Talk to Your Agents — and Hear Them Answer in 12 Languages
Two small icons just landed on every agent card: a microphone and a speaker. They look simple, but between them they turn an agent from something you type at into something you can talk with — and listen to in a language of your choice.
Here’s what each one does and how it works under the hood.
The microphone: talk to an agent
Click the mic on an agent card and a voice chat opens. The flow is deliberately plain:
- Click the mic, speak, click stop. Your browser records a short clip (up to a minute) as 16 kHz audio.
- Your speech is transcribed. The audio goes to BotTrottr, which runs it through NVIDIA’s Nemotron speech-recognition model to get text.
- The agent answers like it always does. The transcript is handed to the agent as an ordinary task, so it uses its own model, system prompt, skills, and connected tools. Nothing about the agent changes — it just received its task by voice.
- The reply is spoken back. The answer is turned into speech by NVIDIA Magpie-TTS and played in the dialog, with the transcript and reply shown as text alongside.
Because step 3 is the agent’s normal run, voice works with every agent you’ve built — a CloudWatch agent, a research agent, a finance agent. You’re not talking to a separate “voice bot”; you’re talking to the same agent, through a different door. The dialog even shows timings for each stage (transcribe → think → speak) so you can see where the time goes.
The microphone is currently an admin preview while we tune latency and cost.
The speaker: hear every result read aloud
The speaker icon sits right next to the mic. It’s a toggle:
- Off — the agent behaves as before.
- On — the button looks pressed in, and every time the agent finishes a run, its result is read aloud automatically.
The setting is remembered per agent in your browser, so your daily-briefing agent can talk while your log-analysis agent stays quiet. Errors and the echoed task line are skipped — only the actual answer is spoken — and if you need to interrupt, there’s a Stop button right there.
Under the hood, the result text is sent to BotTrottr, cleaned up for speech (markdown and links stripped), and synthesized with NVIDIA Magpie-TTS Multilingual. Long results are read from the beginning, cut at a sentence boundary, so you hear the useful part without waiting through a wall of text.
Translate any result into 12 languages
Switch the speaker on and a “Speak in” dropdown appears under the card’s buttons. Leave it on As written and the agent’s text is read as-is, with the language auto-detected. Or pick one of the 12 supported languages:
English · Spanish · French · German · Italian · Vietnamese · Mandarin Chinese · Hindi · Japanese · Arabic · Portuguese (Brazil) · Korean
Pick one and BotTrottr does two steps for you:
- Translate the result into that language with NVIDIA Riva-Translate.
- Speak the translation using a native voice for that language.
That second part matters. A text-to-speech voice only controls how words sound — hand a Japanese voice English text and you get gibberish. So the dropdown translates first, then speaks. Change the language and the last result is re-read in the new one, which makes it easy to compare.
A few things we like about this:
- One agent, any audience. Build the agent once in English; let a colleague in São Paulo or Seoul hear the same result in their language.
- Nothing to configure. No translation agent to wire up, no language codes to type — it’s a dropdown.
- Works on any result. Reports, summaries, alerts, anything an agent writes.
Try it
The speaker and language picker are live in the BotTrottr demo — open an agent, flip the speaker on, choose a language, and run it. Signed-in admins get the same controls on their own agents; broader rollout is coming once we’ve watched it run at scale.
Under the hood, for the curious
| Piece | What it uses |
|---|---|
| Speech → text | NVIDIA Nemotron ASR |
| The agent’s reply | Whatever model the agent is set to use |
| Translation | NVIDIA Riva-Translate-4B |
| Text → speech | NVIDIA Magpie-TTS Multilingual (12 languages) |
Voice in, voice out, any language — on the agents you already have.