Text-to-Speech (TTS).

Speech synthesis has existed for decades, but for a long time it sounded robotic. The leap came with neural models that learn from many hours of recordings how stress, pauses and intonation fit together. Today's systems take a text and return an audio file, often in selectable voices, languages and speaking styles. Many also provide per-word timestamps, which allow subtitles to be aligned precisely with the voice. The reverse process, turning speech into text, is called speech-to-text.

There are two modes of operation. Cloud services such as ElevenLabs or the speech services of the large cloud providers offer high quality without your own hardware, billed by characters or minutes. The text leaves your premises. Local models, available as open source on GitHub, run on your own computer or server: no ongoing fees per character, no data leaving the building, but setup effort and, depending on the model, a powerful graphics card.

The use cases for SMEs are concrete: a narrator voice for explainer and promotional videos, phone announcements and on-hold messages, a read-aloud feature for website copy, speech output in a chatbot or voice assistant, and audio versions of articles. TTS also matters for digital accessibility, although screen readers usually handle speech output themselves. Above all, a website must be cleanly structured so that these tools read it correctly.

In Switzerland the limits lie particularly in dialect. Good systems handle standard German very well, but Swiss German hardly at all or only in individual variants, and results are inconsistent. Proper names, place names and technical terms are also often stressed incorrectly and have to be corrected via phonetic spelling or adjusted spelling. And as soon as the voice of a real person is reproduced, voice cloning begins, with legal questions of its own. For the licence of the voice and model used: read the terms before commercial use.