ASR Error Correction with BERT, ELECTRA, and a Fuzzy‑Phoneme Generator: Methods, Experiments, and Future Directions

The talk introduces the ASR error correction problem faced by Xiaomi's Xiao‑Ai voice assistant, where inaccuracies in the Automatic Speech Recognition (ASR) stage produce erroneous query texts that hinder downstream Natural Language Understanding (NLU).

Typical ASR mistakes include homophonic substitutions, misrecognition of rare characters, and confusion caused by background noise, illustrated with examples such as "生僻字" being transcribed as "升壁纸".

Problem settings are defined: only queries longer than six characters are considered, context and audio information are ignored, a one‑to‑one correction paradigm is adopted, and only unsupervised corpora are used for training.

Related work covers the BERT model (masked language modeling) and its limitations for error detection, the ELECTRA generator‑discriminator framework, and the Soft‑Masked BERT approach that combines a BiGRU error detector with BERT for soft masking.

The proposed solution mirrors the generator‑discriminator design: a fuzzy‑phoneme generator creates synthetic ASR errors based on phonetic similarity levels, and a correction discriminator jointly processes character embeddings from a pretrained BERT and phoneme embeddings, followed by a softmax classification layer.

Training data comprise nearly 100 million Chinese sentences from Wikipedia, Zhihu, news crawls, and Xiao‑Ai user logs; a fuzzy‑phoneme generator simulates errors by replacing tokens according to five phonetic similarity levels and a non‑standard pinyin scheme.

Experimental results show that baseline BERT fine‑tuning yields 9.3 % F1, while adding vocabulary filtering, recursive prediction, correction training, and phoneme features raises performance to 77.6 % F1. Error examples demonstrate successful correction of mis‑recognized queries such as "忙种" → "芒种" and "基因" → "精英".

Future work includes exploring seq2seq BERT‑Decoder models (e.g., GPT), incorporating contextual attention for long‑tail domain knowledge, and designing lightweight update mechanisms to keep the model current with emerging hot topics.