#### ✨ A studio that contains visible f0 editor, speaker mix timeline editor and other features (Where the Onnx models are used) : [MoeVoiceStudio](https://github.com/NaruseMioShirakana/MoeVoiceStudio)
**This project differs fundamentally from VITS, as it focuses on Singing Voice Conversion (SVC) rather than Text-to-Speech (TTS). In this project, TTS functionality is not supported, and VITS is incapable of performing SVC tasks. It's important to note that the models used in these two projects are not interchangeable or universally applicable.**
The purpose of this project was to enable developers to have their beloved anime characters perform singing tasks. The developers' intention was to focus solely on fictional characters and avoid any involvement of real individuals, anything related to real individuals deviates from the developer's original intention.
This project is an open-source, offline endeavor, and all members of SvcDevelopTeam, as well as other developers and maintainers involved (hereinafter referred to as contributors), have no control over the project. The contributors have never provided any form of assistance to any organization or individual, including but not limited to dataset extraction, dataset processing, computing support, training support, inference, and so on. The contributors do not and cannot be aware of the purposes for which users utilize the project. Therefore, any AI models and synthesized audio produced through the training of this project are unrelated to the contributors. Any issues or consequences arising from their use are the sole responsibility of the user.
This project is run completely offline and does not collect any user information or gather user input data. Therefore, contributors to this project are not aware of all user input and models and therefore are not responsible for any user input.
This project serves as a framework only and does not possess speech synthesis functionality by itself. All functionalities require users to train the models independently. Furthermore, this project does not come bundled with any models, and any secondary distributed projects are independent of the contributors of this project.
# Warning: Please ensure that you address any authorization issues related to the dataset on your own. You bear full responsibility for any problems arising from the usage of non-authorized datasets for training, as well as any resulting consequences. The repository and its maintainer, svc develop team, disclaim any association with or liability for the consequences.
1. This project is exclusively established for academic purposes, aiming to facilitate communication and learning. It is not intended for deployment in production environments.
2. Any sovits-based video posted to a video platform must clearly specify in the introduction the input source vocals and audio used for the voice changer conversion, e.g., if you use someone else's video/audio and convert it by separating the vocals as the input source, you must give a clear link to the original video or music; if you use your own vocals or a voice synthesized by another voice synthesis engine as the input source, you must also state this in your introduction.
3. You are solely responsible for any infringement issues caused by the input source and all consequences. When using other commercial vocal synthesis software as an input source, please ensure that you comply with the regulations of that software, noting that the regulations of many vocal synthesis engines explicitly state that they cannot be used to convert input sources!
4. Engaging in illegal activities, as well as religious and political activities, is strictly prohibited when using this project. The project developers vehemently oppose the aforementioned activities. If you disagree with this provision, the usage of the project is prohibited.
5. If you continue to use the program, you will be deemed to have agreed to the terms and conditions set forth in README and README has discouraged you and is not responsible for any subsequent problems.
The singing voice conversion model uses SoftVC content encoder to extract speech features from the source audio. These feature vectors are directly fed into VITS without the need for conversion to a text-based intermediate representation. As a result, the pitch and intonations of the original audio are preserved. Meanwhile, the vocoder was replaced with [NSF HiFiGAN](https://github.com/openvpi/DiffSinger/tree/refactor/modules/nsf_hifigan) to solve the problem of sound interruption.
- Feature input is changed to the 12th Layer of [Content Vec](https://github.com/auspicious3000/contentvec) Transformer output, And compatible with 4.0 branches.
- To support the 4.0 model and incorporate the speech encoder, you can make modifications to the `config.json` file. Add the `speech_encoder` field to the "model" section as shown below:
- download model at [medium.pt](https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/medium.pt), the model fits `whisper-ppg`
- or download model at [large-v2.pt](https://openaipublic.azureedge.net/main/whisper/models/81f7c96c852ee8fc832187b0132e569d6c3065a3252ed18e56effd0b6a73e524/large-v2.pt), the model fits `whisper-ppg-large`
- download model at [chinese-hubert-large-fairseq-ckpt.pt](https://huggingface.co/TencentGameMate/chinese-hubert-large/resolve/main/chinese-hubert-large-fairseq-ckpt.pt)
- download model at [WavLM-Base+.pt](https://valle.blob.core.windows.net/share/wavlm/WavLM-Base+.pt?sv=2020-08-04&st=2023-03-01T07%3A51%3A05Z&se=2033-03-02T07%3A51%3A00Z&sr=c&sp=rl&sig=QJXmSJG9DbMKf48UDIU1MfzIro8HQOf3sqlNXiflY1I%3D), the model fits `wavlmbase+`
Diffusion model references [Diffusion-SVC](https://github.com/CNChTu/Diffusion-SVC) diffusion model. The pre-trained diffusion model is universal with the DDSP-SVC's. You can go to [Diffusion-SVC](https://github.com/CNChTu/Diffusion-SVC)'s repo to get the pre-trained diffusion model.
While the pretrained model typically does not pose copyright concerns, it is essential to remain vigilant. It is advisable to consult with the author beforehand or carefully review the description to ascertain the permissible usage of the model. This helps ensure compliance with any specified guidelines or restrictions regarding its utilization.
To avoid video memory overflow during training or pre-processing, it is recommended to limit the length of audio clips. Cutting the audio to a length of "5s - 15s" is more recommended. Slightly longer times are acceptable, however, excessively long clips may cause problems such as `torch.cuda.OutOfMemoryError`.
To facilitate the slicing process, you can use [audio-slicer-GUI](https://github.com/flutydeer/audio-slicer) or [audio-slicer-CLI](https://github.com/openvpi/audio-slicer)
In general, only the `Minimum Interval` needs to be adjusted. For spoken audio, the default value usually suffices, while for singing audio, it can be adjusted to around `100` or even `50`, depending on the specific requirements.
Although this project has resample.py scripts for resampling, mono and loudness matching, the default loudness matching is to match to 0db. This can cause damage to the sound quality. While python's loudness matching package pyloudnorm does not limit the level, this can lead to sonic boom. Therefore, it is recommended to consider using professional sound processing software, such as `adobe audition` for loudness matching. If you are already using other software for loudness matching, add the parameter `-skip_loudnorm` to the run command:
After enabling loudness embedding, the trained model will match the loudness of the input source; otherwise, it will match the loudness of the training set.
*`all_in_mem`: Load all dataset to RAM. It can be enabled when the disk IO of some platforms is too low and the system memory is **much larger** than your dataset.
*`vocoder_name`: Select a vocoder. The default is `nsf-hifigan`.
##### diffusion.yaml
*`cache_all_data`: Load all dataset to RAM. It can be enabled when the disk IO of some platforms is too low and the system memory is **much larger** than your dataset.
*`duration`: The duration of the audio slicing during training, can be adjusted according to the size of the video memory, **Note: this value must be less than the minimum time of the audio in the training set!**
*`batch_size`: The amount of data loaded to the GPU for a single training session can be adjusted to a size lower than the video memory capacity.
*`timesteps`: The total number of steps in the diffusion model, which defaults to 1000.
*`k_step_max`: Training can only train `k_step_max` step diffusion to save training time, note that the value must be less than `timesteps`, 0 is to train the entire diffusion model, **Note: if you do not train the entire diffusion model will not be able to use only_diffusion!**
-`-lg` | `--linear_gradient`: The cross fade length of two audio slices in seconds. If there is a discontinuous voice after forced slicing, you can adjust this value. Otherwise, it is recommended to use the default value of 0.
-`-f0p` | `--f0_predictor`: Select a F0 predictor, options are `crepe`, `pm`, `dio`, `harvest`, default value is `pm`(note: f0 mean pooling will be enable when using `crepe`)
-`-a` | `--auto_predict_f0`: automatic pitch prediction, do not enable this when converting singing voices as it can cause serious pitch issues.
-`-cm` | `--cluster_model_path`: Cluster model or feature retrieval index path, if left blank, it will be automatically set as the default path of these models. If there is no training cluster or feature retrieval, fill in at will.
-`-cr` | `--cluster_infer_ratio`: The proportion of clustering scheme or feature retrieval ranges from 0 to 1. If there is no training clustering model or feature retrieval, the default is 0.
-`-eh` | `--enhance`: Whether to use NSF_HIFIGAN enhancer, this option has certain effect on sound quality enhancement for some models with few training sets, but has negative effect on well-trained models, so it is disabled by default.
-`-shd` | `--shallow_diffusion`: Whether to use shallow diffusion, which can solve some electrical sound problems after use. This option is disabled by default. When this option is enabled, NSF_HIFIGAN enhancer will be disabled
-`-usm` | `--use_spk_mix`: whether to use dynamic voice fusion
-`-lea` | `--loudness_envelope_adjustment`：The adjustment of the input source's loudness envelope in relation to the fusion ratio of the output loudness envelope. The closer to 1, the more the output loudness envelope is used
-`-fr` | `--feature_retrieval`：Whether to use feature retrieval If clustering model is used, it will be disabled, and `cm` and `cr` parameters will become the index path and mixing ratio of feature retrieval
-`-ks` | `--k_step`: The larger the number of k_steps, the closer it is to the result of the diffusion model. The default is 100
-`-od` | `--only_diffusion`: Whether to use Only diffusion mode, which does not load the sovits model to only use diffusion model inference
-`-se` | `--second_encoding`：which involves applying an additional encoding to the original audio before shallow diffusion. This option can yield varying results - sometimes positive and sometimes negative.
If you are satisfied with the previous results, or if you do not feel you understand what follows, you can skip it and it will have no effect on the use of the model. The impact of these optional settings mentioned is relatively small, and while they may have some impact on specific datasets, in most cases the difference may not be significant.
During the training of the 4.0 model, an f0 predictor is also trained, which enables automatic pitch prediction during voice conversion. However, if the results are not satisfactory, manual pitch prediction can be used instead. Please note that when converting singing voices, it is advised not to enable this feature as it may cause significant pitch shifting.
- Set `auto_predict_f0` to `true` in `inference_main.py`.
Introduction: The clustering scheme implemented in this model aims to reduce timbre leakage and enhance the similarity of the trained model to the target's timbre, although the effect may not be very pronounced. However, relying solely on clustering can reduce the model's clarity and make it sound less distinct. Therefore, a fusion method is adopted in this model to control the balance between the clustering and non-clustering approaches. This allows manual adjustment of the trade-off between "sounding like the target's timbre" and "have clear enunciation" to find an optimal balance.
- Train on a machine with good CPU performance. According to extant experience, it takes about 4 minutes to train each speaker on a Tencent Cloud machine with 6-core CPU.
- Specify `cluster_infer_ratio` in `inference_main.py`, where `0` means not using clustering at all, `1` means only using clustering, and usually `0.5` is sufficient.
Introduction: As with the clustering scheme, the timbre leakage can be reduced, the enunciation is slightly better than clustering, but it will reduce the inference speed. By employing the fusion method, it becomes possible to linearly control the balance between feature retrieval and non-feature retrieval, allowing for fine-tuning of the desired proportion.
- Specify `cluster_infer_ratio` in `inference_main.py`, where `0` means not using feature retrieval at all, `1` means only using feature retrieval, and usually `0.5` is sufficient.
### [](https://colab.research.google.com/github/svc-develop-team/so-vits-svc/blob/4.1-Stable/sovits4_for_colab.ipynb) [sovits4_for_colab.ipynb](https://colab.research.google.com/github/svc-develop-team/so-vits-svc/blob/4.1-Stable/sovits4_for_colab.ipynb)
The generated model contains data that is needed for further training. If you confirm that the model is final and not be used in further training, it is safe to remove these data to get smaller file size (about 1/3).
Introduction: This function can combine multiple models into one model (convex combination or linear combination of multiple model parameters) to create mixed voice that do not exist in reality
2. If you force a multi-speaker model, it is critical to make sure there are the same number of speakers in each model. This will ensure that sounds with the same SpeakerID can be mixed correctly.
The start time must be the same as the end time of the previous one. The first start time must be 0, and the last end time must be 1 (time ranges from 0 to 1).
- Modify `"NyaruTaffy"` in `path = "NyaruTaffy"` in [onnx_export.py](https://github.com/svc-develop-team/so-vits-svc/blob/4.0/onnx_export.py) to your project name, `path = "aziplayer"`（onnx_export_speaker_mix makes you can mix speaker's voice）
Note: For Hubert Onnx models, please use the models provided by MoeSS. Currently, they cannot be exported on their own (Hubert in fairseq has many unsupported operators and things involving constants that can cause errors or result in problems with the input/output shape and results when exported.)
|[2106.06103](https://arxiv.org/abs/2106.06103) | VITS (Synthesizer)| Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech | [jaywalnut310/vits](https://github.com/jaywalnut310/vits) |
|[2111.02392](https://arxiv.org/abs/2111.02392) | SoftVC (Speech Encoder)| A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion | [bshall/hubert](https://github.com/bshall/hubert) |
|[2204.09224](https://arxiv.org/abs/2204.09224) | ContentVec (Speech Encoder)| ContentVec: An Improved Self-Supervised Speech Representation by Disentangling Speakers | [auspicious3000/contentvec](https://github.com/auspicious3000/contentvec) |
|[2110.13900](https://arxiv.org/abs/2110.13900) | WavLM (Speech Encoder) | WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing | [microsoft/unilm/wavlm](https://github.com/microsoft/unilm/tree/master/wavlm) |
|[2305.17651](https://arxiv.org/abs/2305.17651) | DPHubert (Speech Encoder) | DPHuBERT: Joint Distillation and Pruning of Self-Supervised Speech Models | [pyf98/DPHuBERT](https://github.com/pyf98/DPHuBERT) |
|[DOI:10.21437/Interspeech.2017-68](http://dx.doi.org/10.21437/Interspeech.2017-68) | Harvest (F0 Predictor) | Harvest: A high-performance fundamental frequency estimator from speech signals | [mmorise/World/harvest](https://github.com/mmorise/World/blob/master/src/harvest.cpp) |
|[aes35-000039](https://www.aes.org/e-lib/online/browse.cfm?elib=15165) | Dio (F0 Predictor) | Fast and reliable F0 estimation method based on the period extraction of vocal fold vibration of singing voice and speech | [mmorise/World/dio](https://github.com/mmorise/World/blob/master/src/dio.cpp) |
|[8461329](https://ieeexplore.ieee.org/document/8461329) | Crepe (F0 Predictor) | Crepe: A Convolutional Representation for Pitch Estimation | [maxrmorrison/torchcrepe](https://github.com/maxrmorrison/torchcrepe) |
|[DOI:10.1016/j.wocn.2018.07.001](https://doi.org/10.1016/j.wocn.2018.07.001) | Parselmouth (F0 Predictor) | Introducing Parselmouth: A Python interface to Praat | [YannickJadoul/Parselmouth](https://github.com/YannickJadoul/Parselmouth) |
|[2010.05646](https://arxiv.org/abs/2010.05646) | HIFIGAN (Vocoder) | HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis | [jik876/hifi-gan](https://github.com/jik876/hifi-gan) |
|[1810.11946](https://arxiv.org/abs/1810.11946.pdf) | NSF (Vocoder) | Neural source-filter-based waveform model for statistical parametric speech synthesis | [openvpi/DiffSinger/modules/nsf_hifigan](https://github.com/openvpi/DiffSinger/tree/refactor/modules/nsf_hifigan)
|[2006.08195](https://arxiv.org/abs/2006.08195) | Snake (Vocoder) | Neural Networks Fail to Learn Periodic Functions and How to Fix It | [EdwardDixon/snake](https://github.com/EdwardDixon/snake)
|[K-means](https://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=01D65490BADCC216F350D06F84D721AD?doi=10.1.1.308.8619&rep=rep1&type=pdf) | Feature K-means Clustering (PreProcessing)| Some methods for classification and analysis of multivariate observations | This repo |
For some reason the author deleted the original repository. Because of the negligence of the organization members, the contributor list was cleared because all files were directly reuploaded to this repository at the beginning of the reconstruction of this repository. Now add a previous contributor list to README.md.
*Some members have not listed according to their personal wishes.*
