SATB separation guide

Hear every voice in the choir.

Cantor is a browser-based AI choir stem separator. It turns a mixed vocal harmony recording into individual soprano, alto, tenor, and bass tracks for rehearsal, transcription, and closer listening.

Request beta access Hear a real result

What Cantor is for

Cantor is designed for choir directors, section leaders, singers, arrangers, teachers, and music researchers who need to hear one voice part inside a polyphonic choral recording. A director can create rehearsal references, a singer can focus on their line, and a transcriber can inspect overlapping parts.

It is a separation tool, not a score generator. The result is audio, with one stem for each selected voice.

Supported recordings and formats

Voices: Soprano, alto, tenor, and bass, including recordings that contain only a selected subset.
File formats: WAV, FLAC, MP3, M4A, OGG, and OPUS within the current upload limit.
Online sources: Supported public YouTube links can be used in the private application.
Output: Separate WAV stems with browser preview and download controls.
Processing: The complete source or a selected time region can be processed.

Where results can be weaker

Separation is most reliable when SATB parts are clearly sung and the recording has limited room echo, distortion, and accompaniment. When a source also contains instruments or a prominent lead vocalist, it may first need separate stages to isolate the vocal ensemble before Cantor can extract the SATB parts.

Stage 1 Full music mix Choir, lead vocal, and instruments overlap.

Stage 2 Isolate ensemble vocals Instrumental and lead-vocal separation may introduce artefacts.

Stage 3 Extract SATB parts Cantor works with the audio preserved by the earlier stages.

Every additional separation stage can remove or alter useful vocal detail. Instrument bleed, missing harmonics, softened consonants, phase artefacts, and traces of the lead vocal can therefore carry into the SATB result. For the best quality, use the cleanest choir or backing-vocal recording available and avoid unnecessary preprocessing.

Results can also be weaker with dense unison singing, unusual voice ranges, heavy pitch correction, clipping, crowd noise, or strong reverberation. Overlapping voices may leave traces in another stem.

Cantor is a research preview. Listen critically before using a stem for teaching, performance decisions, publication, or research conclusions.

A real Cantor audio example

This is an actual 30-second separation, not a reconstruction or mock-up. Start with the mixture, then compare each generated voice stem.

0:00

0:30

Source audio: Cantoría Dataset, licensed CC BY 4.0. Separation by Cantor.

Privacy and retention

Audio jobs and account data are private. Uploaded audio and generated stems are retained for up to 30 days so an account holder can reopen or rerun a result. The account holder can delete a job's stored audio sooner. Audio is not used for model training without separate, explicit consent.

Cantor records aggregate daily referral-source totals for public marketing pages. It does not store a visitor's full referring URL, IP address, device fingerprint, or live-presence record for this measurement. Read the full privacy notice.

Current beta availability

Cantor is in a limited private beta. Access is currently free and granted by request while capacity, output quality, and the correction workflow are tested. Joining the list does not create an account immediately.

Request beta access Ask a question