Summary
When generating long-form audio (~10 minutes) in Single-Speaker TTS mode using an input prompt of ~9,580 characters, the voice fails to stay consistent. Every ~120 seconds (2 minutes), the voice noticeably shifts or switches to a completely different speaker identity.
Environment & Configuration
- Model: Gemini Flash TTS (e.g., gemini-2.5-flash / gemini-3.8-flash [bitte im Dropdown prüfen])
- Mode: Single-Speaker Audio Generation
- Voices tested: Kore, Rasalgethi
- Input length: ~9,580 characters (single continuous text prompt)
- Output audio length: ~10 minutes
Steps to Reproduce
- Select the Gemini Flash TTS model in Single-Speaker mode.
- Select either
KoreorRasalgethias the voice profile. - Input a continuous text script of approximately 9,580 characters and generate the audio in one pass.
- Listen to the generated audio around the 2:00, 4:00, 6:00, and 8:00 minute marks.
Expected Behavior
The selected voice (Kore or Rasalgethi) maintains a consistent timbre, pitch, pacing, and speaker identity across the entire ~10-minute audio file, as advertised for Single-Speaker TTS.
Actual Behavior
Approximately every 120 seconds, the voice abruptly changes or drifts into a completely different voice profile, despite being locked to a single speaker (Kore or Rasalgethi). It appears the backend is chunking the ~9,580-character input into 2-minute audio segments without carrying over the voice conditioning state between chunks.