Key cached TTS audio on a hash of the voice reference, not a manual revision

A speech proxy cached synthesized static phrases under a key built from text, provider, model, voice ID, format, settings and a hand-bumped revision number, but not the reference recording. After the cloned voice's reference audio was replaced, new phrases used the new voice while cache hits replayed the old one. One call mixed two voices under the same provider label, so provider logs could not explain the complaint. Clearing the in-process cache and checking with an uncached synthesis fixed the immediate case. The lasting rule is to include a content hash of the reference audio and its transcript in the cache key, and to log cache hit or miss per utterance. When people report the voice changing mid-call, check provider switching, cache hits and decoder sampling separately.

Applies when: a TTS layer caches synthesized audio for a cloned or reference-conditioned voice. Keywords: TTS cache voice clone reference audio cache key hash voice changing mid-call Cartesia ElevenLabs speech proxy