backchannel

Full-duplex models are everywhere. Agreement on what “full-duplex” means? Less so.

Does a cascade count if it delivers the same interaction? What’s keeping builders from using these models in production?

We look at the recent wave, the terminology debate, and the engineering tradeoffs.

x.com/heybackchannel…

#BackchannelSignals

Inspired by @_MarcosValera's list of voice AI people worth following, here are a few builders behind the projects we've covered this week.

>> @RealRichomie is part of the AveraLabs team behind Awaken, which adapts half-duplex speech models for full-duplex conversation.
>> @rissa_cao co-founded Fish Audio, whose Live 3 knows to keep going after a “yeah” and stop for a “wait.”
>> @sharathkrml zones out while reading, so he built Read Along, a Mac app whose cursor follows your voice—even across a handwritten page.
>> @anders_thuesen spent a year building Duo with his team, a full-duplex speech model with a preview they say runs on a laptop.
>> @hubeiqiao spent eight months taking Joe Speaking, his IELTS and CELPIP speaking coach, from alpha to 1.0 as a solo builder.
>> @iam_nish7 is one of three builders behind Higgs & Hrrm, giving Minecraft villagers voices, memories and gossip.
>> @BKJinAI founded Adventists AI, the team behind Sophon-Realtime, a full-duplex voice model that can laugh and sigh.
>> @_devp builds voice AI at Boardy and explores what an AI should show you while you talk.
>> @yiz_be_building leads Besimple, whose voice benchmarks test whether models catch the details a task depends on.
>> @rishflips built Twring so the coding agent you already run can ring your phone when it needs you.
>> @johnrobinsn built MacVoice, which runs the speech pipeline locally and uses Apple's echo cancellation so you can interrupt it on laptop speakers.

Keep exploring—your next experiment might start with something someone here is building.

#BackchannelSignals

Raise your wrist. Edit a list as you speak. Pause to apply.

@johnhartquist’s Apple Watch demo brings “raise to talk” to Jev (@typesafeai), inspired by @kunalabcdani’s voice coffee-ordering demo.

As @DeepgramAI Flux transcribes the audio, his server repeatedly asks Jev three questions in one request: what action, which existing item (if any), and what text? His design previews changes while you talk and applies them when you pause, with higher intent-confidence requirements for deletion.

To handle multiple items, he reuses @allietheicon’s shopping-cart pattern: ask again with what’s already been found, until Jev picks “nothing else.”

#BackchannelSignals

Boubou turns a laptop into a local voice server for the house.

Built by @Jinyu_Chai, @runboubou’s open-source Rust runtime is now in beta. Phones and robots on the same Wi-Fi can use models running on the host.

• Phone — the serve command gives you a QR code. Scan it and talk through your browser.
• Models — run Liquid’s LFM3-Audio-2B or LFM3-Omni-2B, or combine a Qwen LLM with Qwen3-TTS, Pocket TTS or Phonon.
• Robot or app — modular pipelines expose OpenAI-compatible Realtime and speech endpoints.

The useful separation: devices can become voice interfaces without running the models themselves. Each server supports one active voice session, with silence-based turns and barge-in.

Audio travels directly between device and host; inference stays local. The default setup uses a hosted HTTPS certificate helper for browser microphone access, which can be skipped with --self-signed or --offline.

MIT-licensed code; models retain their own licenses.

github.com/runboubou/boub…

#BackchannelSignals