An unofficial PyTorch implementation of [VALL-E](https://valle-demo.github.io/), utilizing the [EnCodec](https://github.com/facebookresearch/encodec) encoder/decoder.
> **Note** It is highly recommended to utilize [mrq/ai-voice-cloning](https://git.ecker.tech/mrq/ai-voice-cloning) with `--tts-backend="vall-e"` to handle transcription and dataset preparations.
4. Customize your configuration and define the dataset by modifying `./data/config.yaml`. Refer to `./vall_e/config.py` for details. If you want to choose between different model presets, check `./vall_e/models/__init__.py`.
Included is a helper script to parse the training metrics. Simply invoke it with, for example: `python3 -m vall_e.plot yaml="./training/valle/config.yaml"`
You can specify what X and Y labels you want to plot against by passing `--xs tokens_processed --ys loss stats.acc`
As training under `deepspeed` and Windows is not (easily) supported, under your `config.yaml`, simply change `trainer.backend` to `local` to use the local training backend.
Keep in mind that creature comforts like distributed training or `float16` training cannot be verified as working at the moment with the local trainer.
During experimentation, I've found I can comfortably train on a 4070Ti (12GiB VRAM). Howver, VRAM use is predicated on your dataset; a mix of large and small utterances will cause VRAM usage to spike and can trigger OOM conditions during the backwards pass if you are not careful.
Unfortunately, efforts to train a *good* foundational model seems entirely predicated on a good dataset. My dataset might be too fouled with:
* too short utterances: trying to extrapolate longer contexts seems to utterly fall apart from just the `text` being too long.
* too tightly trimmed utterances: there being little to no space at the start and end might harm associating `<s>` and `</s>` tokens with empty utterances.
* a poorly mapped phoneme mapping: I naively crafted my own phoneme mapping, where a HuggingFace tokenizer might supply a better token mapping.
As the core of VALL-E makes use of a language model, various LLM architectures can be supported and slotted in. Currently supported:
*`transformer`: a basic attention-based transformer implementation, with attention heads + feed forwards.
*`retnet`: using [TorchScale's RetNet](https://github.com/microsoft/torchscale/blob/main/torchscale/architecture/retnet.py) implementation, a retention-based approach can be used instead.
- Its implementation for MoE can also be utilized.
This will export the latest checkpoints, for example, under `./data/ckpt/ar+nar-retnet-8/fp32.pth`, to be loaded on any system with PyTorch, and will include additional metadata, such as the symmap used, and training stats.
*`--ar-temp`: sampling temperature to use for the AR pass. During experimentation, `0.95` provides the most consistent output, but values close to it works fine.
*`--nar-temp`: sampling temperature to use for the NAR pass. During experimentation, `0.2` provides clean output, but values upward of `0.6` seems fine too.
And some experimental sampling flags you can use too (your mileage will ***definitely*** vary):
*`--max-ar-context`: Number of `resp` tokens to keep in the context when inferencing. This is akin to "rolling context" in an effort to try and curb any context limitations, but currently does not seem fruitful.
*`--min-ar-temp` / `--min-nar-temp`: triggers the dynamic temperature pathway, adjusting the temperature based on the confidence of the best token. Acceptable values are between `[0.0, (n)ar-temp)`.
+ This simply uplifts the [original implementation](https://github.com/kalomaze/koboldcpp/blob/dynamic-temp/llama.cpp#L5132) to perform it.
+ **!**NOTE**!**: This does not seem to resolve any issues with setting too high/low of a temperature. The right values are yet to be found.
*`--top-p`: limits the sampling pool to top sum of values that equal `P`% probability in the probability distribution.
*`--top-k`: limits the sampling pool to the top `K` values in the probability distribution.
*`--repetition-penalty`: modifies the probability of tokens if they have appeared before. In the context of audio generation, this is a very iffy parameter to use.
*`--length-penalty`: (AR only) modifies the probability of the stop token based on the current sequence length. This is ***very*** finnicky due to the AR already being well correlated with the length.
- utilize an approach similar to [FasterDecoding/Medusa](https://github.com/FasterDecoding/Medusa/) with additional heads for decoding N+1, N+2, N+3 AR tokens
+ this requires a properly trained AR, however.
* work around issues with extending context past what's trained (despite RetNet's retention allegedly being able to defeat this):
- "sliding" AR input, such as have the context a fixed length.
+ the model may need to be trained for this with a fancy positional embedding injected OR already trained with a sliding context window in mind. Naively sliding the context window while making use of the RetNet implementation's positional embedding doesn't seem fruitful.
- [EnCodec](https://github.com/facebookresearch/encodec) is licensed under CC-BY-NC 4.0. If you use the code to generate audio quantization or perform decoding, it is important to adhere to the terms of their license.
- This implementation was originally based on [enhuiz/vall-e](https://github.com/enhuiz/vall-e), but has been heavily, heavily modified over time.
title={Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers},
author={Wang, Chengyi and Chen, Sanyuan and Wu, Yu and Zhang, Ziqiang and Zhou, Long and Liu, Shujie and Chen, Zhuo and Liu, Yanqing and Wang, Huaming and Li, Jinyu and others},
journal={arXiv preprint arXiv:2301.02111},
year={2023}
}
```
```bibtex
@article{defossez2022highfi,
title={High Fidelity Neural Audio Compression},
author={Défossez, Alexandre and Copet, Jade and Synnaeve, Gabriel and Adi, Yossi},
