PROCTER: PROnunciation-aware ConTextual adaptER for personalized speech recognition in neural transducers

TL;DR

End-to-end ASR struggles with infrequent, user-specific words whose pronunciations vary. The authors present PROCTER, a pronunciation-aware contextual adapter for RNN-T that jointly encodes graphemic and phonemic representations of contextual entities and biases the audio encoder through a multi-layer, cross-attention mechanism. Trained in an adapter fashion with about 1.5M parameters (roughly 1% of the RNN-T), PROCTER delivers substantial gains, achieving up to $57.2\%$ NE-WERR on rare personalized entities and $6.9\%$ NE-WERR in zero-shot device-name scenarios, outperforming text-only contextual adapters. This approach enables robust personalization for ASR with limited additional training and computation, and points to future enhancements like neural grapheme-to-phoneme conversion and extending biasing to decoder outputs.

Abstract

End-to-End (E2E) automatic speech recognition (ASR) systems used in voice assistants often have difficulties recognizing infrequent words personalized to the user, such as names and places. Rare words often have non-trivial pronunciations, and in such cases, human knowledge in the form of a pronunciation lexicon can be useful. We propose a PROnunCiation-aware conTextual adaptER (PROCTER) that dynamically injects lexicon knowledge into an RNN-T model by adding a phonemic embedding along with a textual embedding. The experimental results show that the proposed PROCTER architecture outperforms the baseline RNN-T model by improving the word error rate (WER) by 44% and 57% when measured on personalized entities and personalized rare entities, respectively, while increasing the model size (number of trainable parameters) by only 1%. Furthermore, when evaluated in a zero-shot setting to recognize personalized device names, we observe 7% WER improvement with PROCTER, as compared to only 1% WER improvement with text-only contextual attention

Figures (3)

• Figure 1: Architecture of PROCTER.
• Figure 2: Proposed PROCTER biasing adapter
• Figure 3: Attention visualization