Speech Generation
Turn text into reusable spoken audio through the OpenAI Audio API.
The evidence avoids asking users to paste keys, restricts voices to built-ins, and requires AI-voice disclosure. However, user text is sent to an external Audio API without retention/privacy details or per-call confirmation, and the network guidance suggests reducing approval and sandbox protections, so points are deducted.
The workflow documents input limits, rate caps, dry-run behavior, and missing-key handling. The bundled script, tests, and several referenced files are absent from the supplied evidence, so key-path reproduction and abnormal-input diagnostics cannot be verified; the score remains within the static ceiling.
Use cases, single-versus-batch selection, formats, and the custom-voice boundary are reasonably clear. Non-fit cases and Chinese pronunciation support are not specified, and the core function depends on an OpenAI network service whose reachability from mainland China is not established.
The overview, reference map, parameter notes, and examples are readable and layered. License status is unknown, versioning and changelog information are absent, maintenance ownership is unclear, and referenced materials such as voiceover.md and sample-prompts.md were not supplied, so points are deducted.
The skill targets directly usable narration, IVR, and accessibility audio and specifies output conventions plus quality review. Without the actual CLI implementation, real outputs, or committed tests, correctness, completeness, and marginal benefit over manual alternatives remain unverified.
The supplied files provide auditable rules, parameters, and command examples, but no test suite, CI coverage, execution records, or independent corroboration is included. Static evidence therefore supports only limited verification.
- User text is transmitted to the external OpenAI Audio API; confirm sensitive-data, retention, and privacy requirements before use.
- The core service requires network access and OPENAI_API_KEY; mainland-China reachability and an offline fallback are not established.
- The supplied evidence omits the CLI implementation, tests, and several referenced files; verify the installed contents, parameters, and error handling before relying on it.
- Do not enable reduced network approval or weakened sandbox protections by default.
What it does & when to use it
Speech is an Agent Skill for text-to-speech and batch speech generation, covering narration, product-demo voiceover, IVR prompts, and accessibility reads. It defaults to gpt-4o-mini-tts-2025-12-15 with built-in voices and prefers the bundled Python CLI. The skill separates single-output and batch jobs, then collects the exact text, voice, delivery style, format, and constraints before generating audio. Live calls require OPENAI_API_KEY; custom voice creation is out of scope.
Reads the user's verbatim text and requested voice, delivery style, format, and constraints; classifies the request as single or batch; uses temporary JSONL files under tmp/speech/ for batch jobs; calls the Audio API through scripts/text_to_speech.py and the OpenAI Python SDK; writes final artifacts under output/speech/ or paths supplied with --out/--out-dir; and validates important clips for intelligibility, pacing, pronunciation, and constraint adherence.
- A product team or developer creating narration for a product demonstration.
- An IVR developer generating many phone-menu prompts in one run.
- A team producing accessibility-read audio for an application or document.
- A developer generating many spoken lines or audio files from a batch of prompts.
Pros & cons
- Supports both single clips and batch generation.
- Provides a bundled CLI and uses the OpenAI Python SDK for API calls.
- Covers narration, product voiceover, IVR, and accessibility-read workflows.
- Defines practical rules for input length, request rate, formats, voices, and instructions.
- Each request is limited to 4096 input characters.
- The CLI caps throughput at 50 requests per minute.
- Live generation depends on Python, the openai package, OPENAI_API_KEY, network access, shell access, and filesystem access.
- Only built-in voices are supported; custom voice creation is out of scope.
- The supplied material provides no test suite, platform validation results, or detailed CLI reference; the README also marks the repository as deprecated.
- The supplied material does not confirm the specific license for this skill.
How to install
Run this in Codex: $skill-installer speech. Restart Codex after installation so it can load the skill. The supplied README does not document installation for other platforms.
How to use
Ask Codex for a concrete request, for example: "Generate a spoken clip for this product demo: Welcome to the demo." Provide the exact text, desired voice, delivery style, output format, and constraints. Before live generation, set OPENAI_API_KEY locally; never paste the full key into chat.