UTAR Institutional Repository

Voice to text conversion app with speaker recognition

Ang, Sea Zhe (2022) Voice to text conversion app with speaker recognition. Final Year Project, UTAR.

[img]
Preview
PDF
Download (3108Kb) | Preview

    Abstract

    This project is a final year project of a computer science student. Voice recognition is a field that is still quite underdeveloped. There are still a lot of obstacles that are needed to be overcome before voice recognition system can identify all the speakers correctly under all kind of conditions. This would be helpful in speaker verification field, and also a speech recognition system that is personalized to the user. In this paper, Mel Frequency Cepstral Coefficient and delta of it are used to describe the vocal traits of a person. Mel Frequency Cepstral Coefficient is popular in this field to describe the phenomes of voice. After that, Gaussian Mixture Model is used to represent each speaker or each pair of speakers. In the first part of experiments using self-generated datasets, the total number of users that are tested in this paper is 5. 25 voice recordings, where 5 of them belongs to each speaker are used as the input to the system for single speaker identification. For two simultaneous speaker identifications, 65 voice recordings where 40 of them are artificially mixed are used as the input to the system for two simultaneous speaker identifications. In the second part of experiments using LibriSpeech datasets, the total number of users that are tested in this paper is 20 including me and speakers from LibriSpeech dataset. There are a total of 20 single speaker models. Not only that, 5 single word speaker models are trained to detect each short word is spoken by who. Finally, a Universal Background Model is also built for speaker verification. All the models are built using Gaussian Distributions Technique. The experiments that are done in second part of experiments are single speaker identification, non-overlapped multi-speaker identification with speech extraction with known speakers, speaker verification and speaker verification with noise estimation and speech extraction with unknown which is the main part of the project, Voice To Text Conversion With Speaker Recognition.

    Item Type: Final Year Project / Dissertation / Thesis (Final Year Project)
    Subjects: Q Science > Q Science (General)
    T Technology > T Technology (General)
    Divisions: Faculty of Information and Communication Technology > Bachelor of Computer Science (Honours)
    Depositing User: ML Main Library
    Date Deposited: 15 Jan 2023 21:20
    Last Modified: 15 Jan 2023 21:20
    URI: http://eprints.utar.edu.my/id/eprint/4637

    Actions (login required)

    View Item