Back to feed
arXiv cs.CL
arXiv cs.CL
7/8/2026
Revisiting the Relation Between Language Model Perplexity and ASR Word Error Rate for Modern End-to-End Speech Recognition

Revisiting the Relation Between Language Model Perplexity and ASR Word Error Rate for Modern End-to-End Speech Recognition

Short summary

This paper re-examines the historically linear log-log relationship between language model perplexity (PPL) and ASR word error rate (WER) in the context of modern end-to-end speech recognition systems. The authors investigate whether external LMs still improve current ASR systems, how encoder context length affects the PPL-WER relation, and how LLM perplexities fit existing trends. They also show that internal language model subtraction in attention-based encoder-decoder systems alters the observed PPL-WER relation, indicating the decoder's internal LM must be accounted for when evaluating external LM quality.

  • •Revisits the PPL-WER linear relation assumption for modern end-to-end ASR systems
  • •Examines impact of external LMs, encoder context length, and LLM perplexities on the trend
  • •Shows internal LM subtraction changes the observed relation, complicating external LM quality interpretation

Generated with AI, which can make mistakes.

#research-breakthrough
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

Encoding EEG Signals to Examine Human-Like Next-Word Prediction Behaviour in Language Models
arXiv cs.CL
Do LLMs Need Architectural Changes for Simultaneous Speech Translation? A Prefix-to-Prefix Data Driven Approach
arXiv cs.CL
Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents
arXiv CS.AI
An Emergent Mirage: Is Emergent Misalignment and Realignment Indeed a Robust Phenomenon?
arXiv cs.CL