Text-to-Speech and Voice Output
Give your product a voice: TTS models and voices, streaming audio, pronunciation control, and the consent rules around voice cloning.
Text-to-speech used to mean a flat robotic voice you'd only put up with for accessibility. Current TTS models sound natural enough that users often can't tell, which opens up real product uses: help articles people can listen to, voice replies in an assistant, audio versions of reports, spoken alerts on a factory floor. It also opens up real risks, because a model that can sound like anyone can be made to sound like someone specific.
In this lesson you'll add voice output to Deskmate: choose a TTS model and voice, stream audio to the browser so it starts playing almost immediately, rewrite screen text so it sounds right spoken aloud, control how product names and numbers are pronounced, and understand the consent rules around custom and cloned voices.
Where voice output earns its place
Before adding a speaker icon everywhere, decide which job voice output is doing. Each job has different latency and cost needs:
| Job | Example | Latency need | Generate when |
|---|---|---|---|
| Read-aloud | "Listen to this article" on a help page | Start within a second or two | Ahead of time, then cache |
| Voice replies | An assistant answering out loud | As fast as possible | On demand, streamed |
| Notifications | Spoken alerts, phone reminders | Low | On demand, short |
| Audio content | A weekly summary as a short audio briefing | None | In a background job |
The key distinction is static versus dynamic text. A help article changes rarely, so you can generate its audio once and serve it to thousands of listeners. An assistant's reply is different every time, so you pay for every generation and latency matters. Most products have both, and treating them the same is either slow or expensive.
Deskmate's first voice feature is read-aloud for knowledge-base articles. It helps customers with visual impairments or reading difficulties, people on their phones, and anyone following setup steps with their hands on the keyboard. It's static text, so it's also the cheapest place to start.