I built a voice chat with two talking dogs

I built a real-time voice chat with two talking dogs. It is live for everyone on thedoodlecast.com.

For the last few months I've been running a side project called The Doodle Cast — an AI-generated video podcast starring two animated dogs, Rusty and Oreo. Rusty is the deadpan operator; Oreo is chaos with a golden coat. It's a sandbox where I get to push pipelines for image-to-video generation, dialogue writing with LLMs, and voice synthesis into a production workflow.

The latest addition is something I've wanted for a while: a live voice chat where you can actually talk to one of the dogs, out loud, and hear them talk back in character — in real time.

What it actually does

You open the link on your phone or laptop, tap the mic, and speak. The page captures your voice through the browser's built-in speech recognition, sends the transcript to a local LLM that's been primed with each character's personality and show bible, streams the response back, and pipes it into a TTS model that's been cloned to match each dog's voice from hours of existing Doodle Cast dialogue. Round-trip latency is good enough that it feels like a real conversation — not a walkie-talkie.

Each character has a distinct profile:

• Rusty — measured, strategic, quietly judgmental. Will answer honestly but treat most questions like a tactical briefing.
• Oreo — enthusiastic to the point of malfunction. Do not mention squirrels, the vacuum, or anything involving water unless you want a meltdown.

The stack, briefly

It's deliberately modest — no Realtime APIs, no WebRTC gymnastics. The whole thing is held together with:

• The browser's built-in SpeechRecognition for input (Chrome, Safari/iOS, Edge).
• A locally hosted LLM behind a streamed SSE endpoint on the Doodle Cast API, tuned with the character's system prompt on every turn.
• ElevenLabs Flash v2.5 for the voice — it's fast enough for conversational turn-taking and the voice clones hold up well over short lines.
• Playback through the Web Audio API (decodeAudioData + BufferSource) instead of <audio> elements, because iOS Safari really does not enjoy the handoff between microphone recording and audio playback, and the buffer-source path survives it.
• Session scoping and rate limiting so the compute budget stays sane.

It started out invite-only while I watched how people actually used it. It held up, so it's now open to everyone, live on thedoodlecast.com. Every conversation still costs a little CPU and a little TTS budget, which is what the rate limiting is for.

Try it now

No application needed anymore: open the cast chat on thedoodlecast.com, pick a character, tap the mic, and just talk. No account, no tracking beyond what's needed for rate limiting.

What to expect

1. Works on an iPhone, iPad, or a Chrome/Edge desktop.
2. Pick Rusty or Oreo, tap the mic, and just talk. They answer in character, in their own voices.
3. Conversations are rate-limited so the budget stays sane; if it queues for a moment, that is the home GPU thinking.

If something goes sideways (the mic won't unlock, a response freezes, audio cuts out mid-sentence), I genuinely want to hear about it. This is still early, and iPhones are where most of the interesting edge cases live.