Skip to content

This repository demonstrates an ultra-low latency architecture orchestrating Rasa (Agent), Rime (TTS), and Deepgram (ASR). Features a custom Python orchestrator, real-time rich CLI dashboard, and CALM-based banking logic.

Notifications You must be signed in to change notification settings

RasaHQ/rasa-rime-voice-demo

Repository files navigation

Rasa Banner

🎁 Unwrap the Future: Voice Orchestration

A low-latency voice assistant architecture orchestrating Rasa, Rime, and Deepgram.

Hello Rasa Join Community Rasa Docs

This repository demonstrates an ultra-low latency architecture that orchestrates Rasa Pro (Agent), Rime (TTS), and Deepgram (ASR). It features a custom Python orchestrator, a real-time rich CLI dashboard, and CALM-based banking logic to achieve sub-second, human-like conversational turns.


🚀 The Architecture

We achieve fluid, human-like voice interaction by orchestrating best-in-class specialized services rather than using a monolithic black box.

  1. Input: User Audio (Simulated via Rime "Abbie" voice).
  2. ASR (Ears): Deepgram Nova-2 for lightning-fast speech-to-text.
  3. Brain: Rasa Pro executes deterministic business logic (CALM) to ensure secure money transfers.
  4. TTS (Voice): Rime Mist v2 generates high-fidelity agent audio with <300ms latency.
  5. Orchestrator: A custom Python client manages the real-time traffic to ensure immediate playback.

🛠️ Setup & Usage

Prerequisites

  • Python 3.10+
  • uv (for fast dependency management)

1. Installation

Install all dependencies into a strictly versioned virtual environment:

make install

2. Configuration

Create a .env file with your credentials:

# Required for Rasa Pro
RASA_LICENSE=your-rasa-pro-license-key-here

# Required for NLU/Reasoning
OPENAI_API_KEY=your-openai-api-key-here

# Required for the Voice Demo
DEEPGRAM_API_KEY=your-deepgram-api-key-here
RIME_API_KEY=your-rime-api-key-here

3. Generate Audio Assets

Instead of using a live microphone (which is risky for demos), we pre-generate high-quality user prompts using Rime:

make generate-audio

4. Train the Brain

Train the Rasa CALM model to handle the banking logic:

make train

🎤 Run the Flash Talk Demo

To run the full orchestration, open 3 separate terminal tabs:

Tab 1: The Action Server (Handles custom business logic)

make run-actions

Tab 2: The Agent (Rasa Pro Core)

make run-rasa

Tab 3: The Live Client (The Orchestrator & Visual Dashboard)

make demo

You will see a rich CLI dashboard and hear the conversation flow between the User and the Agent in real-time.



🚀 Two Ways to Get Started with Rasa

🤖 Start Building 🧠 Join the Conversation
Try Hello Rasa Agent Engineering Community
The fastest way to prototype. An interactive playground to build CALM (Conversational AI with Language Models) agents in your browser.

No setup required
No NLU training needed
Built-in Copilot
The home for people building real-world AI agents. A vendor-neutral space to discuss architectures, memory, and orchestration.

🤝 Meet other builders
🛠️ Share agent patterns
🎓 Learn form the best

Connect with us

About

This repository demonstrates an ultra-low latency architecture orchestrating Rasa (Agent), Rime (TTS), and Deepgram (ASR). Features a custom Python orchestrator, real-time rich CLI dashboard, and CALM-based banking logic.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published