Hello,
We use Gemini TTS (Gemini API, model gemini-3.8-flash-lite-tts) to voice lines from stage plays, so that actors can rehearse their parts.
About 1% of lines are refused with HTTP 400 content_blocked: “Request blocked for an unspecified policy reason. Please modify your input and retry.” Long, well-known speeches are refused far more often: about 1 in 6 in our tests. The refusals are not deterministic: the same line sometimes passes on an identical retry. The refused texts contain nothing harmful. They are classic stage speeches, often public domain, like the three below.
Example 1: Gotthold Ephraim Lessing, Minna von Barnhelm (1767) (public domain). Model gemini-3.8-flash-lite-tts, voice Leda, language de-DE, no style, refused twice in a row:
Wer kann denn in den verzweifelten großen Städten schlafen? Die Karossen, die Nachtwächter, die Trommeln, die Katzen, die Korporals – das hört nicht auf zu rasseln, zu schreien, zu wirbeln, zu mauen, zu fluchen; gerade, als ob die Nacht zu nichts weniger wäre als zur Ruhe. – Eine Tasse Tee, gnädiges Fräulein? –
Example 2: William Shakespeare, Twelfth Night (c. 1601), act 3 scene 4 (public domain). Model gemini-3.8-flash-lite-tts, voice Autonoe, language en-US, no style, refused twice in a row:
That defence thou hast, betake thee to’t. Of what nature the wrongs are thou hast done him, I know not, but thy intercepter, full of despite, bloody as the hunter, attends thee at the orchard end. Dismount thy tuck, be yare in thy preparation, for thy assailant is quick, skilful, and deadly.
Example 3: Luigi Pirandello, La giara (1917) (public domain). Model gemini-3.8-flash-lite-tts, voice Sulafat, language it-IT, no style, refused twice in a row:
È bianca. E me l’insegnò mio padre a farla bianca. Riconoscerete la sua potenza quando ci starete a bollire in mezzo. Ma laggiú è nera. Se accosti due dita, non le stacchi piú; e se t’attacco il labbro col naso, resti abissino per tutta la vita.
Possibly related: Incorrect PROHIBITED_CONTENT classification (a public-domain Hindi sentence blocked in TTS).
Questions:
- Why are these refused? Is there a recitation or copyright-style check on well-known texts? If so, how should public-domain texts, and scripts the user holds the rights to, be handled?
- Is there a supported way to avoid these refusals when voicing scripts for rehearsal (an attestation, a setting, another model)?
- May we retry a refused line split into its sentences, with the same words in the same order and nothing added or removed, as separate requests? The error message says “modify your input and retry”, and we want to be sure this is within the terms before we do it.
Thank you!