Accelerating Diffusion-based Singing Voice Conversion through Consistency Distillation

This is an implement of Consistency Models for accelerating diffusion-based singing voice conversion. The overall architecture follows "Leveraging Content-based Features from Multiple Acoustic Models for Singing Voice Conversion" (NeurIPS 2023 Workshop on Machine Learning for Audio), only a slightly modification is applied on acoustic model. Specifically,

The acoustic model is a conformer which generates a coarse spectrogram and a diffusion decoder based on Bidirectional Non-Causal Dilated CNN which polish the former spectrogram for better. This is similar to CoMoSpeech: One-Step Speech and Singing Voice Synthesis via Consistency Model
To accelerate diffusion model, we apply consistency distillation from Consistency Models. For teacher model, the diffusion schedule of the diffusion decoder follows karras diffusion. For distilling teacher model, the condition encoder and the conformer part of acoustic model are frozen while the diffusion decoder model is updated via exponential moving average. See Figure above for details.

There are five stages in total:

Data preparation
Features extraction
Teacher Model Training
Consistency Distillation
Inference/conversion

1. Data Preparation

Dataset Download

By default, we utilize the five datasets for training: M4Singer, Opencpop, OpenSinger, SVCC, and VCTK. How to download them is detailed here.

Configuration

Specify the dataset paths in exp_config.json. Note that you can change the dataset list to use your preferred datasets.

    "dataset": [
        "m4singer",
        "opencpop",
        "opensinger",
        "svcc",
        "vctk"
    ],
    "dataset_path": {
        // TODO: Fill in your dataset path
        "m4singer": "[M4Singer dataset path]",
        "opencpop": "[Opencpop dataset path]",
        "opensinger": "[OpenSinger dataset path]",
        "svcc": "[SVCC dataset path]",
        "vctk": "[VCTK dataset path]"
    },

2. Features Extraction

Content-based Pretrained Models Download

By default, we utilize the Whisper and ContentVec to extract content features. How to download them is detailed here.

Configuration

Specify the dataset path and the output path for saving the processed data and the training model in exp_config.json:

    // TODO: Fill in the output log path
    "log_dir": "[Your path to save logs and checkpoints]",
    "preprocess": {
        // TODO: Fill in the output data path
        "processed_dir": "[Your path to save processed data]",
        ...
    },

Run

Run the run.sh as the preproces stage (set --stage 1).

cd Amphion
sh egs/svc/DiffComoSVC/run.sh --stage 1

Note: The CUDA_VISIBLE_DEVICES is set as "0" in default. You can change it when running run.sh by specifying such as --gpu "1".

3. Teacher Model Training

Configuration

Set the distill in config/comosvc.json to false for teacher model training, you can also specify the detailed configuration for conformer encoder and diffusion process here:

"comosvc":{
            "distill": false,
            // conformer encoder
            "input_dim": 384,
            "output_dim": 100,
            "n_heads": 2,
            "n_layers": 6,
            "filter_channels":512,
            // karras diffusion
            "P_mean": -1.2,
            "P_std": 1.2,
            "sigma_data": 0.5,
            "sigma_min": 0.002,
            "sigma_max": 80,
            "rho": 7,
            "n_timesteps": 40,
        },

We provide the default hyparameters in the exp_config.json. They can work on single NVIDIA-24g GPU. You can adjust them based on you GPU machines.

"train": {
        "batch_size": 32,
        ...
        "adamw": {
            "lr": 2.0e-4
        },
        ...
    }

Run

Run the run.sh as the training stage (set --stage 2). Specify a experimental name to run the following command. The tensorboard logs and checkpoints will be saved in [Your path to save logs and checkpoints]/[YourExptName].

cd Amphion
sh egs/svc/DiffComoSVC/run.sh --stage 2 --name [YourExptName]

Note: The CUDA_VISIBLE_DEVICES is set as "0" in default. You can specify it when running run.sh such as:

cd Amphion
sh egs/svc/DiffComoSVC/run.sh --stage 2 --name [YourExptName] --gpu "0,1,2,3"

4. Consistency Distillation

Configuration

Set the distill in config/comosvc.json to true for teacher model training, and specify the teacher_model_path for consistency distillation. You can also specify the detailed configuration for conformer encoder and diffusion process here:

"model": {
    "teacher_model_path":"[Your_teacher_model_checkpoint].bin",
    ...
    "comosvc":{
                "distill": true,
                // conformer encoder
                "input_dim": 384,
                "output_dim": 100,
                "n_heads": 2,
                "n_layers": 6,
                "filter_channels":512,
                // karras diffusion
                "P_mean": -1.2,
                "P_std": 1.2,
                "sigma_data": 0.5,
                "sigma_min": 0.002,
                "sigma_max": 80,
                "rho": 7,
                "n_timesteps": 40,
            },

We provide the default hyparameters in the exp_config.json. They can work on single NVIDIA-24g GPU. You can adjust them based on you GPU machines.

"train": {
        "batch_size": 32,
        ...
        "adamw": {
            "lr": 2.0e-4
        },
        ...
    }

Run

cd Amphion
sh egs/svc/DiffComoSVC/run.sh --stage 2 --name [YourExptName]

Note: The CUDA_VISIBLE_DEVICES is set as "0" in default. You can specify it when running run.sh such as:

cd Amphion
sh egs/svc/DiffComoSVC/run.sh --stage 2 --name [YourExptName] --gpu "0,1,2,3"

5. Inference/Conversion

Pretrained Vocoder Download

We fine-tune the official BigVGAN pretrained model with over 120 hours singing voice data. The benifits of fine-tuning has been investigated in our paper (see this demo page). The final pretrained singing voice vocoder is released here (called Amphion Singing BigVGAN).

Run

For inference/conversion, you need to specify the following configurations when running run.sh:

Parameters	Description	Example
`--infer_expt_dir`	The experimental directory which contains `checkpoint`	`[Your path to save logs and checkpoints]/[YourExptName]`
`--infer_output_dir`	The output directory to save inferred audios.	`[Your path to save logs and checkpoints]/[YourExptName]/result`
`--infer_source_file` or `--infer_source_audio_dir`	The inference source (can be a json file or a dir).	The `infer_source_file` could be `[Your path to save processed data]/[YourDataset]/test.json`, and the `infer_source_audio_dir` is a folder which includes several audio files (.wav, .mp3 or *.flac).
`--infer_target_speaker`	The target speaker you want to convert into. You can refer to `[Your path to save logs and checkpoints]/[YourExptName]/singers.json` to choose a trained speaker.	For opencpop dataset, the speaker name would be `opencpop_female1`.
`--infer_key_shift`	How many semitones you want to transpose.	`"autoshfit"` (by default), `3`, `-3`, etc.

For example, if you want to make opencpop_female1 sing the songs in the [Your Audios Folder], just run:

cd Amphion
sh egs/svc/DiffComoSVC/run.sh --stage 3 --gpu "0" \
    --infer_expt_dir [Your path to save logs and checkpoints]/[YourExptName] \
    --infer_output_dir [Your path to save logs and checkpoints]/[YourExptName]/result \
    --infer_source_audio_dir [Your Audios Folder] \
    --infer_target_speaker "opencpop_female1" \
    --infer_key_shift "autoshift"

Specially, you can configurate the inference steps for teacher model by setting inference at exp_config(student model is always one-step sampling):

    "inference": {
        "comosvc": {
            "inference_steps": 40
        }
    }

Reference

https://github.com/zhenye234/CoMoSpeech

https://github.com/openai/consistency_models