cottom/rvc-train-voice-model

RVC v2 voice model training from any recordings: vocal extraction, harmony and reverb removal, denoise, breath removal, RMVPE, HuBERT, training and index.

Public

8 runs

Playground API Examples README Versions

Run cottom/rvc-train-voice-model with an API

Use one of our client libraries to get started quickly. Clicking on a library will take you to the Playground tab where you can tweak different inputs, see the results, and copy the corresponding code to use in your own project.

Input schema

The fields you can use to run this model with an API. If you don't give a value for a field its default value will be used.

Field	Type	Default value	Description
dataset	string		One recording, or a .zip of recordings (wav, mp3, flac, m4a, ogg, ...). Songs, a cappella and speech are all cleaned to the dry lead voice first. 1 minute minimum, 10+ recommended, 30 maximum.
sample_rate	None	40k	Model sample rate.
epochs	integer	0 Max: 1000	Training epochs. 0 picks about 4,000 steps for the cleaned dataset (100-500 epochs).
batch_size	integer	8 Min: 2 Max: 32	Segments per training step.
separation	None	auto	Vocal extraction. auto: per 10-second window, only where there is a backing track or bleed. always: everywhere. never: train on the audio as uploaded.
lead_vocal_only	None	auto	Remove harmonies and backing vocals (the main cause of cracked, hoarse notes). auto: wherever a backing track was found.
remove_reverb	boolean	True	Remove reverb and echo (the second cause of cracked notes).
denoise	boolean	True	Gate steady noise (hiss, hum) against the recording's quietest moments.
remove_breaths	boolean	True	Silence breaths: quiet unvoiced stretches longer than a consonant.
residue_floor_db	number	-50 Min: -90 Max: -30	Delete spectral content this far below the voice (separation residue). -90 keeps everything.
time_limit_minutes	number	28 Min: 10 Max: 180	Whole-run budget. Training stops early, keeping its latest saved weights, to finish inside it. Replicate stops a prediction after 30 minutes of running.

{
  "type": "object",
  "title": "Input",
  "required": [
    "dataset"
  ],
  "properties": {
    "epochs": {
      "type": "integer",
      "title": "Epochs",
      "default": 0,
      "maximum": 1000,
      "minimum": 0,
      "x-order": 2,
      "description": "Training epochs. 0 picks about 4,000 steps for the cleaned dataset (100-500 epochs)."
    },
    "dataset": {
      "type": "string",
      "title": "Dataset",
      "format": "uri",
      "x-order": 0,
      "description": "One recording, or a .zip of recordings (wav, mp3, flac, m4a, ogg, ...). Songs, a cappella and speech are all cleaned to the dry lead voice first. 1 minute minimum, 10+ recommended, 30 maximum."
    },
    "denoise": {
      "type": "boolean",
      "title": "Denoise",
      "default": true,
      "x-order": 7,
      "description": "Gate steady noise (hiss, hum) against the recording's quietest moments."
    },
    "batch_size": {
      "type": "integer",
      "title": "Batch Size",
      "default": 8,
      "maximum": 32,
      "minimum": 2,
      "x-order": 3,
      "description": "Segments per training step."
    },
    "separation": {
      "enum": [
        "auto",
        "always",
        "never"
      ],
      "type": "string",
      "title": "separation",
      "description": "Vocal extraction. auto: per 10-second window, only where there is a backing track or bleed. always: everywhere. never: train on the audio as uploaded.",
      "default": "auto",
      "x-order": 4
    },
    "sample_rate": {
      "enum": [
        "40k",
        "48k"
      ],
      "type": "string",
      "title": "sample_rate",
      "description": "Model sample rate.",
      "default": "40k",
      "x-order": 1
    },
    "remove_reverb": {
      "type": "boolean",
      "title": "Remove Reverb",
      "default": true,
      "x-order": 6,
      "description": "Remove reverb and echo (the second cause of cracked notes)."
    },
    "remove_breaths": {
      "type": "boolean",
      "title": "Remove Breaths",
      "default": true,
      "x-order": 8,
      "description": "Silence breaths: quiet unvoiced stretches longer than a consonant."
    },
    "lead_vocal_only": {
      "enum": [
        "auto",
        "always",
        "never"
      ],
      "type": "string",
      "title": "lead_vocal_only",
      "description": "Remove harmonies and backing vocals (the main cause of cracked, hoarse notes). auto: wherever a backing track was found.",
      "default": "auto",
      "x-order": 5
    },
    "residue_floor_db": {
      "type": "number",
      "title": "Residue Floor Db",
      "default": -50,
      "maximum": -30,
      "minimum": -90,
      "x-order": 9,
      "description": "Delete spectral content this far below the voice (separation residue). -90 keeps everything."
    },
    "time_limit_minutes": {
      "type": "number",
      "title": "Time Limit Minutes",
      "default": 28,
      "maximum": 180,
      "minimum": 10,
      "x-order": 10,
      "description": "Whole-run budget. Training stops early, keeping its latest saved weights, to finish inside it. Replicate stops a prediction after 30 minutes of running."
    }
  }
}

Output schema

The shape of the response you’ll get when you run this model with an API.

Schema

{
  "type": "array",
  "items": {
    "type": "string",
    "format": "uri"
  },
  "title": "Output"
}