The paper investigates methods of compressing part-of-speech models used for pre-annotating utterances supplied to ondevice grapheme-to-phoneme converter. The performance of Transformer-based part-of-speech models is analyzed under different compression regimes. The evaluation is done with respect to French, German and Italian datasets that consist of TTS input prompts. The study shows that a proper selection of a compression method reduces the model size significantly without deteriorating the grapheme-to-phoneme conversion performance. 

https://ieeexplore.ieee.org/document/9746710