OpenAI Voice Labs — learning path
This is the English learning index. It organizes the material into numbered modules in a hands-on workshop format: each module states its prerequisites, expected outcome, executable steps, checkpoints, and next destination.
If you only want an overview, return to the main README. If you want to build, start directly with either standalone tutorial.
How to use this path
Choose one of three approaches:
- Run and investigate: prepare the environment, run the finished solution, and study its engineering decisions.
- Build with guidance—recommended: start from a compiling starter branch, implement one slice at a time, and use versioned checkpoints to compare or recover.
- Rebuild from zero: use Path C in each lab and create directories and files in the documented order.
Each standalone tutorial includes the essential setup context: API project, billing, credentials, installation, and secret protection.
Read the workshop guide to choose a path, clone a starter, and inspect checkpoints without losing your work.
Modules
| Module | Tutorial | Starting point | Time | Verifiable outcome |
|---|---|---|---|---|
| 01 — Text to Speech | Open the complete tutorial | Starter branch | 2–3 h | Validated text returns playable, downloadable audio |
| 02 — Realtime Voice Agent | Open the complete tutorial | Starter branch | 3–4 h | A live conversation supports turns, mute, and interruption |
You can complete Module 02 without Module 01. Comparing them helps explain why a bounded TTS request and a live Realtime session need different architectures.
Module completion contract
Continue only when you can answer “yes” to all four questions:
- Outcome: does the described behavior work?
- Explanation: can I explain why this architecture was selected?
- Protection: is the API key still server-only and outside Git?
- Evidence: did the checkpoint command finish successfully?
Automated tests make no paid OpenAI calls. Manual voice tests are always explicit and should use short content, a controlled environment, and a monitored budget.
Checkpoints and finished solutions
Each lab keeps its finished implementation on main. The learning journey uses:
- behavioral checkpoints inside the tutorial;
- validation commands after each slice;
- tests as repeatable evidence;
- one compiling starter branch per lab;
- read-only
workshop/*-v1-step-*references for comparison and recovery; - Path C for rebuilding from an empty directory.
No workshop reference contains credentials, .env.local, node_modules, or build output. Breaking changes should create a v2 family while preserving the v1 edition used by existing classes and links.
Visual conventions
- Terminal: the exact directory and command to run;
- Expected outcome: what must appear before continuing;
- Checkpoint: an objective completion condition;
- Why: the engineering decision, not only syntax;
- Pitfall: a common error and how to recognize it;
- Production: a limitation the lab must not hide.
Start now
- Lab 01 — Text to Speech from zero to deployment.
- Lab 02 — Realtime Voice Agent from zero to deployment.
Workshop support
- Troubleshooting for environment, cache, imports, or Pages.
- Recording plan for real demonstrations without exposed credentials.
- UX and DevEx audit for the before-and-after evidence.