Experiencing the same problem here.
We hit the same finish-reason and truncation symptoms running gemini-3.1-flash-tts-preview for non-conversational audio tour generation (single speaker, no multi-speaker config). Spent the last week or so quantifying it across endpoints — sharing the data in case it’s useful for the team triaging this.
Setup: 10 real production tour scripts, 1421–1705 chars each, voice = Charon. Single-input synthesis, not multi-speaker. Same SDK config across runs (timeout=90, retry=None).
Results across the four endpoints that accept this model:
endpoint: aiplatform.googleapis.com/.../streamGenerateContent (SSE)
success rate: 80% full audio, 20% silently short
finish reasons on success: mix of STOP and OTHER
notes: OP’s “OTHER on truncation” 100% reproduces here\
────────────────────────────────────────
endpoint: texttospeech.googleapis.com/.../streamingSynthesize (Cloud TTS bidi)
success rate: 3/10 to 5/10 across re-runs
finish reasons on success: STOP when works
notes: rest fail with 400 InvalidArgument: violates Vertex AI’s usage guidelines, support code 54702341 =
Unspecified
────────────────────────────────────────
Is there any way we can reach out to the team or get updates on whether this is being worked on and we should wait for a fix?