Text Translation Service: Intermittent Hindi Word Insertion in French Output — Request for Escalation

Problem description

When translating English to French with the Translator text API in LLM mode (gpt-5.4-mini deployment on a Foundry resource), the output sometimes contains one Hindi word in otherwise correct French. Examples: "संकेत" for "signal", "प्रक्रिया" for "process", "समाधान" for "resolution". The request sets the source language to en and the target to fr, and the response reports "language": "fr". This looks like a bug in the LLM translation mode: standard NMT never produces it on the same text.

Environment

• Azure AI Translator, Foundry resource, region West US 3
• Deployment: gpt-5.4-mini
• API: POST /translator/text/translate?api-version=2026-06-06
• Direction: en to fr, tone formal

Reproduction

The same request (specified in the support case), sent 30 times each:

• Text with HTML tags, textType html: 5 failing
• Same text without tags, textType html: 1 failing
• Same text without tags, textType plain: 0 failing

A longer HTML document (about 6,000 characters) fails in more than half the requests, with or without the tone parameter.

Failing request IDs (UTC, 19 Oct 2026)

• 9aa75cbc-2a54-4cb2-8338-3e6a8c94292c.USW2.1019T0743, 07:43:06
• 3c1a24b8-81b8-4def-8796-d9d5397a15c6.USW2.1019T0743, 07:43:10
• ee826d03-c02c-4e52-8bc0-8babd03cbba5.USW2.1019T0743, 07:43:47 (no tags in the input)

Impact

The reliability of translations in LLM mode is not prod-OK (but we need it because NMT is not qualitative enough)

Request

Support case 2610190030002097 agreed the issue comes from the LLM translation path. We would like it escalated to the Azure engineering team, with a reference we can follow and an expected timeline for a fix.