Upload Convbased Studio models and assets: README.md
Browse files
README.md
CHANGED
|
@@ -12,76 +12,30 @@ tags:
|
|
| 12 |
- multi-speaker
|
| 13 |
- convolution
|
| 14 |
- encoder-decoder
|
| 15 |
-
- aishell
|
| 16 |
-
- vctk
|
| 17 |
license: apache-2.0
|
| 18 |
datasets:
|
| 19 |
-
- aishell
|
| 20 |
-
- thchs30
|
| 21 |
-
- primewords
|
| 22 |
- vctk
|
| 23 |
library_name: pytorch
|
| 24 |
---
|
| 25 |
|
| 26 |
# Convbased
|
| 27 |
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
-
|
| 41 |
-
-
|
| 42 |
-
-
|
| 43 |
-
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
- **解码器**: 声学特征解码器
|
| 48 |
-
- **判别器**: 对抗训练判别器
|
| 49 |
-
- **损失函数**: 组合损失(Mel频谱损失 + KL散度损失 + 特征匹配损失)
|
| 50 |
-
|
| 51 |
-
## 训练曲线
|
| 52 |
-
|
| 53 |
-
模型训练过程中的各项损失函数变化:
|
| 54 |
-
|
| 55 |
-

|
| 56 |
-
*判别器总损失*
|
| 57 |
-
|
| 58 |
-

|
| 59 |
-
*生成器特征匹配损失*
|
| 60 |
-
|
| 61 |
-

|
| 62 |
-
*KL散度损失*
|
| 63 |
-
|
| 64 |
-

|
| 65 |
-
*Mel频谱损失*
|
| 66 |
-
|
| 67 |
-

|
| 68 |
-
*生成器总损失*
|
| 69 |
-
|
| 70 |
-
## 训练数据集
|
| 71 |
-
|
| 72 |
-
本模型使用以下高质量中文语音数据集进行训练:
|
| 73 |
-
|
| 74 |
-
| 数据集名称 | 时长(小时) | 描述 |
|
| 75 |
-
|-------------------|-------------|------|
|
| 76 |
-
| data_aishell | 178 | 中文普通话语音识别数据集 |
|
| 77 |
-
| data_thchs30 | 30 | 清华大学中文语音数据集 |
|
| 78 |
-
| primewords_md_2018| 178 | 中文语音合成数据集 |
|
| 79 |
-
| VCTK | 44 | 英文多说话人数据集 |
|
| 80 |
-
| 四川方言 | 4 | 四川话方言数据 |
|
| 81 |
-
| 闽南语 | 3 | 闽南话方言数据 |
|
| 82 |
-
| 粤语 | 3 | 粤语方言数据 |
|
| 83 |
-
| 温州方言 | 7 | 温州话方言数据 |
|
| 84 |
-
| 噪声 | 20 | 噪声环境语音数据 |
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
*本模型致力于推进中文语音合成技术的发展,该底模已用于微调大部分模型于 [Convbased](https://weights.chat/)*
|
|
|
|
| 12 |
- multi-speaker
|
| 13 |
- convolution
|
| 14 |
- encoder-decoder
|
|
|
|
|
|
|
| 15 |
license: apache-2.0
|
| 16 |
datasets:
|
|
|
|
|
|
|
|
|
|
| 17 |
- vctk
|
| 18 |
library_name: pytorch
|
| 19 |
---
|
| 20 |
|
| 21 |
# Convbased
|
| 22 |
|
| 23 |
+
本项目专注于训练高质量的预训练底模,为语音转换任务提供强大的基础模型支持。
|
| 24 |
+
|
| 25 |
+
| 特征提取 | 声码器 | 40k | 48k |
|
| 26 |
+
|-----------|--------|-----|-----|
|
| 27 |
+
| contentvec | hifigannsf | ❌ | ✅ |
|
| 28 |
+
| contentvec | sifigan | ❌ | ✅ |
|
| 29 |
+
| contentvec | refinegan | ❌ | ❌ |
|
| 30 |
+
| contentvec | hifiganmrf | ❌ | ❌ |
|
| 31 |
+
| spin | hifigannsf | ❌ | ❌ |
|
| 32 |
+
| spin | sifigan | ❌ | ✅ |
|
| 33 |
+
| spin | refinegan | ❌ | ❌ |
|
| 34 |
+
| spin | hifiganmrf | ❌ | ❌ |
|
| 35 |
+
| chinese-hubert-base | hifigannsf | ❌ | ✅ |
|
| 36 |
+
| chinese-hubert-base | sifigan | ❌ | ❌ |
|
| 37 |
+
| chinese-hubert-base | refinegan | ❌ | ❌ |
|
| 38 |
+
| chinese-hubert-base | hifiganmrf | ❌ | ❌ |
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
*致力于推进中文语音合成技术的发展,该底模已用于微调大部分模型于 [Convbased Studio](https://weights.chat/)*
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|