Automatic Fluency Assessment Method for Spontaneous Speech without Reference Text

The automatic fluency assessment of spontaneous speech without reference text is a challenging task that heavily depends on the accuracy recognition (ASR). Considering this scenario, it necessary to explore an method combines ASR. This mainly due fact in addition acoustic features being essential for assessment, output by ASR may also contain potentially information. However, most existing studies are based solely audio features, utilizing textual information, which lead limited understanding features. To address this, we propose multimodal output. Specifically, first relevance and fine-tune Wav2Vec2.0 model using multi-task learning jointly optimize task, resulting both results Then, obtained from fine-tuned fed into model, attention mechanisms obtain more reliable results. Finally, experiments PSCPSF Speechocean762 dataset suggest our proposed performs well different scenarios.