Barkify: an unoffical repo for training Bark, a text-prompted generative audio model by suno-ai.
Bark has two GPT style models which is compatible for prompting and other tricks from NLP. Bark realize a great real world tts result but the repo itself doesn't a training recipe. We want to conduct some experiments or train this model. Here we release our basic training code which might be a guidance of training for open source community.
We do our experiment on LJspeech. Follow the instrcutions in process.ipynb.
For Chinese, we test a famous steamer named 峰哥亡命天涯. It shows an acceptable result but worse than our other TTS repo.
For English, we test LibriTTS dataset. It works fine and basic items in our roadmap have been proved.
Stage1 stands for text to semantic and stage2 stands for semantic to acoustic.
You should config paramters in the configs/barkify.yaml. We use one A100 to train our model (both S1&S2).
# training stage 1 or 2
python trainer.py start_path=/path/to/your/work_env stage=1 name=<dataset>
python trainer.py start_path=/path/to/your/work_env stage=2 name=<dataset>
Directly use infer.ipynb and follow the instrcutions to infer your model.
We have already achieve the following items and we will release our code soon.
- Construct a basic training code for bark-like generative model
- Test one speaker scenario
- Test multi speaker scenario
- Test speaker semantic prompting
- Test speech/audio acoustic prompting
- Test variable length data(as we use a fixed length now)
These items are pretty data-hungry or rely on massive GPUs.
So we are open to any sponsors or collaborators to finish these jobs.
You could contact us by QQ: 3284494602 or email us at 3284494602@qq.com
- Long-form generation(which may be longer than 1min.)
- Support more language(especially for ZH)
- Paralanguage modeling in the text input
- Speaker generation by text prompts
- Emotion/Timbre/Rhythm controlling by text/acoustic prompts
- Add/Remove background noise(which might be important for downstream tasks)