Speaker tag [Sxx] sometimes missing on continuous/single-speaker audio, causing full segment collapse

Summary

On some audio clips, the model returns valid transcribed text but omits the [Sxx] speaker tag entirely, even though the documented output format is [start_time][Sxx]text[end_time]. Observed on continuous single-speaker Arabic audio (monologue/lecture style, clear audio, no background noise/music). Other single-speaker (English) and multi-speaker (Arabic) files consistently return correct [Sxx] tags.

Reproduction

vLLM, fp8 quantized (vllm serve OpenMOSS-Team/MOSS-Transcribe-Diarize --trust-remote-code):

curl -X POST http://localhost:8055/v1/audio/transcriptions
-F model="OpenMOSS-Team/MOSS-Transcribe-Diarize"
-F file=@"sample_52s_arabic.wav"
-F response_format="json"
-F temperature="0"
-F max_completion_tokens="65536"

Response (no speaker tags):
{"text":"[0.00]القارئين معك[2.00][2.00]زيضلك من يوم ولدك[3.88]...[51.28]"}

Expected:
[0.00][S01]القارئين معك[2.00]...

Ruled out

• Not audio quality: clip is clean speech, no music/noise/reverb.
• Not speaker count: a separate single-speaker English clip tags correctly.
• Not language: a separate multi-speaker Arabic clip tags correctly.
• Not prompt-related: passing the documented diarization prompt explicitly gives byte-identical output to not passing it.
• Not a one-off: re-running gives slightly different transcribed text (temp=0, fp8) but still omits tags both times.
• No silence gaps to key off either — each segment's end time exactly equals the next one's start time.

Question

Is this a confidence-based fallback where the diarization head skips tagging rather than risk a wrong one? Is there a way to force tagging, or a confidence signal we can check client-side?

Happy to share the audio file privately for reproduction.

Start by reproducing the issue with the documented vLLM command, the /v1/audio/transcriptions endpoint, and sample_52s_arabic.wav, then compare the output with and without the diarization prompt. No repository files or tests are named; done would require identifying whether missing [Sxx] tags are an intended fallback and documenting or exposing a way to force tagging or inspect confidence.