ComfyUI Node: AudioSeparation

AudioSeparation:

The AudioSeparation node is designed to isolate different components of an audio track, such as bass, drums, other instruments, and vocals. This node leverages advanced machine learning models to perform source separation, allowing you to extract individual elements from a mixed audio file. This can be particularly useful for remixing, creating karaoke tracks, or analyzing specific parts of a song. The node processes the audio by resampling it to match the model's expected sample rate, normalizing the waveform, and then applying a pre-trained model to separate the sources. The result is a set of isolated audio tracks that can be used independently or recombined as needed.

AudioSeparation Input Parameters:

audio

This parameter accepts an audio file in the form of a dictionary containing the waveform and sample rate. The waveform is a tensor representation of the audio signal, and the sample rate is the number of samples per second. This input is required for the node to function.

chunk_fade_shape

This optional parameter defines the shape of the fade applied to the chunks of audio during processing. The available options are "linear", "half_sine", "logarithmic", and "exponential". The default value is "linear". This parameter affects how smoothly the transitions between chunks are handled, which can impact the quality of the separated audio.

chunk_length

This optional parameter specifies the length of each audio chunk in seconds. The default value is 10.0 seconds. Adjusting this parameter can influence the performance and accuracy of the source separation, as shorter chunks may process faster but could result in less accurate separation.

chunk_overlap

This optional parameter determines the overlap between consecutive audio chunks, expressed as a fraction of the chunk length. The default value is 0.1 (10%). Overlapping chunks can help to ensure smoother transitions and reduce artifacts in the separated audio.

AudioSeparation Output Parameters:

Bass

This output provides the isolated bass component of the input audio. It is one of the four separated sources and can be used independently or in combination with the other outputs.

Drums

This output provides the isolated drums component of the input audio. It is one of the four separated sources and can be used independently or in combination with the other outputs.

Other

This output provides the isolated "other" instruments component of the input audio. It includes all elements that are not classified as bass, drums, or vocals. It is one of the four separated sources and can be used independently or in combination with the other outputs.

Vocals

This output provides the isolated vocals component of the input audio. It is one of the four separated sources and can be used independently or in combination with the other outputs.

AudioSeparation Usage Tips:

• For best results, ensure that the input audio is of high quality and has a consistent sample rate.
• Experiment with different chunk_fade_shape options to find the one that produces the smoothest transitions for your specific audio.
• Adjust the chunk_length and chunk_overlap parameters to balance between processing time and separation accuracy. Shorter chunks may process faster but could result in less accurate separation.

AudioSeparation Common Errors and Solutions:

"Input audio sample rate does not match model sample rate"

• Explanation: The sample rate of the input audio does not match the sample rate expected by the model.
• Solution: Ensure that the input audio is resampled to the model's expected sample rate before processing.

"Model failed to separate sources"

• Explanation: The model encountered an issue while trying to separate the audio sources.
• Solution: Check the input audio for any anomalies or corruption. Ensure that the audio is of high quality and try adjusting the chunk_length and chunk_overlap parameters.

"CUDA out of memory"

• Explanation: The GPU does not have enough memory to process the audio.
• Solution: Reduce the chunk_length or process the audio on a machine with more GPU memory. Alternatively, try running the node on a CPU if GPU resources are limited.