Hi! We use Gemini TTS models in a production pipeline generating fully voiced training scenarios (multiple characters, long scripts, different accents & languages). We’re aiming for a fully automated script-to-audio flow, and we’ve hit four recurring limitations. I’d love best-practice guidance where it’s a usage issue, and honest signal where it’s a model limitation (and whether upcoming releases address it):
1. Long-input cutoff. Long scripts stop generating after a few minutes of audio, forcing us to chunk. What are the actual limits, and is there a recommended long-form pattern?
2. Run-to-run voice inconsistency (our biggest blocker). Same voice name + same config produces an audibly different voice on each run. Chunked audio therefore doesn’t match, which breaks automation, every batch needs human review and re-runs. Is there any seed/voice-pinning mechanism, or one planned?
3. In-generation degradation. Within a single long generation, pacing gets slower and the accent drifts towards generic as the generation progresses. Known issue? Is there a “stable zone” input length you’d recommend?
4. Emotional ceiling. We need genuinely extreme registers, real shouting, a voice breaking, menacing whispers, stammering. Style instructions plateau well below that, and some inline tags get read aloud instead of interpreted. Is there a definitive list of supported tags per model version, and anything beyond short style instructions (reference audio conditioning, intensity parameters) now or on the roadmap?
Happy to share audio samples of each failure mode if useful.
Thanks!