---
name: Language conditioning is mandatory for parakeet TDT
description: Parakeet TDT 1.1B requires language embedding at inference — skipping it causes 84% WER instead of 37%
type: feedback
---

Always pass language IDs when running inference on Parakeet TDT 1.1B (hybrid-tdt-gemma) checkpoints. The LanguageEmbedding module is load-bearing — without it, WER jumps from ~37% to ~84% because the model outputs Latin transliterations instead of Indic script.

**Why:** The model was trained with language embeddings conditioned on the encoder from step 0. Running without them puts the model in a mode it was never trained in.

**How to apply:** Register `LanguageEmbedding` module before loading state_dict, patch encoder forward to inject `_current_lang_ids`, and set lang IDs per batch during transcription. Use `strict=False` is NOT sufficient — must explicitly register and load the lang embed weights.