German AI Voices

The sentence holds its meaning until the end

German brackets its clauses. A separable verb throws its prefix to the end of the sentence — Ich mache das Fenster auf — and a subordinate clause sends the finite verb there too, so a listener cannot know what happened until the last word arrives. Written German can afford this because a reader's eye moves ahead. A listener's ear cannot, and that is the single biggest difference between writing German for the page and writing it for audio.

The practical consequence is that Schachtelsätze, however elegant, are the enemy of a synthetic voice. Break the long ones. A voice that has to hold an intonation arc across four nested clauses will flatten somewhere in the middle, and the flattening is the moment the listener stops believing there is a person there.

Compound nouns need their seams found

Geschwindigkeitsbegrenzung is one word and four stresses, and reading it correctly means finding the internal boundaries before deciding where the emphasis goes. The engine does this well on established compounds and less well on fresh ones, which matters because German writers coin compounds freely and fiction coins them constantly. A newly invented compound is the German equivalent of an invented name in Japanese: the thing most likely to come out wrong.

There is no pronunciation dictionary, so where a coinage is mangled the lever is the text. Hyphenating a long compound — Drachen-Beschwörungs-Ritual — gives the engine the seams explicitly, and the listener hears no difference. The subtitles do show it: the SRT export on Plus and Pro follows your script text, so the hyphens appear there too. Use it sparingly and only where auditioning shows you need it.

Three standards, and no children

Ten of the twenty voices are German-language voices proper, and they cover Standard German as spoken in Germany, in Austria and in Switzerland. These are three standards rather than three accents, and they differ in vowel quality, in intonation and in vocabulary. It is worth saying clearly what the Swiss voices are not: Swiss Standard German is the formal register used in broadcast, not Schweizerdeutsch dialect, which is a different thing and is not available here at all.

The other ten are multilingual voices whose first language is not German. They speak it competently and they are what lets a twenty-character Hörspiel cast without repeats, but they are not a fallback reached only by large casts. Each character is matched into the part of the pitch range that suits its age, and wherever the German voices are thin in that band a multilingual voice is the first choice. Twenty-one of the forty library characters start on one in German, so a four-hander can land on one exactly as a full ensemble does. Audition per character rather than assuming a small cast is safe.

German also has no true children's voices. A child character goes to the highest-pitched adult voice of its gender — an Austrian man's voice for Milo, a Swiss woman's for the gender-neutral Pip — which reads as young at best rather than as a child. For a Kinderhörspiel that is the constraint to design around — a strong adult narrator and teenage rather than child protagonists is the version of the format that works today.