Voice Module - GPT_Sovits Configuration Guide ✨

Bailing's Important Notice

The API interfaced in this section is GPT_Sovits with the entry point as api_v2.py API interface.

Now GPT_Sovits has been updated with an inference specialization package, abbreviated as GSVI, with the entry point as GSVI.py.

Please note that the GSVI interface cannot be used for this module, as its API interface is incompatible with this module's implementation

Meow ha~! Welcome to this GPT_Sovits voice module guide! I'm Mo Bailing, a cat mage who loves researching magic.

Below is the configuration instructions for this module: This module is used to connect to GPT_Sovits TTS. The sample configuration file is in the template folder~

Before starting, everyone needs to copy the template template_configs/gpt_sovits.toml to configs/gpt_sovits.toml, then adjust the configuration file according to your preferences~

toml

[pipeline]
# pipeline preset configuration
default_preset = "default"  # Default preset name to use

# pipeline platform preset mapping
[pipeline.platform_presets]
qq = "default"       # QQ platform uses default preset

[tts]
# GPT-SoVITS API configuration
host = "127.0.0.1"
port = 9880

# Voice synthesis basic configuration
media_type = "wav"        # Audio format: wav

# GPT-SoVITS model parameters
top_k = 12                # top k sampling
top_p = 1.0              # top p sampling
temperature = 1.0        # Temperature coefficient
batch_size = 1           # Batch size
batch_threshold = 0.75   # Batch threshold
text_split_method = "cut5" # Text segmentation method
repetition_penalty = 1.35 # Repetition penalty coefficient
sample_steps = 32        # VITS sampling steps
super_sampling = false    # Whether to enable super sampling

[tts.models]
# Custom character presets
[tts.models.presets]
# Preset character configuration example
[tts.models.presets.default]
name = "Default Character"
gpt_model = ""
sovits_model = ""
ref_audio_path = "path to reference audio" # Reference audio path
aux_ref_audio_paths = []                   # Auxiliary reference audio path list
prompt_text = "this is a ref audio"        # Prompt text
text_language = "auto"                     # Text language
prompt_language = "zh"                     # Prompt language
speed_factor = 1.0                         # Speech speed control

[tts.models.presets.custom1]
name = "kanami"
gpt_model = "GPT_SoVITS/pretrained_models/XiangNaimei-e5.ckpt"
sovits_model = "GPT_SoVITS/pretrained_models/XiangNaimei_e4_s248.pth"
ref_audio_path = "GPT_SoVITS/pretrained_models/voices/平淡.WAV"
aux_ref_audio_paths = []                                              # Auxiliary reference audio path list
prompt_text = "嗨！我每天都会像这样做发声训练的。"                       # Prompt text
text_language = "zh"                                                  # Text language
prompt_language = "zh"                                                # Prompt language
speed_factor = 1.0                                                    # Speech speed control

pipeline Configuration

This part specifies the presets used by platforms.

default_preset: Default preset name to use.

pipeline.platform_presets Presets for Different Platforms

qq: Preset name used by QQ platform~

Bailing's Tip!

The key values here need to match the names after . in [tts.models.presets].

For example, if you want to use the custom1 preset on the QQ platform, you need to configure it as custom1 here, and also have the [tts.models.presets.custom1] configuration in [tts.models.presets]~

Key Name Explanation~

The key names here are platform names, which are also adapter identifier names.

For example, MaiBot Napcat Adapter uses qq as the default identifier, so here it should be set as qq=your preset

tts Configuration 🎙️

Next, this part is used to configure the GPT-SoVITS API interface.

host: GPT-SoVITS API host address, usually 127.0.0.1, i.e., local address.
port: GPT-SoVITS API port, default is 9880.
media_type: Audio format, only supports wav format. Bailing recommends not changing this.

Advanced Configuration

If you're not very familiar with GPT_Sovits yet, you can leave these unchanged for now.

top_k: top k sampling, default is 12.
top_p: top p sampling, default is 1.0.
temperature: Temperature coefficient, default is 1.0.
batch_size: Batch size, default is 1.
batch_threshold: Batch threshold, default is 0.75.
text_split_method: Text segmentation method, default is cut5.
repetition_penalty: Repetition penalty coefficient, default is 1.35.
sample_steps: VITS sampling steps, default is 32.
super_sampling: Whether to enable super sampling, default is false.

tts.models Configuration 📂

This part is where you configure your model files!

tts.models.presets Exclusive Character Configuration

This part is where you configure preset characters.

As Mo Bailing reminded everyone earlier, the preset names must exactly match the names after . in [pipeline.platform_presets]!

Let's take default as an example:

You need to add a configuration like [tts.models.presets.default] (if it already exists, no need to add it again). Then configure the items like this:

name: Character's name, fill it however you like~
gpt_model: Exclusive GPT model path for this character. Please replace it with your own model before use and check the path.
sovits_model: Exclusive SoVITS model path for this character. Similarly, replace it with your own model and check the path.
ref_audio_path: Reference audio path. Please be sure to replace it with your own reference audio and ensure it exists with the correct path.
prompt_text: Prompt text. The text needs to be the same as the content in the reference audio, and the language type needs to correspond to the prompt_language below.
text_language: The language of the text you want to synthesize, default is automatic detection (auto).
prompt_language: Language of the prompt text, default is Chinese (zh).
speed_factor: Speech speed control, default is 1.0.

TIP

For text_language and prompt_language, you can choose from these:

zh: Chinese-English mixed
ja: Japanese-English mixed
yue: Cantonese-English mixed
ko: Korean-English mixed
auto: Multilingual mixed
auto_yue: Multilingual mixed (including Cantonese)
all_zh: Pure Chinese
en: Pure English
all_ja: Pure Japanese
all_yue: Pure Cantonese
all_ko: Pure Korean

This content is copied from the GPT_Sovits code section. If there are issues, please refer to the GPT_Sovits code.

Voice Module - GPT_Sovits Configuration Guide ✨ ​

pipeline Configuration ​

pipeline.platform_presets Presets for Different Platforms ​

tts Configuration 🎙️ ​