Despite using STT with languageCode set to 'ja' for Japanese transcription, I sometimes receive tra | ElevenLabs | Page 1

AI-generated Answer for ttiger4287

Short answer: This is expected when Scribe v2 detects multiple languages. Setting languageCode to ja does not hard-lock output to Japanese; Scribe v2 automatically detects and transcribes whatever language it hears, and Scribe v2 Realtime can switch languages mid-stream.

What the docs say

Scribe v2 provides automatic multi‑language transcription; a single file can contain multiple languages and the model will detect and transcribe them accordingly. (elevenlabs.io)
Batch STT (POST /v1/speech-to-text) accepts an optional language_code (ISO‑639‑1 or ‑3, e.g., ja or jpn). The page documents temperature but does not claim language_code forces a single-language output. (elevenlabs.io)
Realtime STT exposes language_code as an optional query param and supports automatic language detection; you can also request the detected language in each committed_transcript_with_timestamps event. Realtime additionally supports “text conditioning” via the previous_text field (which can bias the next output). (elevenlabs.io)
Product guide: you may select a primary language, or leave Detect on; “any languages within the audio will be automatically detected.” (elevenlabs.io)

How to keep output Japanese-only (workarounds)

Batch: Pass language_code=ja (or jpn) on POST /v1/speech-to-text; keep temperature at 0 if you want more deterministic phrasing. This hints the model but does not disable detection of other languages present in the audio. (elevenlabs.io)
Realtime:
- Set language_code=ja in the websocket query. Enable include_language_detection=true and drop any committed_transcript_with_timestamps segments whose language_code isn’t ja. (elevenlabs.io)
- Avoid sending previous_text in English/Italian; this conditions the model and can steer output toward that language. (elevenlabs.io)
If your recordings actually contain English/Italian utterances (intros, code‑switching, background speech), Scribe v2 will transcribe those as such by design. Remove those sections or segment audio before transcription. (elevenlabs.io)
If certain Japanese terms are being misrecognized into English/Italian, add them as keyterms to bias recognition. (elevenlabs.io)

Limitations

There is no documented “force single language” or “no language switching” flag beyond providing language_code; automatic detection remains active. (elevenlabs.io)

References

Transcription overview and features (Scribe v2) (elevenlabs.io)
Scribe v2 blog: automatic multi‑language transcription, keyterms, entity detection (elevenlabs.io)
Batch STT API (POST /v1/speech

#Despite using STT with languageCode set to 'ja' for Japanese transcription, I sometimes receive tra