Free AI Audio Splitter

A finished job means the model produced files, not that it recreated the original studio stems. Dense arrangements, distorted singing, chorus effects, live audience noise, and reverb can make clean separation difficult. Quiet artifacts may become more noticeable when a track plays alone. The output can still be useful for listening and practice, but critical production work needs your own quality judgment. We do not promise lossless conversion, perfect vocal removal, or studio-quality isolation. Try a cleaner source or a shorter section if a result is disappointing. This version provides one model configuration so the first workflow stays straightforward.

Your audio is uploaded and processed by the Demucs model hosted on Replicate. The website saves the separated tracks in private storage for playback and download through your account. Original uploads are scheduled for cleanup after the task ends. Free-account results are available for 24 hours after completion, then cleaned up. You can also delete completed results earlier. Closing the browser does not immediately remove the files from the service. Downloads saved to your device remain under your control and are not removed when you delete a task. See the privacy notice for provider processing and retention details.

Upload an MP3 or WAV between 0.5 seconds and five minutes, up to 30 MB. Each completed task produces two 44.1 kHz stereo WAV files with 16-bit PCM samples. Conversion may involve resampling and quantization, so the original encoding is not preserved. Export audio from videos first. MP3 output and separate instrument stems are not included. Re-export invalid files from your audio application before trying again. No local model service needs to be installed or started. A temporary progress-query failure keeps the existing task for another check; avoid submitting it again.

Yes. After Google sign-in, the Free plan includes 10 minutes per month, Pro includes 250, and Studio includes 600. Usage counts the source audio duration once, rounded up to the next whole second, not each output track. Starting a task reserves its duration; a confirmed failed task releases that reservation. Deleting completed results does not restore used minutes. Annual plans also refresh monthly, and unused monthly minutes do not roll over. The tool shows your remaining monthly allowance. See the pricing page if you need more minutes; the free plan does not provide unlimited processing.

The available downloads are vocals and instrumental only. There are no separate buttons for drums, bass, guitar, or piano, and this workflow is not a dedicated noise-reduction or speech-enhancement tool. Do not expect it to clean every noisy voice recording. If your goal is simply to shorten a track, use the audio cutter instead; that operation selects a time range without trying to recognize instruments. If your goal is to understand which kind of splitting you need, the tool guide explains the difference. Keeping these tasks separate makes the controls easier to understand and the output easier to evaluate.