Skip to content

Analytics cookies

Optional analytics for visits, tool use and purchase clicks. No audio, filenames or download links. Change your choice in the footer.

Optional · Privacy
Tools

Vocals & backing tracks

No-upload separatorFree vocals and instrumental WAVs on your device. Experimental.Vocal removerKeep the backing music. Make room to sing.Vocal isolatorIsolate the voice to study phrasing, diction and harmonies.Acapella extractorHear the voice on its own.Backing trackMake an instrumental or karaoke backing track. Free WAV download.

Stems & instruments

Stem splitterExplore vocals, drums, bass and other.Drum extractorIsolate drums, or use the other stems for a drumless mix.Bass extractorDownload an isolated bass line for learning or arranging.

Tempo

Tap tempoTap a live rhythm with your mouse, phone or spacebar.MetronomeSet a steady practice click with accents and subdivisions.

Practice & sound

Audio practiceLoop, slow down or change the pitch of saved audio. No upload.Bass boostLift the low end, compare and save a WAV. No upload.
Audio guidesPricing & free limits
Audio guidesPricingAccountSeparate audio
← BACK TO THE TOOL

Privacy policy

Last updated November 12, 2026

This policy describes the free separation, tempo, bass boost and audio practice tools at vocalseparator.com. Listening examples use publicly licensed audio and recorded processing results. Playing an example does not upload your files. Selecting a file for cloud separation first previews it in your browser. Uploading starts when you confirm your audio rights and choose Separate audio, Split into 4 stems, Prepare bass practice or the corresponding extract/remove action on a separation page. Tap tempo measures taps locally. The BPM counter can analyze an audio file on your device, and bass boost applies EQ and prepares a WAV locally, without uploading audio. Audio practice plays your saved file locally, with waveform selection, loop, speed and pitch controls, and MP3/WAV practice export. The result workbench reads your existing stems and applies mixing and practice changes on your device. These local tools do not record from your microphone. Audio, filenames, tap times, BPM estimates and boost settings are not saved to browser storage. Clear the file or leave the page to release local previews and exported audio. Practice settings are kept only in memory unless you explicitly choose to save them as described below.

What we process

The no-upload separator at /local-separator runs on your device. Audio, filenames and results are not sent to us, an inference provider or analytics. It does not use cloud job allowances or create a separation job on our servers. The UVR MDX-Net model is downloaded from a byte-verified, fixed-revision Hugging Face mirror of the official GitHub release. That host receives ordinary network information for those requests, such as your IP address. Our hosting provider serves the page and runtime files. Only public model files are stored in its browser model cache; use Clear model cache to remove them. Closing the tab releases its audio and results, except files you have explicitly saved. After preparing the engine, the current tab can process and save audio offline; offline page reloads are not supported. It never falls back to a cloud upload. Its optional Share anonymous usage switch is off by default and applies only to the current tab. If you enable it, fixed action counts are sent to our first-party counter while online. They include engine readiness, separation outcomes, playback, practice controls and download requests, never audio, filenames, settings or error text. Offline actions are not queued or sent later. The switch can be turned off at any time; this page uses the site-wide analytics choice for optional Google Analytics, Microsoft Clarity and PVUV visits and purchase events. Buying opens Waffo Pancake in a separate tab; no audio is attached. Local entitlement is checked against our server ledger using your verified account or an essential, HttpOnly guest receipt cookie kept for up to one year. The cookie stores a random credential, never audio or a paid flag. Clearing it requires signing in with your checkout email to restore a confirmed purchase. The purchase record is retained for access, accounting and dispute handling. The 35-day retention and privacy signals described below also apply to these optional counts.

Metronome clicks are generated on your device without an audio file or microphone. Its settings stay in memory. Choosing Use in metronome includes the displayed BPM in the destination URL to fill the tempo; that URL may appear in browser history and hosting access logs. It contains no audio, filename or tap times, and does not start playback.

Choosing Save practice setup stores a local file fingerprint (or a hashed result identifier), duration, practice part, speed, pitch, track levels, mute/solo settings, named passages, loop setting, A/B points and save time in this browser. The fingerprint matches the same file when you select it again; audio, source URLs and filenames are not stored. These records are not sent to our server or analytics. Up to 20 setups are kept for 30 days after saving, with expired records removed the next time they are read. You can forget an individual setup or all saved setups in Audio Practice. Clearing this site’s browser storage also removes them. Restoring requires selecting the file again, or recovering an unexpired result, and choosing Restore saved setup; it does not autoplay. Saving settings does not extend the lifetime of separation results. A local clear revision, containing no audio or personal data, prevents a pending save in another tab from undoing Forget all.

When you choose free cloud separation, we process the file, its filename, size and duration. No registration or email address is required. We use a random browser identifier for your daily allowance and a keyed hash of your IP address to limit abuse; application job records do not store the raw IP. Our hosting providers may maintain operational security logs under their own policies.

Before you select a file, we check current free processing availability using that essential browser identifier and network hash. Checking availability does not upload audio, reserve processing or use a free job. The same checks run again when you start separation.

Optional accounts use Neon Auth for email or Google sign-in. With Google, your Google account identifier, verified email and basic profile identify your account. Signing in does not grant extra free uses. When you process a free cloud file while signed in, or choose Save to my account, we link that temporary result to your account so you can find it on another device. The link and job metadata expire within the original 25-hour job record period; downloads still expire 55 minutes after processing. We do not copy the audio or extend its storage when you save a result to your account.

Where your file goes

Paid accounts use Neon Auth to verify your email and manage login sessions. Neon Postgres stores your account identifier, purchase references, minute ledger, private batch metadata and refund status. Your selected payment provider, Waffo Pancake or Creem, handles billing details, wallet authorization and card data on its hosted checkout. We receive verified payment references and your checkout email to deliver your minutes; we do not receive full card details. Necessary account cookies keep you signed in. Private account pages, checkout tokens and batch names are excluded from our Google Analytics page tracking.

Uploads are sent to private Cloudflare R2 storage. Replicate receives a temporary signed link to run audio separation. The website runs on Vercel, and an Upstash-compatible Redis service stores short-lived job records and usage counters. These providers may process data outside your country. We do not create a public audio library or use your uploads to train our own models.

Cloud model attribution and audio rights

Cloud separation uses Deezer Spleeter. Copyright Deezer SA; MIT license. The Spleeter JOSS paper states that the source code and pretrained models were released under MIT. The current Spleeter README explicitly labels the code MIT; the maintainers have not provided written confirmation of commercial use of the weights. Training-data rights have not been verified. See the cloud model notice and MIT text. A model license does not grant you rights to the source audio: you must own it or have permission to process it.

When bass practice is enabled, our server reads the existing four stems to make listening and practice mixes. This does not send another separation request. These mixes are kept in private R2 storage and share the original result’s download deadline. They are scheduled for deletion when that deadline passes and removed by a subsequent daily cleanup. Download expiry does not mean immediate physical deletion. The one-day lifecycle backup described below also applies. We remove temporary mixing files during normal completion or failure cleanup.

Retention and downloads

The practice workbench reads existing WAV stems in your browser and prepares local MP3 or WAV audio with your chosen levels, speed, pitch and passage. It makes no new separation request and uploads no mix. Prepared audio stays in memory until you leave; save it before closing the page or the result links expire. Settings stay in memory unless you explicitly save them as described above.

Free uploads and processing copies are scheduled for deletion 23 hours after an upload begins. Cleanup runs once a day and processes due files on its next run, normally around 23–48 hours after the upload begins. A one-day R2 lifecycle rule also expires stored files as a backup; physical deletion is asynchronous and can happen after that eligibility time. If cleanup is interrupted, deletion may take longer. Download access expires separately from physical file deletion.

Paid uploads and processing copies use separate private storage prefixes and are deleted during paid maintenance after their jobs finish, or after an abandoned upload has been inactive for a day. Completed paid WAV results are available for seven days after each file is ready; access then ends and maintenance schedules deletion. An eight-day lifecycle backstop is used for paid results, independently of the free one-day rule. Physical deletion is asynchronous. Account, payment and minute-ledger records remain while the account or purchased balance is active and for legitimate accounting, dispute and legal requirements. Contact support for access, correction or deletion requests; records required for those obligations may be retained.

Replicate removes API input parameters, output files and logs after approximately one hour. Free cloud separation downloads expire 55 minutes after processing finishes, independently of the upload deletion schedule. Job metadata expires after 25 hours, and daily usage counters expire within 48 hours. Failure-refund markers expire within 25 hours of the latest confirmed status check. We also keep a cumulative, site-wide processing count without audio or browser identifiers to track overall usage. Read Replicate’s retention policy.

Service reliability records

We keep anonymous daily totals of task outcomes, audio duration and reported processing time to detect failures and control costs. These totals contain no audio, filenames, download links, IP addresses or browser identifiers, and expire within 35 days of their last update. Maintenance summaries are kept for up to seven days. Temporary upload reservations expire within 15 minutes. Job records also hold the processing status and temporary result links so a refresh can recover completed work; saving a link does not extend the file’s lifetime.

Optional usage measurement

When enabled, we count tool views and actions such as file selection, playback, download preparation, practice download link requests and links between tools to improve them. Counts are grouped by tool and a fixed page label. For audio practice, we count file selection and playback, setup save/restore/removal, clip and pitch preparation, playback, download requests and export failures; loop and speed actions are also counted on result pages. These events contain no filenames, loop positions, speed values or semitone changes. For the connected workbench, we count opening, actual playback, setup save/restore, export preparation, preview, download requests and failures. The first export failure on a page may include a fixed category, such as an audio read timeout or an encoding failure, without the original error text. Saved passage names, fingerprints and mix settings are not sent. For stem mixes, we count preparation requests, ready mixes, playback, download requests and failures; track levels, mute and solo settings are not sent. For the BPM counter, we count actions such as starting a tap or audio measurement, finding a reading, an unsuccessful analysis and copying; tap times and BPM values are not sent in these events. For the metronome, we count start requests, successful starts, tempo changes, taps, interruptions and failures; events contain no tempo, beat pattern or volume values. For bass boost, we count selecting a file, preparing or playing a result, requesting a download and failures; audio and EQ settings are not sent. A download request does not tell us whether a file was saved. You can also choose to give fixed-choice result feedback, including whether it helped and one optional improvement category. We do not collect free-text feedback in these counters. Public example playback and clicks back to the tool are counted separately from your own processing results and downloads. These first-party counters do not contain audio, filenames, download links, job credentials, referrers or user identifiers, and do not use analytics cookies or browser storage. Each action is counted at most once per loaded tool page; a refresh starts a new visit. They do not identify unique people or prove a file was saved to disk.

Daily totals expire within 35 days. A keyed network hash is used only for analytics request limits and expires within two minutes. We honor Do Not Track and Global Privacy Control signals for this measurement. Hosting providers may still keep operational logs. Analytics being unavailable does not prevent you from using the audio tool.

Google Analytics

Purchase-click and visible-offer counts require the same analytics consent and respect Do Not Track and Global Privacy Control. Our first-party totals contain only the UTC day, a fixed entry label (such as pricing, results or a free limit), and fixed account/checkout stages such as sign-in attempted, account ready, checkout link ready or checkout failed; they expire within 35 days. They contain no account, order or payment identifiers, email, audio, filenames or checkout links. Marked QA clicks are excluded. Account pages remain excluded from Google Analytics. Separately, essential purchase records establish checkout creation, server-confirmed payments and credited minutes, with QA records kept apart. These operational records do not depend on analytics consent and are not sent to Google Analytics.

We may use Google Analytics 4 to understand visit sources and tool use. It loads only after you choose Allow analytics. Declining leaves the tools available. When enabled, Google receives public page paths, titles, a simplified referring website, browser and device information, and fixed action labels such as starting separation, playing a result or requesting a download, choosing a public purchase entry, or following a public navigation or guide link. Navigation records only a fixed destination and placement, never link text or private account routes. Diagnostic labels include fixed file-format categories, size and duration ranges, elapsed-time ranges, processing stages, and whether the workbench was opened for practice or quick export. They do not include exact file sizes or durations. Known public promotion labels from an explicit allowlist may identify campaign source, medium and name; other query values are discarded. Analytics cookies distinguish visits and sessions and are configured to expire after 60 days. Audio, filenames, temporary download URLs, job credentials, tap timings, BPM values, EQ settings, page query strings and fragments are excluded from these events.

Choose Analytics cookies in the footer to change or withdraw your choice. We store only the choice and its expiry in local browser storage for up to 180 days. Withdrawing disables further measurement and clears our analytics cookies; it does not erase data already received by Google. Do Not Track and Global Privacy Control prevent this measurement. Google signals, advertising storage, advertising personalization and automatic enhanced measurement are disabled. Google may process analytics data outside your country. See how Google uses information from sites that use its services. Analytics consent is separate from the first-party tool-use counters and essential job recovery cookies described in this policy; purchase clicks always require consent.

Microsoft Clarity

After you choose Allow analytics, we may also use Microsoft Clarity to understand how people move through public pages. It loads only after that choice. Declining or withdrawing consent, Do Not Track and Global Privacy Control prevent it from loading. Account pages remain excluded. Clarity may record page views, clicks, scrolling and session replays of public pages, together with browser and device information. It does not receive your audio files, temporary download URLs or job credentials. We do not use it for advertising. Withdrawing consent stops further recording and removes Clarity’s first-party cookies from this site; it does not erase recordings already received by Microsoft. Microsoft may process this data outside your country. See Microsoft Clarity terms and the Microsoft Privacy Statement.

Vercel Web Analytics

After you choose Allow analytics, we also use Vercel Web Analytics for aggregate visits to public pages. It receives public page paths, a referring website when it can be sent without private details, browser and device information, and approximate location. We remove query strings and fragments and exclude account pages, private result routes and the no-upload separator. No audio, filenames, job identifiers or download links are sent. This integration sends page views only.

Vercel Web Analytics does not use analytics cookies. Vercel derives a temporary visitor hash from incoming requests and discards the visitor session after 24 hours. Declining or withdrawing analytics consent, Do Not Track, Global Privacy Control and marked QA sessions prevent these page views. Withdrawing consent stops future measurement; it does not delete aggregates already received by Vercel. See Vercel Web Analytics privacy information.

Capacity measurement

When measurement is enabled, we also count requests refused because of capacity, allowance or request limits, grouped by reason and UTC hour. Upload capacity categories distinguish browser, network, site-wide concurrent uploads and temporary byte capacity. These anonymous totals include repeated attempts; they do not identify people. They expire within 35 days and honor the same privacy signals. Only temporary, two-minute network hashes limit collection traffic. We do not attach filenames, audio, job credentials or browser identifiers to these summaries.

Access and controls

Temporary, unguessable credentials protect your job. Anyone you share a signed download URL with can access the file until that URL expires. An essential, HttpOnly cookie for each tool remembers its latest job for up to 25 hours so refreshing this browser can recover its status. It contains a temporary job credential, not the audio. The stem splitter, bass remover, bass extractor and drum extractor share the same latest four-track job. The BPM counter does not use these cookies. Starting with another file clears that recovery entry; closing the tab does not cancel processing. A separate essential cookie stores a signed, random browser identifier for up to 30 days to apply your shared free allowance across both tools and network changes. It does not grant access to your audio. Analytics cookies are optional and controlled separately. We do not fingerprint your device, use advertising cookies or sell uploaded audio. Contact us about a data request and include your approximate upload time, without sending sensitive audio unless requested.

Optional visitor attribution and PVUV.ai — November 12 supplement

After you allow analytics, we may set the first-party, HttpOnly vocal_visitor_v1 cookie for one year to measure the conversion funnel and first source, and link that visitor to your account when you sign in. It is separate from essential allowance and result cookies. Our visitor records retain the first landing path, referring website, utm_source, utm_medium, utm_campaign, utm_term, utm_content and gclid when present, approximate country from hosting request headers, and a device category. Later visits do not replace the first source. An existing account keeps its first visitor binding across devices. We record fixed funnel actions and associate verified server payment and refund records with that source when available. We exclude marked QA, test and configured operator accounts from business reports. Keyed network and browser-agent hashes help flag unusually frequent page loads; requests without a browser-executed page-load event do not count as visitors. These pseudonymous funnel records are separate from the older aggregate counters and their 35-day expiry described above. They are retained for operating and reconciling this funnel; cookie expiry does not delete server records. Contact us for access, correction or deletion, subject to accounting and legal obligations.

For this attribution update, optional GA4 public page locations may include only the five UTM parameters listed above and gclid. Safe, bounded campaign labels are accepted without limiting them to older campaign names. Other query parameters, fragments, email addresses, audio, filenames and private download or checkout links are excluded. This updates the earlier query-free and fixed-promotion description for attribution only. Account pages remain excluded from Google Analytics. The local-separator page follows the site-wide consent choice for visits and local-unlimited purchase events; no audio or filenames are sent. When configured, consented server funnel events may also be sent to Google using a pseudonymous analytics identifier; our payment ledger remains the source of payment truth.

We may also use the free hosted PVUV.ai service for third-party statistics, source attribution and bot filtering, under the same analytics choice and Do Not Track / Global Privacy Control checks. Its script can collect page URLs and campaign parameters, referrers, visitor/session identifiers, browser, device, screen and language information, approximate location, engagement and performance signals, and automation or rendering signals for traffic-quality scoring. PVUV uses first-party cookies for visitor identity, sessions, first source, a site-supplied opaque account identifier and traffic-quality state. Its visitor cookie lasts approximately thirteen months and may be refreshed; a visitor identifier may also be stored in localStorage without automatic expiry. Its collection service receives network information; its analytics records use keyed IP/network hashes and hashed fingerprint information rather than a dedicated plaintext IP field or raw fingerprint values. This is pseudonymous data, not a guarantee of anonymity. The earlier statement about our own tools not fingerprinting does not describe PVUV’s optional traffic-quality checks. After sign-in we supply an opaque identifier, never an email. Verified server payments and refunds can be reported with transaction event identifiers, USD amounts and currency even without a visitor cookie; they may then be unlinked from a browser journey. Audio, file names, card data and checkout links are not event properties. Optional paid AI reports and advertising protection are not enabled by this integration. PVUV’s hosting providers may process information outside your country, and its server retention depends on service settings. See the PVUV.ai privacy policy.

Declining leaves these new optional browser identifiers and scripts off. Withdrawing through Analytics cookies in the footer stops their further browser reporting and removes our visitor cookie and PVUV identifiers from this browser; it does not erase records already received by us, Google, Microsoft or PVUV, or cancel necessary payment processing. You can also clear cookies and localStorage in browser settings; clearing only cookies may leave PVUV’s local visitor identifier.

Privacy questions

You can reach us through the contact page. Please do not send passwords, card details or private audio in an initial support message.

Contact: support@vocalseparator.com. For copyright reports, use the subject “Copyright report”.

support@vocalseparator.com

Audio tools

No-upload separatorTap tempoMetronomeBass boostAudio practiceVocal isolatorVocal removerAcapella extractorBacking trackStem splitterDrum extractorBass extractor

Product

Pricing & free limitsAccount / My resultsAudio guidesHow it worksFAQ

Support & policies

ContactRefundsPrivacyTermsCopyright & abuse
© 2026 Vocal Separator