Implement text analysis solutions 4.2 Implement speech solutions
A company has recorded customer calls in Japanese, Korean, and German.
Study guide skill: Translate speech into other languages by using language models and Foundry Tools
Question B5-019
A company has recorded customer calls in Japanese, Korean, and German. For each audio file it needs an English text translation from one call to the Speech fast transcription endpoint, and it wants to guide the output style with a short instruction. What should you configure?
Select one answer.
Answer: A
LLM Speech enhances the fast transcription API with a multimodal model that can transcribe or translate audio files: you enable enhancedMode, set task to translate with a targetLanguage, and can add an optional prompt to guide the output style. The default fast transcription model supports neither translation nor custom prompting, so setting locales or a prompt on it doesn't produce English text. Diarization only separates the speakers in a recording.
Deep dive
How it works
LLM Speech enhances the fast transcription API with a large language model, adding deep contextual understanding, multilingual support and prompt-tuning. It runs one of two tasks per request: transcribe or translate.
For this requirement, the request definition sets:
- enhancedMode with enabled set to true. Without it you get the default fast transcription model.
- task set to translate, plus targetLanguage set to en. Translation requires the target language code. English is on the supported target list, and Japanese, Korean and German audio are supported inputs.
- An optional prompt: text that guides the output style for transcribe or translate.
The feature availability table explains why nothing else fits. For the fast transcription API, translation and custom prompting are marked unsupported without LLM Speech and supported with it.
One request per file goes to the same transcriptions:transcribe endpoint as plain fast transcription, so the pipeline barely changes.
The other options, and when you'd use them
- Default fast transcription with locales: Locales help recognition in the source language, but without LLM Speech the fast transcription API doesn't support translation.
- Diarization with maxSpeakers 2: Diarization distinguishes between speakers in the conversation. It labels who spoke; it doesn't translate.
- Default model with a "Translate into English" prompt: Custom prompting is also unsupported without LLM Speech. Turn on enhancedMode first.
Exam tip
Fast transcription + translate or prompt = LLM Speech: enhancedMode enabled, task translate, targetLanguage, optional prompt.
References
Last reviewed: