Streaming automatic speech recognition (ASR) outputs text in a spoken form, such as "twenty five dollars" or "March third." Inverse text normalization (ITN) is the process that converts these utterances into properly formatted written text, like "$25" or "March 3rd." This is essential for real-time applications where users expect readable transcriptions as they speak.
The new paper, posted on arXiv, proposes a method that frames streaming ITN as a text-to-text task. The authors combine a pretrained text-to-text language model with a "Read-Tag-Write" policy, which appears to structure how the model processes input tokens and generates normalized output. The abstract notes that existing streaming ITN methods rely on hybrid systems combining neural components, though the sentence is cut off in the source, leaving the exact baseline details unclear.
Because the abstract is truncated, the full evaluation results and implementation specifics are not available from this source alone. The contribution, based on the title and available text, is the introduction of a more efficient simultaneous ITN approach using a pretrained model and a structured policy, rather than a purely hybrid architecture.