DesktopAvatar: Mac app has AI agents' reports spoken aloud

The app grew out of a YouTube channel she opened in July, where she first used OpenAI's TTS with default voices. When Gemini 3.8 Flash TTS arrived, she rebuilt the channel around it with a Voice Design voice; soon after came news of Irodori-TTS being strong, then an ElevenLabs update. She judged Irodori the right pick for this use, despite its lower accuracy than ElevenLabs, because its Japanese reading was better than Gemini's in her view and it runs locally rather than per-use.

The build itself was largely handed to AI, and the code worked. What did not come free were the things only eyes and ears catch: mouths that stayed open, a voice that came out like an answering-machine announcement instead of the character, and a short line like "了解。" that ended with unexplained sound. Fixing them meant watching lip-sync, listening to output, and adjusting timing, seeds and text rather than chasing bugs.

Her own verdict is that she rereads reports less. For reports with real volume, she now has AI write them in HTML and has the voice read them from the top, which she feels costs the least mental load and time.