speech
by openaiUse the speech skill to turn text into spoken audio for narration, voiceover, IVR prompts, accessibility reads, and batch speech generation. It uses the OpenAI Audio API with built-in voices, a bundled CLI, and `OPENAI_API_KEY` for live runs. Custom voice creation is out of scope.
This skill scores 88/100, which means it is a solid directory listing with good practical value for agents. Users should expect a clearly triggerable speech-generation workflow that is more actionable than a generic prompt, with enough CLI and reference detail to support real installs, though it still depends on network access and the OpenAI API for live output.
- Strong triggerability: the frontmatter explicitly scopes use cases like text-to-speech narration, voiceover, accessibility reads, and batch speech generation.
- Operationally clear: SKILL.md provides a decision tree for single vs. batch and a step-by-step workflow, backed by a bundled CLI reference.
- Good agent leverage: supporting references cover voices, audio API parameters, accessibility defaults, and batch usage, reducing guesswork for execution.
- Live generation requires `OPENAI_API_KEY` and network access, so it is not fully self-contained for offline use.
- Custom voice creation is out of scope, so users needing bespoke voices or advanced audio workflows will need something else.
Overview of speech skill
What the speech skill does
The speech skill turns text into spoken audio for narration, voiceover, IVR prompts, accessibility reads, and batch speech generation. It is best when you need reproducible audio output from a prompt, not a freeform “make it sound nice” request.
Who should use it
Use speech if you need the speech install to fit a real workflow: product demos, app onboarding, accessibility assets, or many short clips from structured text. It is a strong match when you care about voice choice, pacing, output format, and consistent generation across runs.
What makes it different
The speech guide is built around the OpenAI Audio API and the bundled CLI, so it favors deterministic use over ad hoc prompting. It uses built-in voices, supports single or batch jobs, and expects OPENAI_API_KEY for live generation. Custom voice creation is out of scope.
How to Use speech skill
Install and locate the workflow
Install with npx skills add openai/skills --skill speech. After that, read SKILL.md first, then references/cli.md for command details, references/audio-api.md for model and parameter limits, and references/prompting.md or references/voice-directions.md for better instruction writing. For quick context, check agents/openai.yaml and references/sample-prompts.md.
Turn a rough goal into a usable prompt
The speech usage pattern works best when you give the skill the exact text to read, the target voice, the delivery style, output format, and any pronunciation constraints. A strong request looks like: “Generate a 45-second product demo voiceover from this script, use cedar, keep it warm and steady, output mp3, and emphasize the product name on first mention.” That is better than “make this sound professional,” because it gives the skill concrete synthesis controls.
Single vs batch workflow
The skill is designed for two paths: one clip or many clips. If you have multiple lines, prompts, or files, treat it as batch and prepare a temporary JSONL file under tmp/, then run the CLI once and delete the JSONL after use. If you have one script, use the single-file path. This decision matters because the skill’s structure and validation steps change with output volume.
What to check before you run
For best results, verify the text verbatim, not just the theme. Confirm the voice, file format, speed, and whether the output must be neutral, expressive, or accessibility-first. The main repository file to inspect for execution is scripts/text_to_speech.py; do not modify it unless the repository maintainer instructs you to.
speech skill FAQ
Is the speech skill only for narration?
No. The speech skill also fits voiceover, accessibility reads, IVR prompts, and short audio prompts. It is less useful for custom voice cloning or creative voice design, which this repo does not cover.
Do I need the CLI to use speech?
For reliable speech usage, yes. The bundled CLI is the intended path for live generation, while --dry-run is useful for checking invocation shape without making an API call. If you only write a generic prompt, you lose the structure that makes the skill reproducible.
Is this beginner friendly?
Yes, if you can provide the exact text and a basic voice direction. The speech install is simple, but the output quality depends on how clearly you define pacing, tone, format, and pronunciation. Beginners usually succeed faster when they start with a short clip and one voice.
When should I not use this skill?
Do not use speech if you need custom voice creation, heavy post-production, or a workflow that depends on modifying the bundled script. It is also a poor fit if you cannot use networked OpenAI API calls or do not have an OPENAI_API_KEY.
How to Improve speech skill
Give the skill fewer ambiguities
The biggest quality gain in speech skill output comes from removing guesswork. Provide the exact text, not a summary; name the intended listener; and specify whether the read should sound like narration, support messaging, accessibility, or an IVR prompt. If a term is hard to pronounce, spell it out or add a pronunciation note.
Tune one variable at a time
When the first pass is close but not right, change only one thing: voice, speed, or instruction style. That makes iteration cleaner than rewriting the whole prompt. For example, if the timing feels rushed, keep the text and voice fixed and adjust only the speed from 1.0 to 0.95.
Use output constraints that matter
The speech guide works better when constraints are operational, not vague. Say “mp3 for quick playback,” “wav for review,” or “steady and neutral for accessibility.” For batch jobs, keep each line narrowly scoped so the skill can preserve consistent delivery across outputs.
Read the right references first
If you want better results from speech for Design Implementation, prioritize references/accessibility.md for neutral reads, references/voiceover.md for presentation-style delivery, and references/sample-prompts.md for prompt shape. These files help you write instructions that the CLI and API can execute without extra interpretation.