So a few months ago, I used gpt°hard word-4o-mini-tts°hard word to make audio°hard word for Trakaido°hard word Spanish°hard word. This week, I wanted to fill in a few new words. But the newer version°hard word of that model ... was much worse. About 15% of audio°hard word was unusable°hard word, compared to 1% before.
No, the old model can’t be used. The undated°hard word gpt°hard word-4o-mini-tts°hard word alias°hard word that the Sep°hard word 1 batch°hard word called now points to gpt°hard word-4o-mini-tts°hard word-2025-12-15, which is the snapshot°hard word that produced today’s repeats°hard word. The older gpt°hard word-4o-mini-tts°hard word-2025-03-20 was shut down on July 23, 2026. Other developers°hard word have reported that the 12-15 snapshot°hard word ignores°hard word TTS°hard word instructions°hard word, which fits the repeats°hard word we saw. So nova°hard word Mar°hard wordía is gone, and redoing°hard word all es°hard word audio°hard word with Realtime°hard word is the right call.
Fortunately°hard word, it's only about $2 to regenerate°hard word all the audio°hard word. If I had done°hard word $500 of audio°hard word in the old system, I would be much more unhappy°hard word.
In general, "audio°hard word" is a relative°hard word weakness°hard word of LLM°hard word systems. Other than the influence°hard word of LONDON, I'm not sure why ...