Customer Service IVR Voice Pack

The delivery problem for phone-tree prompts

In a hotline, the painful part is not one welcome message. It is keeping welcome, main menu, hours, hold, transfer, queue, invalid key, after-hours, and error prompts in one consistent brand voice. Manually writing scripts, calling TTS, auditioning clips, and redoing one item can easily mix slot names, files, pronunciation, and billing. This skill addresses that problem by turning a switchboard-ready IVR voice pack into tracked prompt items with a constrained generation, review, and redo workflow.

How the skill works

The core inputs are the company or line name, menu branches, target language, and a pronunciation table whenever person names, product terms, or extension numbers appear. By default it creates about twelve prompts covering common phone-tree slots. The pack uses one voice_id for the whole set, and each prompt is synthesized once so a changed clip can be redone independently. The key steps are building a slot ledger, selecting a catalog voice, reading the live TTS information card, submitting each prompt through beatra.speech.synthesize, polling tasks to terminal state, and delivering slot IDs, real duration, MIME type, artifact URL, resolved model, and billing facts. Voice cloning is only used when the user explicitly authorizes it, and it requires balance or top-up confirmation before one submission and then freezes the voice_id.

Boundaries and cautions

It does not produce hold music, brand jingles, or general narration; use dedicated BGM or voiceover tools for those. Before any paid generation, confirm the voice, prompt count, weighted character count, and estimate. A changed text creates a new task ID, so updates are not free redos. If balance is insufficient, keep the top-up entry unchanged and retry only after the user confirms, using the same frozen ID.