A High-Fidelity Multi-Speaker Multi-Lingual VQTTS System for Low-Resource Text-to-Speech
This research paper presents a novel VQTTS-based system for multi-speaker multi-lingual text-to-speech synthesis, achieving high fidelity and naturalness, especially in low-resource scenarios, by utilizing VQ acoustic features, speaker and language embeddings, and a cross-lingual synthesis approach.
대규모 언어 모델 기반 고급 다국어 TTS 시스템, Fish-Speech: 음성 복제 및 실시간 처리에 중점
Fish-Speech는 기존 TTS 시스템의 한계를 극복하기 위해 LLM과 듀얼 AR 아키텍처를 활용하여 고품질의 다국어 음성 합성, 특히 음성 복제 및 실시간 처리에 뛰어난 성능을 보이는 혁신적인 TTS 프레임워크입니다.
Fish-Speech: A Novel TTS Framework Using LLMs for Multilingual Synthesis and Enhanced Voice Cloning
Fish-Speech is a novel TTS framework leveraging LLMs and a dual autoregressive architecture to achieve high-quality, multilingual speech synthesis and improved voice cloning capabilities, outperforming traditional G2P-based methods.
LINA-SPEECH: A Parameter-Efficient Text-to-Speech Synthesis Model Using Gated Linear Attention and Initial-State Tuning
LINA-SPEECH, a novel text-to-speech synthesis model, utilizes Gated Linear Attention and Initial-State Tuning to achieve comparable or superior performance to larger, more complex models while maintaining parameter efficiency.
Very Attentive Tacotron: Achieving Robustness and Unbounded Length Generalization in Autoregressive Transformer-Based Text-to-Speech
Very Attentive Tacotron (VAT), a novel autoregressive Transformer-based TTS model, overcomes robustness and length generalization limitations of previous models by incorporating an alignment mechanism with interpolated relative position biases (IRPBs), achieving high-quality speech synthesis for virtually any length without word omissions or repetitions.
대규모 언어 모델을 음성 생성에 활용하는 방법: 후기 융합 기반 TTS 시스템 및 멀티모달 LLM 개발 연구
사전 훈련된 대규모 언어 모델(LLM)을 매개변수 효율적 미세 조정(PEFT) 및 Mixture-of-Experts 아키텍처를 통해 음성 생성 기능을 갖춘 텍스트 음성 변환(TTS) 시스템 및 텍스트-음성 멀티모달 LLM으로 발전시킬 수 있습니다.
VQTTS: Improving Neural Text-to-Speech Synthesis with Self-Supervised Vector-Quantized Acoustic Features
VQTTS, a novel text-to-speech synthesis system, leverages self-supervised vector-quantized acoustic features and a redesigned acoustic model to achieve state-of-the-art performance in speech naturalness.
ControlSpeech: A Novel Text-to-Speech System for Simultaneous Zero-Shot Speaker Cloning and Style Control Using a Decoupled Codec
ControlSpeech is a novel TTS system that leverages a decoupled codec and a novel Style Mixture Semantic Density (SMSD) module to achieve simultaneous zero-shot speaker cloning and flexible style control, addressing limitations of previous models that could not independently manipulate content, timbre, and style.
Continuous Speech Tokenizer for Improved Text-to-Speech Synthesis
Continuous speech tokenization enhances text-to-speech synthesis by preserving more audio information than traditional discrete methods, leading to improved speech continuity, quality, and robustness to sampling rate variations.
SALAD: Exploring Continuous Representation in Zero-Shot Text-to-Speech Synthesis using Per-Token Latent Diffusion
Quantizing inherently continuous modalities like audio for text-to-speech synthesis may be suboptimal, and continuous representation learning using per-token latent diffusion models like SALAD offers a competitive alternative with superior intelligibility.
© 2024 by Linnk AI