Blog

Technical blog posts on speech AI, machine learning, and deep learning

LoRA：低秩适配的高效参数微调方法

Low-Rank Adaptation of Large Language Models

LoRA：低秩适配的高效参数微调方法 简介 LoRA (Low-Rank Adaptation) 是一种高效的大模型参数微调技术，通过低秩分解实现了高效参数微调，且不需要新增 speaker id。这种方法在保持模型性能的同时，大幅降低了可训练参数的数量。 核心思想 LoRA的核心思想是通过低秩矩阵分解来近似全参数微调中的权重更新。具体来说，对于预训练模型的权重矩阵，LoRA假设其更...

Posted by 朱晓旭 on July 30, 2025
LoRA 参数微调 深度学习 大模型 低秩分解 Parameter Efficient

【语音合成前端】Unified Mandarin TTS Front-end Based on Distilled BERT Model 论文解读

Unified Mandarin TTS Front-end Based on Distilled BERT Model

Posted by 朱晓旭 on April 16, 2021
Front End BERT TinyBERT knowledge distillation

【语音合成前端】Flavored Tacotron Conditional Learning for Prosodic-linguistic Features 论文解读

Flavored Tacotron Conditional Learning for Prosodic-linguistic Features

Posted by 朱晓旭 on April 16, 2021
Front End English prosody conditioning Tacotron2

【语音合成前端】A unified sequence-to-sequence front-end model for Mandarin text-to-speech synthesis 论文解读

A unified sequence-to-sequence front-end model for Mandarin text-to-speech synthesis

Posted by 朱晓旭 on April 16, 2021
Front End self attention TinyBERT seq2seq

【多音字消歧】Knowledge Distillation from Bert in Pre-Training and Fine-Tuning for Polyphone Disambiguation 论文解读

A Mask-based Model for Mandarin Chinese Polyphone Disambiguation

Posted by 朱晓旭 on April 12, 2021
Front End Polyphone Disambiguation mask vector BERT fine-tuning knowledge distillation

【多音字消歧】A Bi-directional LSTM Approach for Polyphone Disambiguation in Mandarin Chinese论文解读

A Bi-directional LSTM Approach for Polyphone Disambiguation in Mandarin Chinese

Posted by 朱晓旭 on March 17, 2021
Front End Polyphone Disambiguation BLSTM POS

【多音字消歧】A Mask-based Model for Mandarin Chinese Polyphone Disambiguation 论文解读

A Mask-based Model for Mandarin Chinese Polyphone Disambiguation

Posted by 朱晓旭 on March 16, 2021
Front End Polyphone Disambiguation mask vector Modified Focal Loss

【声码器】HIFI-GAN论文解读

hifi-gan

Posted by 朱晓旭 on March 4, 2021
TTS Vocoder HIFI-GAN MelGAN

多线程与线程安全

Multithread

进程(Process) 狭义的进程是正在运行的程序的实例；广义的进程是一个具有一定独立功能的程序关于某个数据集合的一次运行活动，是操作系统动态执行的基本单元。进程间数据不共享(因此才会有进程间通讯这个课题，进程间通讯的方法有：匿名管道，有名管道，消息队列，信号量，信号，共享内存，套接字等等)。 线程(Thread) 线程是程序执行流的最小单位；一个标准的线程由线程ID、当前指令指针(PC)...

Posted by 朱晓旭 on January 7, 2021
线程安全 多线程

转置卷积剖析

Transpose Convolution

问题 之前遇到过多次转置卷积的运算，查了就忘，这次在PQMF算法C语言实现过程中，对转置卷积采用矩阵乘加速，就详细记录下来分析过程。这里采用语音领域常用的一维转置卷积为例，高维会更简单。 一维转置卷积矩阵乘分析 输入输出信信息 input: [batch size, input channel, input length] weight(kernel): [output channel, ...

Posted by 朱晓旭 on December 24, 2020
conv transpose1d transpose conv