Developers have unveiled Shizuku-Duplex, a new Japanese voice model that can listen while it speaks, a capability that could move spoken AI conversations beyond strict turn-taking.
Voice AI · Model announcement
Shizuku-Duplex aims to listen while it speaks in Japanese
Developers say the model can process incoming speech while generating its own reply, a step beyond the strict turn-taking of most voice assistants. The claim is not yet independently verified.
How a conversational turn flows
Turn-based assistants versus a full-duplex model
Turn-based (most voice assistants)
User finishes speaking
→
System processes input
→
System replies, deaf to input
Full-duplex (Shizuku-Duplex, as described)
Model speaks
+
Model keeps listening
→
Handles interruptions and "uh-huh" signals
What the developers claim
Processes speech while generating its own reply
Speaks fluent Japanese
Handles overlapping speech more naturally
First Japanese full-duplex model since GPT-Live
What remains open
No independent verification yet
No benchmarks, latency or model size provided
Telling a user's interruption from background noise
Avoiding awkward overlaps in live talk
Key question for the coming weeks
Can Shizuku-Duplex hold natural, low-latency Japanese dialogue outside controlled demos?
Real-time performance
Does it stay responsive outside scripted demonstrations?
Evaluation data
Will the developers release the data behind the "first since GPT-Live" claim?
Third-party tests
Until independent results appear, treat the capabilities as the developers' own description.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...