice agents chain speech recognition, LLM reasoning, speech synthesis and turn detection, and rely on silence thresholds to guess when a user is done. The result: a caller pausing mid phone number gets cut off, and a quick "yeah" to show they are listening makes the agent stop as if interrupted. Fish Audio CEO Rissa Cao says the industry fix so far has been "more rules and more thresholds," and argues a model should learn this from real conversations, the way language models learn grammar.
Architecture and Mechanism
Live 3 is full duplex: it keeps listening while it speaks, bringing conversational timing into the speech model itself. Fish Audio says it was trained on real human conversations to learn when to acknowledge, wait, respond or stop. When a caller reads "415... 555... 0132" in chunks, it may say "mm-hmm" or "okay, 415" and wait for the rest; it keeps talking through a backchannel but stops at "Wait, actually, I just need Thursday"; and it waits while a caller keeps revising an order.
Benchmarks and Measured Results
No standardized benchmark or third-party evaluation has been published. The only quantitative figure is that Live 3 can decide to respond in as little as 186 ms (vendor figure). The blog shares just 3 excerpts from real test calls. For context, ElevenLabs Scribe v3 Live (November 10) added turn detection to realtime transcription and Nuance Labs released the full-duplex audiovisual model Nuplex 1.0 on November 9; vendors are all attacking the "when to speak" problem, but there is no shared evaluation set yet.
Developer Quick Start
Sign up at fish.audio/agents for $5 of free credit with no card required. Speech recognition, LLM and voice are billed together per minute at a limited-time $0.035; switching to a larger LLM is billed at that model's token rate, and inbound or outbound phone calls add $0.01 per minute. English is the only language today, so Chinese customer service is not yet supported. It fits ordering, bookings and support callbacks where callers pause and change their minds; run simulation tests on your own real call recordings before going live.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.