This is part of talk-llama-fast. Don't use installation guide at current page here, it is outdated and left for legacy. Full and actual instruction how to install is here: https://github.com/Mozer/talk-llama-fast
Old and original readme:
This project is inspired by silero-api-server and utilizes XTTSv2.
This server was created for SillyTavern but you can use it for your needs
Feel free to make PRs or use the code for your own needs
There's a google collab version you can use it if your computer is weak.
If you are looking for an option for normal XTTS use look here https://github.com/daswer123/xtts-webui
Recently I have little time to do this project, so I advise you to get acquainted with a similar project
You can keep track of all changes on the release page
- Make it possible to change generation parameters through the generation request and through a different endpoint
Simple installation :
pip install xtts-api-server
This will install all the necessary dependencies, including a CPU support only version of PyTorch
I recommend that you install the GPU version to improve processing speed ( up to 3 times faster )
python -m venv venv
venv\Scripts\activate
pip install xtts-api-server
pip install torch==2.1.1+cu118 torchaudio==2.1.1+cu118 --index-url https://download.pytorch.org/whl/cu118
python -m venv venv
source venv\bin\activate
pip install xtts-api-server
pip install torch==2.1.1+cu118 torchaudio==2.1.1+cu118 --index-url https://download.pytorch.org/whl/cu118
# Clone REPO
git clone https://github.com/daswer123/xtts-api-server
cd xtts-api-server
# Create virtual env
python -m venv venv
venv/scripts/activate or source venv/bin/activate
# Install deps
pip install -r requirements.txt
pip install torch==2.1.1+cu118 torchaudio==2.1.1+cu118 --index-url https://download.pytorch.org/whl/cu118
# Launch server
python -m xtts_api_server
A Dockerfile is provided to build a Docker image, and a docker-compose.yml file is provided to run the server with Docker Compose as a service.
You can build the image with the following command:
mkdir xtts-api-server
cd xtts-api-server
docker run -d daswer123/xtts-api-server
OR
cd docker
docker compose build
Then you can run the server with the following command:
docker compose up # or with -d to run in background
python -m xtts_api_server
will run on default ip and port (localhost:8020)
Use the --deepspeed
flag to process the result fast ( 2-3x acceleration )
usage: xtts_api_server [-h] [-hs HOST] [-p PORT] [-sf SPEAKER_FOLDER] [-o OUTPUT] [-t TUNNEL_URL] [-ms MODEL_SOURCE] [--listen] [--use-cache] [--lowvram] [--deepspeed] [--streaming-mode] [--stream-play-sync]
Run XTTSv2 within a FastAPI application
options:
-h, --help show this help message and exit
-hs HOST, --host HOST
-p PORT, --port PORT
-d DEVICE, --device DEVICE `cpu` or `cuda`, you can specify which video card to use, for example, `cuda:0`
-sf SPEAKER_FOLDER, --speaker-folder The folder where you get the samples for tts
-o OUTPUT, --output Output folder
-mf MODELS_FOLDERS, --model-folder Folder where models for XTTS will be stored, finetuned models should be stored in this folder
-t TUNNEL_URL, --tunnel URL of tunnel used (e.g: ngrok, localtunnel)
-ms MODEL_SOURCE, --model-source ["api","apiManual","local"]
-v MODEL_VERSION, --version You can download the official model or your own model, official version you can find [here](https://huggingface.co/coqui/XTTS-v2/tree/main) the model version name is the same as the branch name [v2.0.2,v2.0.3, main] etc. Or you can load your model, just put model in models folder
--listen Allows the server to be used outside the local computer, similar to -hs 0.0.0.0
--use-cache Enables caching of results, your results will be saved and if there will be a repeated request, you will get a file instead of generation
--lowvram The mode in which the model will be stored in RAM and when the processing will move to VRAM, the difference in speed is small
--deepspeed allows you to speed up processing by several times, automatically downloads the necessary libraries
--streaming-mode Enables streaming mode, currently has certain limitations, as described below.
--streaming-mode-improve Enables streaming mode, includes an improved streaming mode that consumes 2gb more VRAM and uses a better tokenizer and more context.
--stream-play-sync Additional flag for streaming mod that allows you to play all audio one at a time without interruption
You can specify the path to the file as text, then the path counts and the file will be voiced
You can load your own model, for this you need to create a folder in models and load the model with configs, note in the folder should be 3 files config.json
vocab.json
model.pth
If you want your host to listen, use -hs 0.0.0.0 or use --listen
The -t or --tunnel flag is needed so that when you get speakers via get you get the correct link to hear the preview. More info here
Model-source defines in which format you want to use xtts:
local
- loads version 2.0.2 by default, but you can specify the version via the -v flag, model saves into the models folder and usesXttsConfig
andinference
.apiManual
- loads version 2.0.2 by default, but you can specify the version via the -v flag, model saves into the models folder and uses thetts_to_file
function from the TTS apiapi
- will load the latest version of the model. The -v flag won't work.
All versions of the XTTSv2 model can be found here the model version name is the same as the branch name [v2.0.2,v2.0.3, main] etc.
The first time you run or generate, you may need to confirm that you agree to use XTTS.
Streaming mode allows you to get audio and play it back almost immediately. However, it has a number of limitations.
You can see how this mode works here and here
Now, about the limitations
- Can only be used on a local computer
- Playing audio from the your pc
- Does not work endpoint
tts_to_file
onlytts_to_audio
and it returns 1 second of silence.
You can specify the version of the XTTS model by using the -v
flag.
Improved streaming mode is suitable for complex languages such as Chinese, Japanese, Hindi or if you want the language engine to take more information into account when processing speech.
--stream-play-sync
flag - Allows you to play all messages in queue order, useful if you use group chats. In SillyTavern you need to turn off streaming to work correctly
API Docs can be accessed from http://localhost:8020/docs
By default the speakers
folder should appear in the folder, you need to put there the wav file with the voice sample, you can also create a folder and put there several voice samples, this will give more accurate results
You can change the folders for speakers and the folder for output via the API.
The following post is a quote by user Material1276 from reddit
Some suggestions on making good samples
Keep them about 7-9 seconds long. Longer isn't necessarily better.
Make sure the audio is down sampled to a Mono, 22050Hz 16 Bit wav file. You will slow down processing by a large % and it seems cause poor quality results otherwise (based on a few tests). 24000Hz is the quality it outputs at anyway!
Using the latest version of Audacity, select your clip and Tracks > Resample to 22050Hz, then Tracks > Mix > Stereo to Mono. and then File > Export Audio, saving it as a WAV of 22050Hz
If you need to do any audio cleaning, do it before you compress it down to the above settings (Mono, 22050Hz, 16 Bit).
Ensure the clip you use doesn't have background noises or music on e.g. lots of movies have quiet music when many of the actors are talking. Bad quality audio will have hiss that needs clearing up. The AI will pick this up, even if we don't, and to some degree, use it in the simulated voice to some extent, so clean audio is key!
Try make your clip one of nice flowing speech, like the included example files. No big pauses, gaps or other sounds. Preferably one that the person you are trying to copy will show a little vocal range. Example files are in here
Make sure the clip doesn't start or end with breathy sounds (breathing in/out etc).
Using AI generated audio clips may introduce unwanted sounds as its already a copy/simulation of a voice, though, this would need testing.
- Thanks to the author Kolja Beigel for the repository RealtimeTTS , I took some of its code for my project.
- Thanks erew123 for the note about creating samples and the code to download the models
- Thanks lendot for helping to fix the multiprocessing bug and adding code to use multiple samples for speakers