View Proposal

---

Proposer

Zi Hau Chin

Title

Enhancing Voice Synthesis Control in a Transformer-Based Speech Synthesis System through Probabilistic Latent Space Manipulation

Goal

To advance control and realism in voice synthesising.

Description

This project explores an innovative method to enhance voice synthesis in a transformer-based speech synthesis system. Instead of discrete tokens, the speech synthesis component will receive probability distributions from the language model's latent space. This aims to provide finer-grained control over speaker characteristics. The core work involves: 1. Modifying the system to generate and accept these probabilistic inputs. 2. Synthesizing speech using both the original and new methods. 3. Comparing speaker identity preservation and voice quality quantitatively using speaker recognition models (e.g., EER, speaker embedding similarity) and potentially human evaluation.

Resources

Dedicated server for model training, evaluation, and application hosting.

Background

Python, NLP, TTS

Url

Difficulty Level

Challenging

Ethical Approval

None

Number Of Students

1

Supervisor

Zi Hau Chin

Keywords

ai, voice synthesizer, speaker recognition

Degrees

Bachelor of Science in Computing Science