Gradium Examples

Small, standalone demo apps built directly on the Gradium APIs.

---

This is a collection of example apps built on Gradium's streaming speech-to-text, text-to-speech, and hosted LLM APIs - used directly, without any framework in between.

If you want a batteries-included voice-agent framework instead, see Gradbot. These examples are for learning the raw APIs and copying patterns into your own app.

Each folder is self-contained: its own README, dependencies, and run instructions.

Examples

Example	What it does
genz-translator	Speak normally and read your words rewritten into GenZ slang, live and karaoke-style. Gradium streaming STT → LLM rewrite → tokens streamed to the browser.
tts-latency-race	Race six streaming Text-to-Speech models from Gradium, ElevenLabs, Cartesia and Fish Audio on the same sentence, fired at the same instant. One waveform per model and time to first audio measured the way a listener hears it.

Agent skills

The skills directory contains reusable instructions for coding agents rather than standalone demo applications. Install a skill in Codex or Claude Code, describe what you want, and let the coding agent generate and test the project in your workspace.

Skill	What it builds
gradium-live-avatar-agent	A minimal live avatar voice agent using Gradium Voice Design, STT, and TTS with LiveKit orchestration and a LemonSlice animated face.
gradium-pruna-designed-avatar	A reviewed, non-realtime talking-avatar video using a bespoke Gradium voice and Pruna p-video-avatar.

Gradium Live Avatar Agent

This skill turns three creative inputs into an extensible realtime agent:

1. A reference image for the avatar.
2. A description of the voice that should accompany the image.
3. The agent's purpose and behavior, such as a language teacher, coach, tutor, concierge, or game character.

The generated project uses:

• Gradium Voice Design to create and audition an original voice;
• Gradium STT and TTS for the live speech pipeline;
• LiveKit for realtime orchestration, room access, and agent dispatch;
• LemonSlice to animate the supplied portrait and publish synchronized audio and video.

The result is deliberately minimal: one worker, one avatar-only call surface, server-side configuration, focused tests, and a clear extension point for tools or function calling.

Install it for Codex:

mkdir -p ~/.codex/skills
cp -R skills/gradium-live-avatar-agent ~/.codex/skills/

Or install it for Claude Code:

mkdir -p ~/.claude/skills
cp -R skills/gradium-live-avatar-agent ~/.claude/skills/

Then provide the three creative inputs in one prompt:

Use $gradium-live-avatar-agent to build a live avatar agent.

Reference image: ./portrait.png
Voice: A warm, low voice with a lightly textured tone, measured pacing,
and calm confidence. English.
Agent purpose: A language coach who helps adults practise conversational
English with concise corrections and encouraging follow-up questions.

If an input is missing, the skill asks for it. After the creative direction is settled, it asks about LiveKit Inference credits or a custom LLM, explains where to configure credentials, generates one Gradium voice candidate, and waits for approval before promoting that voice and completing the agent.

You need server-side Gradium and LemonSlice API keys plus LiveKit credentials. Configure them in the generated project's environment file or deployment secret store; never paste credentials into the prompt or expose them to browser code.

Gradium + Pruna Designed Avatar

This skill creates a polished talking-avatar video from three creative inputs:

1. A portrait image for the avatar.
2. A description of the original voice to design and audition.
3. The exact spoken script, or the clip's goal and audience if you want help writing it.

Gradium Voice Design creates the approved voice, Gradium TTS renders the final speech audio, and Pruna p-video-avatar animates the portrait from that audio. The skill checks the finished MP4 for duration, audio, speech accuracy, lip movement, face stability, framing, and visual artifacts. Unlike the LiveKit and LemonSlice skill above, this workflow renders a video asynchronously; it does not create a live or interruptible conversation.

Install it for Codex:

mkdir -p ~/.codex/skills
cp -R skills/gradium-pruna-designed-avatar ~/.codex/skills/

Or install it for Claude Code:

mkdir -p ~/.claude/skills
cp -R skills/gradium-pruna-designed-avatar ~/.claude/skills/

Then provide the creative inputs in one prompt:

Use $gradium-pruna-designed-avatar to create a talking-avatar video.

Portrait: ./portrait.png
Voice: A warm, softly textured voice with measured pacing and calm confidence.
Language: English.
Script: Welcome. In the next minute, I will walk you through the three ideas
that matter most.

The skill creates one Gradium voice candidate, asks you to approve its audition, then promotes that voice and submits one Pruna render. Both operations can use paid credits, so it does not generate extra candidates or rerender without your approval. Configure GRADIUM_API_KEY and PRUNA_API_KEY in your environment or secret store; never put either key in a prompt, client-side bundle, or committed file.

Getting started

Every example needs a Gradium API key:

export GRADIUM_API_KEY=your_gradium_key

Then pick an example, cd into it, and follow its README. Most use uv:

cd genz-translator
uv sync
uv run uvicorn app:app --port 8402

License

MIT