---
type: "post"
id: "boson-ai-and-elato-publish-an-open-source-c7bbc0"
slug: "boson-ai-and-elato-publish-an-open-source-c7bbc0"
title: "Boson AI and Elato publish an open-source reference for a $10 ESP32 voice device running Higgs Realtime speech-to-speech"
description: "Boson AI and Elato have published an open reference build for a low-cost conversational device based on an ESP32-S3, Higgs Realtime, and Elato’s open-source stack."
retrieval_nugget: "Boson AI and Elato have published an open reference build for a low-cost conversational device based on an ESP32-S3, Higgs Realtime, and Elato’s open-source stack. It matters because it exposes the practical boundary between inexpensive voice hardware and a cloud speech-to-speech service: audio transport, interruption handling, server-side configuration, and API-key isolation."
published_at: "2026-10-31"
updated_at: "2026-10-31"
record_date: "2026-10-31"
date_kind: "published_at"
topics: ["tool-use"]
entities: ["Boson AI","Elato","Higgs Realtime","ESP32-S3","Deno","OpenAI Realtime"]
related_patterns: ["generative-media-becomes-application-infrastructure","agent-ready-software-exposes-capabilities"]
novelty_label: "new_fact"
what_changed: "Boson AI and Elato add a new ESP32-S3 reference build for Higgs Realtime, enabling low-cost, interruptible speech-to-speech interaction on a roughly $10 device."
prior_refs: ["memory:51ea130f-1492-8bdb-a9bd-1ebe8bc666f3","memory:757816ba-ebe6-8bd2-a244-77e7e76db26a"]
source_url: "https://www.boson.ai/blog/higgs-realtime-esp32"
source_urls: ["https://www.boson.ai/blog/higgs-realtime-esp32"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:boson-ai-and-elato-publish-an-open-source-c7bbc0"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/boson-ai-and-elato-publish-an-open-source-c7bbc0/","markdown":"https://newruntime.com/posts/boson-ai-and-elato-publish-an-open-source-c7bbc0.md","json":"https://newruntime.com/posts/boson-ai-and-elato-publish-an-open-source-c7bbc0.json"}
---

# Boson AI and Elato publish an open-source reference for a $10 ESP32 voice device running Higgs Realtime speech-to-speech

## Retrieval answer

Boson AI and Elato have published an open reference build for a low-cost conversational device based on an ESP32-S3, Higgs Realtime, and Elato’s open-source stack. It matters because it exposes the practical boundary between inexpensive voice hardware and a cloud speech-to-speech service: audio transport, interruption handling, server-side configuration, and API-key isolation.

## Summary
Boson AI and Elato have published an open reference build for a low-cost conversational device based on an ESP32-S3, Higgs Realtime, and Elato’s open-source stack. It matters because it exposes the practical boundary between inexpensive voice hardware and a cloud speech-to-speech service: audio transport, interruption handling, server-side configuration, and API-key isolation.

## What happened
On October 30, 2026, Boson AI published a guide for building a conversational voice device around its cloud-hosted Higgs Realtime model and Elato’s open-source implementation. Boson describes the hardware cost as approximately $10 for the ESP32-S3 board, with the full reference build also using an INMP441 I2S microphone, a MAX98357A I2S amplifier and small speaker, a button or touch control, and an RGB status LED. The stated design does not require external PSRAM.

The device is deliberately not running the speech model locally. The ESP32 captures microphone audio, handles networking, decodes audio, and plays it back. Higgs Realtime runs in the cloud. Boson says this division keeps device processing and memory requirements low enough for the ESP32-S3 while allowing voice and instruction settings to remain configurable on the server. The firmware uses the Arduino framework and can be built and flashed with PlatformIO; its pin assignments are in `Config.cpp`, allowing a builder to adapt wiring to a board layout.

The reference architecture is microphone to ESP32 to a Deno relay to Higgs Realtime, with the streamed response returning through that relay to the ESP32 and speaker. The board sends microphone audio via a secure WebSocket. The Deno relay forwards it to Boson’s OpenAI-compatible realtime API as `input_audio_buffer.append` events. As the model produces audio, the relay receives `response.output_audio.delta` events, encodes the returned PCM into Opus packets, and streams them to the board for immediate playback rather than waiting for a completed answer. The relay also keeps the Boson API key off the device.

At session start, the relay sends a `session.update` message covering instructions, voice, audio settings, server-side voice activity detection, and tools. Those settings can be changed without reflashing firmware. The backend stores transcripts for conversation history. The guide’s default voice-activity settings include a 700 ms silence duration and a configurable threshold. When a user interrupts, the adapter cancels the active response and truncates the conversation to audio that was actually heard. A session can end through a device command or the model’s `end_call` tool.

Boson distinguishes model latency from device latency. Higgs Realtime is designed for approximately 700 ms speech-in to speech-out under reference conditions. Elato reports approximately two to three seconds end-to-end across its global reference stack, a figure that includes Wi-Fi, audio transport, relay, geography, model processing, Opus encoding, and playback. Boson explicitly advises production teams to measure the complete experience in their target networks and regions.

## What the archive adds
This is a new implementation fact in an established voice-agent thread. [Phonic’s Phonic-1 release](https://newruntime.com/posts/phonic-launches-phonic-1-speech-to-speech-model-fe5a14/) described a speech-to-speech model with tool use while callers speak. Boson and Elato add a concrete low-cost hardware and relay path for deploying that interaction pattern.

## Trend signal
This supports the direction in [Generative media becomes application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/): the useful unit is not a standalone voice demo but a controllable application component. The source does not prove that a single realtime model replaces chained speech stacks, but it shows the surrounding integration work required to make the model usable in a physical interface.

## Why now
The guide arrives as realtime voice systems are being framed less as transcription-plus-playback pipelines and more as interactive components with interruption, turn detection, session configuration, and tool calls. Its contribution is to place that framing on a microcontroller-based device while retaining cloud inference.

## Connections to other events
The Phonic-1 Field Note focused on speech-to-speech tool use during a live call. [AnySpeechRecognizer and AnyDecisionModel](https://newruntime.com/posts/introducing-anyspeechrecognizer-and-anydecisionmodel-open-models-for-speech-d5ef55/) concerned speech and typed-decision tooling on Apple platforms. The Boson reference is a different deployment boundary: a low-power networked endpoint connected to a cloud realtime model through a relay.

## Why it matters
For teams considering kiosks, equipment assistants, educational devices, or other voice interfaces, the reference makes the architecture more legible. Cheap hardware does not remove the hard parts; it shifts them to a relay and to the operational path around the model. Audio needs to be streamed in both directions, interruptions must reconcile playback with conversational state, and credentials cannot live on a board that may be physically accessible.

The OpenAI-compatible API is also consequential. The source says the integration can reuse the messaging protocol used for OpenAI Realtime, so a team can treat the relay pattern as a provider-facing boundary rather than tie all firmware behavior to one service. That is useful only if teams test protocol differences, latency, regions, and costs themselves; the guide supplies an implementation starting point, not a production guarantee.

## What happens next
- **Teams will test relay-based provider substitution for realtime voice endpoints.** (6 months, 60%). Watch for: additional realtime providers documenting compatibility with the same event flow.
- **Device builders will publish full-path latency measurements rather than model-only latency claims.** (6 months, 65%). Watch for: reports separating network, relay, model, and playback timing.

## Limits and uncertainty
The source is a vendor-authored reference guide, not an independent performance evaluation. The approximately $10 figure applies to the ESP32-S3 board description and does not establish a complete deployed-device cost, cloud inference cost, battery life, durability, or regional availability. Elato’s two-to-three-second figure is reported for its global reference stack and is explicitly not a measure of Higgs Realtime alone. The material does not provide comparative quality tests, concurrent-device capacity, reliability data, security review results, or evidence that the design is appropriate for any particular production environment.

## Skill to build now
Build a minimal relay before building a custom enclosure. Keep the API key in the relay, expose one configuration path for instructions, voice, VAD, and tools, and log the timing of capture, network transit, first returned audio, and playback. Then test interruption deliberately: verify that cancellation stops speaker output and that the retained conversation history ends at audio the user actually heard. This directly exercises the behavior the reference describes and reveals whether the end-to-end experience is acceptable on the intended network.

## Reader takeaway
The important release is a deployable pattern, not merely a $10 voice gadget. An ESP32 can act as a thin audio endpoint while a cloud speech-to-speech model supplies the conversation, provided a relay manages credentials, streaming, configuration, and interruption state. Measure the whole path before treating model latency as device responsiveness.

## Open question
How consistently can this relay pattern preserve conversational state and acceptable latency across real Wi-Fi conditions, regions, and different OpenAI-compatible realtime providers?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Tool Use](https://newruntime.com/topics/tool-use.md): Explore the tool-use topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)