UMA-Split: unimodal aggregation for both English and Mandarin non-autoregressive speech recognition
This work addresses a language-specific bottleneck in speech recognition models, making it incremental by extending an existing method to handle multiple languages.
The paper tackles the problem of adapting a unimodal aggregation (UMA) method for non-autoregressive speech recognition to work effectively in both English and Mandarin, addressing issues where UMA struggles with English due to fine-grained tokenization. The proposed UMA-Split method allows each aggregated frame to map to multiple tokens, improving performance across languages.
This paper proposes a unimodal aggregation (UMA) based nonautoregressive model for both English and Mandarin speech recognition. The original UMA explicitly segments and aggregates acoustic frames (with unimodal weights that first monotonically increase and then decrease) of the same text token to learn better representations than regular connectionist temporal classification (CTC). However, it only works well in Mandarin. It struggles with other languages, such as English, for which a single syllable may be tokenized into multiple fine-grained tokens, or a token spans fewer than 3 acoustic frames and fails to form unimodal weights. To address this problem, we propose allowing each UMA-aggregated frame map to multiple tokens, via a simple split module that generates two tokens from each aggregated frame before computing the CTC loss.