Gemini-3.5-transcribe drops the last 2–3 words when audio ends without trailing silence

Gemini API
1

Model: gemini-3.5-transcribe (Interactions API, REST, audio via Files API)
Modes: verbatim and smart (both affected identically)

Problem

For longer audio files (tested: 241 s and 260 s, German speech, MP3 mono 64 kbit/s 44.1 kHz, no decode errors) whose speech runs right up to the end of the file, the transcript is missing the final 2–3 words – the last sentence is cut mid-way. Both modes cut at exactly the same word. gemini-2.5-flash transcribes the same files completely.

What I checked

  • Reproducible: the same file cut at the same word in two separate runs.
  • Not a client parsing issue: the raw interaction contains a single model_output step with one text content item, status completed. The missing words are not anywhere in the response.
  • The whole audio was processed: input tokens ≈ 25 per second over the full duration (6500 audio tokens for 260 s).
  • Workaround: appending 2 s of silence (silent MP3 frames, original audio not re-encoded) makes the transcript complete.
  • A 21 s excerpt of the same file ending is transcribed completely, with or without trailing silence – so it seems to affect longer inputs only.

Interaction IDs (verbatim mode, 2026-10-31)

  • Original file, last words missing: v1_ChY4MGptYXVMbEZ1MnhrZFVQdElLRUlBEhY4MGptYXVMbEZ1MnhrZFVQdElLRUlB
  • Same file + 2 s silence, complete: v1_Chdfa2ptYXR5M080bmlrZFVQM0x2cG9RNBIXX2tqbWF0eTNPNG5pa2RVUDNMdnBvUTQ

Expected

The final words should be transcribed regardless of trailing silence.

Similar report for the live model: Bug : Gemini 3.5 Transcribe Live — final words frequently missing when stopping · Issue #372 · DevEmperor/DictateKeyboard · GitHub

2

Give me code and input here, i can try

3

Hello,

Thanks for sharing the detailed breakdown and interaction IDs, @linohype. Could you please share a non-sensitive ~4-minute sample MP3 file (ending without trailing silence) that triggers the missing words so we can test with that exact recording?