Embedded Speech Technology

End-to-End models in Automatic Speech Recognition simplify the speech recognition process. They convert audio data directly into text representation without exploiting multiple stages and systems. This direct approach is efficient and reduces potential points of error. On the contrary, Sequence-to-Sequence models adopt a more integrative approach where they use distinct models for retrieving the acoustic and language-specific features, which are respectively known as acoustic and language models. This integration allows for better coordination between different speech aspects, potentially leading to more accurate transcriptions. In this thesis, we explore various Speech-to-Text (STT) models, mainly focusing on End-to-End and Sequence-to-Sequence techniques. We also look into using offline STT tools such as Wav2Vec2.0, Kaldi and Vosk. These tools face challenges when handling new voice data or various accents of the same language. To address this challenge, we fine-tune the models to make them better at handling new, unseen data. Through our comparison, Wav2Vec2.0 emerged as the top performer, though with a larger model size. Our approach also proves that using Kaldi and Vosk together creates a robust STT system that can identify new words using phonemes.