Azure LLM Speech returns empty transcripts with custom prompts and drops sentences in multilingual transcription

We are experiencing two separate issues with Azure Speech Fast Transcription in enhanced mode.

Configuration

Region: East US

API version: 2025-10-15

Enhanced mode: enabled

Task: transcribe

Languages: English and Spanish

Issue 1: Previously working requests return empty transcripts when a custom prompt is supplied

Starting 27 October 2026, previously working requests containing enhancedMode.prompt began returning HTTP 200 with empty transcription output:

{
  "combinedPhrases": [{"text": ""}],
  "phrases": []
}

The audio contains audible speech. Omitting the prompt restores nonempty transcription while enhanced mode remains enabled.

The behavior depends on the prompt content. Ordinary instructions such as Transcribe the audio. can produce empty output, while the exact instruction Output must be in lexical format. can still return text.

Expected behavior is to return the transcription when a supported prompt is supplied, or provide an actionable error if the prompt is invalid.

Issue 2: Multilingual transcription drops sentences, including with prompts disabled

Separately, mixed English and Spanish recordings can return incomplete transcripts. Sentences or speaker turns disappear from the middle of a conversation while surrounding speech is transcribed.

Spanish responses are particularly affected: English questions may remain while the corresponding Spanish answers are omitted. Some Spanish speech is also rendered in English despite the task being transcribe.

This occurs with the prompt field omitted, so disabling prompts does not fully resolve the multilingual quality problem.

Expected behavior is to preserve audible sentences, short answers and negations in their original spoken language throughout the recording.