Dumping TTS frames before an audio track connects and killing the call while the buffer drains makes voice agents sound broken.
Wrote a post on how to start and end a Python voice agent call without clipping first or last words:dev.to/pmckeetx/start-and-en…
Sign up for the Dialog-RSN-1 API before November 19th, every developer who builds on it starts with @cartesia Sonic as the voice, and your first $100 is on us.
Under 300ms median first-token latency, plus the highest quality and safety scores of any real-time-capable model.
🔑 Get your API key: bit.ly/4knljLA
The tests run the real agent against a fake realtime server. A fake that speaks the holding line in the same order as the real server caught a hang when the agent failed before the line finished.
Code: github.com/pmckeetx/dialog-r…
An agent behind a tool took 4 to 5 seconds in my runs. That's dead air on a call.
Mark one tool argument "poly_spoken": true and the model writes a line the server speaks while the tool runs. First words in about half a second.
LangChain tool to realtime tool is one line:
{"type": "function", **convert_to_openai_function(tool)}
convert_to_openai_tool is the wrong one here. It nests everything under "function", and the Realtime API wants it flat.
Give your LangChain tools a voice without wiring up STT, turn detection and TTS yourself.
Dialog-RSN-1 hears the caller and writes the reply; the API voices it. LangChain tools and a create_agent do the work. ~270 lines of Python, with tests.
petermckee.dev/blog/build-a-…
“Hello? Is anyone there?” 🤔
That’s what callers say when a voice agent takes a beat too long. AI can write production code, but many agents still can’t tell when to speak on a live call.
Dialog-RSN-1 hears the caller’s audio and handles the whole turn in a single pass, so callers hear a reply in under 1 second. Independent benchmarking by Coval found it the strongest realtime model they’ve tested on production-style calls.
@pmckee, @Addlesee_AI, and @matthen2 take you inside the model in Part 1 of Let’s Dialog: Building the Next Voice Agents.
Grab your spot: bit.ly/4dVUc6B
A browser voice assistant in ~550 lines of JS. Mic audio streams to Dialog-RSN-1, which hears it directly, detects end of turn and streams the reply back as speech. No STT, LLM, TTS chain.
Ships as one <voice-assistant> custom element. Write-up:
petermckee.dev/blog/add-a-vo…
Vite 8 library mode bundles React 19, the CSS and the worklet (as a Blob URL) into one ES module, ~81 kB gzipped, in a Shadow DOM.
Code, plus a test site to drop it into:
github.com/pmckeetx/dialog-r…
Audio: an AudioWorklet turns float32 into 16-bit PCM in 20 ms chunks at the native rate, and the server resamples. Replies are 24 kHz PCM16, scheduled as back-to-back AudioBuffers.
On barge-in the server cancels the reply. The client just flushes its queue.
AI can write code, but on a live call it struggles to know when to speak.
Wed 4 Nov I'm hosting Matt Henderson and Angus Addlesee, who built Dialog-RSN-1, on turn-taking, barge-in and noisy lines.
12 pm ET: poly.ai/webinars/lets-dialog…
AI After Hours is for engineers, architects and product leads in London, in any industry.
You build a working agent at PolyAI HQ with our engineers at your table. Wed 14 Oct, 6 to 9 PM.
poly.ai/events/ai-after-hour…
AI After Hours, London, Wed 14 Oct, 6 to 9 PM.
Bring a laptop. Join a small team with a PolyAI engineer at the table and build a working agent for a bank or insurer in 90 minutes. AWS covers food and drinks.
poly.ai/events/ai-after-hour…
We've been running Dialog-RSN-1 in production for weeks
~250k live production ~2.5M model turns ~100k calls in the last 7 days ~50k on the peak day ~20 customer accounts
Speech-native. no transcript step, sub-300ms median model first-token.
poly.ai/blog/PolyAI-dialog-r…
A caller says "Matthew with two Ts." The ASR types "Matthew with to tease." Your LLM is done.
Dialog-RSN-1 skips the text bottleneck. Raw audio + conversation history + your prompt, so it hears what transcripts get wrong. Mood, noise, accents too.
poly.ai/dialog-rsn-1-early-a…
Most voice AI ends your turn with a silence timer. Pause to think and it talks over you. Dialog-RSN-1 moves that call into the model. Early Access: poly.ai/dialog-rsn-1-early-a…