TopLine is vocal mixing software that learns from the decisions that made finished records: what the engineer did to the raw voice, measured step by step, turned into numbers a machine can choose. The first prototype is built. It decoded a working engineer's mix, then mixed the raw vocal from the decode alone. You can hear its first song on this page.
The name is the product. In a song, the topline is the lead vocal, the part everybody actually hears. TopLine is the engineer for that one part: vocals only, end to end, directed in plain words.
Honest stage report: the first prototype is built and working. It has done one full job end to end: decode an engineer's mix into measured decisions, then render the raw vocal through those decisions, with the measurements published next to the audio. There are no users yet, nothing is for sale, and there is no revenue. The company is new; the problem is not: years of the founder's own catalog sit behind it, and the approach has been in development for over a year. Nothing on this page claims otherwise.
This is Confide, from Razi's album Angels & Demons (2023), in three stages. The raw vocal chains he recorded at home in December 2022. The same vocal after his engineer worked on it, printed without the beat. And the finished record that was released. Listen to the distance between the first stage and the last. That distance is the gap this project exists to close, and every step of it is human work. Then hear the prototype take its first pass at closing it: stage four is in the prototype section below.
Confide as Razi recorded it in December 2022: three vocal chains, a top chain, a bottom chain, and a double-layer harmony, stacked and played together at unity. His own recording, in his own room. Nothing has been added and nothing has been cleaned up. This is what an engineer receives.
The same performance after Andrew Morrell mixed the vocal, printed from his session without the beat and returned in February 2025. Every decision in the craft section below lives in the difference between this file and the first one: the tuning, the EQ, the compression, the space.
The finished record: the mixed vocal seated in the beat, mixed and mastered, and released on Angels & Demons on February 14, 2023. This is the standard a bedroom vocal has to reach before anyone calls it a record.
The gap, measured: the distance between stage one and stage two was measured off these files, not guessed. Pitch: the raw take sits a median of 22.2 cents off the note grid, and the mixed vocal sits 8.4 cents off. Air: the mix lifts the top end to about +10 dB at 16 kHz. Space: a dark reverb that lives in the midrange and keeps ringing after the voice stops. Judgment leaves numbers behind. Learning to choose those numbers is the whole job.
What this is not: none of these three stages is TopLine output. This is the human work TopLine has to learn: a take, the engineer's decisions, and the record they became. TopLine's own first render is stage four, in the prototype section: the same raw chains, run through the chain the measurements produced. Nothing on this page carries a label it has not earned.
The plan is to close that gap in one session. The route today is the one this project exists to replace: months of waiting on one engineer, who also holds the only copies of the project files, and a result that can still come back needing human work. The TopLine route aims at the same reference target every time, reached in one sitting, with revisions directed in plain words.
The test is already designed. The same raw vocal and the same beat go to a working genre engineer and to the prototype, loudness matched before judging, and working producers and engineers score both mixes blind. The bar is win or tie on preference. The protocol and the results get published either way, and the full protocol is Step 04 of the plan.
Every AI mixing tool on the market learned from finished records. TopLine learns from the decisions that made them. The Lab is where those decisions get recovered: a song's vocal chain measured off the files themselves, written down as parameters, and filed as training data. This page is the working record. It grows one decode at a time, and it states only what the files support.
The song at the top of this page, taken apart. Razi's raw vocal chains from December 2022, measured against the mixed acapella his engineer Andrew Morrell returned in February 2025: the same performance, before the work and after it. The three players above are the evidence. Below is the chain the measurements read out of those files, step by step, in the order engineers usually run them. A print cannot prove the order of the middle steps, and each step below is labeled for what it is: measured, estimated, or unconfirmed.
The bottom falls away under the voice: about 14 dB down at 25 Hz, 9 dB down at 40 Hz, back to level by 80 to 125 Hz.
A high-pass filter around 60 to 80 Hz, moderately steep. Measured.
The median distance from the note grid goes from 22.2 cents in the raw take to 8.4 cents in the mixed vocal. Frames landing within 10 cents of the grid rise from 21.7% to 56.9%.
Real correction at moderate strength. Notes are pulled to the grid, not snapped to it. Measured. Whether it was manual work or an automatic tuner, the print cannot say.
Sibilant sounds hold essentially flat from raw to mixed, while the same 5 to 10 kHz band rises about 8 dB on everything else.
A de-esser or dynamic EQ holding sibilance in place while a large top-end lift was applied around it. Estimated, from a stated method.
The gain moves 2 to 4 dB on typical peaks, with the deepest momentary cuts near 6 dB. The phrase-to-phrase loudness spread is roughly halved, from 2.66 dB to 1.35 dB.
One compressor or two stages, plus light rides or a slow leveler. The envelope is measured; the settings are estimated.
About +3 dB across the mids. A presence peak of +5 to +6 dB centered near 2 to 3 kHz. An air shelf still rising at the top, reaching about +10 dB at 16 kHz.
Measured. The top octave reading is partly bounded by what the raw recording itself captured up there.
After the EQ's own contribution is subtracted, a small lift remains in the third and fourth harmonics, about +2 to +3 dB.
Consistent with light saturation or an exciter. Pitch processing can leave a similar trace, so the cause stays unassigned. Estimated, low to medium confidence.
In the break, where the raw stems are digital silence, the mix keeps ringing: about 29 dB of decay across roughly 4.5 seconds. 83% of the tail's energy sits between 200 Hz and 1 kHz.
A dark, low-mid weighted stereo reverb in a long decay class. The decay is measured; the exact time is an estimate, and the patch and pre-delay are unknown.
No tempo-locked repeat stands out in the tail at the song's 112 BPM subdivisions.
Not measurable from the print. The artist confirmed by ear that a slight delay is in the engineer's chain, so presence is artist-confirmed. The time and the level stay estimates, and they are labeled as estimates wherever the chain below uses them.
The raw chains are bit-exact mono. In the mix, the width lives between 500 Hz and 2 kHz, where the reverb returns sit; the dry voice and the top octaves stay mono. The vocal print sits at -16.4 LUFS with no brickwall limiting on it.
Width from the returns, not from a widener on the voice. Measured.
What this print cannot settle: every plugin identity, the exact settings, the reverb patch, whether the tuning was manual or automatic, and the mix bus and mastering chain, because the beat file is not in the archive. That list is part of the decode. A decode that hides its limits would be worthless as training data.
The Lab sits on the founder's own archive: years of sends to one engineer and the work that came back. These are the counts, verified against the files on disk in November 2026. Nothing is rounded up, and nothing enters a count until its decode is written and filed.
The zero is the honest center of this dataset. The decisions are recoverable from the sound; the first decode proves that much. The settings are the ground truth, and the settings live in session files the engineer holds. So the dataset is built in two layers. The seeds teach the mapping from a raw vocal to a finished one, with every inferred line labeled as inference. New pairs get commissioned with the session file named as a deliverable, so the next ground truth arrives with its settings documented.
Confide is decoded end to end and written into the dataset format: measurements first, reconstructions second, the unsettled list attached. It is the template every later decode follows.
Every finding in every decode carries one of four labels: measured, engineer-documented, estimated, unknown. The audio is evidence. The parameters are the dataset, and the format is fixed now so nothing has to be re-decoded later to fit it.
Dreamland, Mellow High, Stalkin, Illmatic, and an untitled song at 90 BPM in G minor get the same pass Confide got. Each one is checked against the files on disk before its record is written.
Decoding widens to the released records the genre runs on, measured from the released print alone. Those decodes set the targets: what a finished vocal measures like. Without the raw side, the chain behind a released record stays labeled inference, never recovered fact.
The same raw vocal and the same beat go to a working genre engineer and to the prototype, loudness matched, judged blind by working producers and engineers. The bar is win or tie, and the results are published either way. The full protocol is Step 04 of the plan.
What the Lab is: the dataset being built in the open, one decode at a time. It has no users and it sells nothing. When a decode lands, this page is where it shows up first. The first prototype running on Decode No. 1 is the next section down.
The prototype's first job was Confide. It took the same three raw vocal chains from stage one and ran them through the chain that Decode No. 1 measured out of the engineer's print, in code, start to finish. No settings were copied from a session file, because no session file exists. The decode was the only recipe it had. The render is below, with what the chain did and how close the measurements land.
The raw performance from stage one, processed by the decoded chain and printed at the same loudness as the engineer's acapella, so the comparison runs on sound rather than volume. This is the vocal on its own: no beat file survives in the archive, so there is no full mix to render yet. This is version 1.5 of the render, November 2026, and it is not a final version: the chain keeps improving, and this exhibit updates with it. Its render notes, including the parts it still gets wrong, are part of the project record.
The bottom falls away under the voice before anything else touches it: a second-order filter at 70 Hz, shaped against the decode's low-end curve.
Fit check: the chain sits 8.6 dB down at 42 Hz against raw. The decode measured the engineer at 9.1 dB down at 40 Hz.
Peaks get caught at the weight the print's envelope says they were caught: 3 to 1, soft knee, fast attack, a release that lets phrases breathe.
Measured gain reduction: up to 5.6 dB on the peaks. The decode read the engineer at 2 to 4 dB typical, with peaks near 6 dB.
A slow gain ride works phrase to phrase, aimed at halving the loudness spread the raw take carries.
The render's phrase spread lands at 1.29 dB. The engineer's print: 1.28 dB. The raw stack: 2.85 dB.
The decode's measured curve, rebuilt filter by filter: body in the mids, the presence peak, and the air shelf still rising at the top.
About +2.5 dB at 300 and 900 Hz, +5.5 dB at 2.5 kHz, and a +9 dB shelf at 7.5 kHz, from the decode's readings of +1.7 to +4.1 dB in the mids and about +10 dB of air at 16 kHz.
Runs after the EQ, the position the print's behavior points to: sibilance held in place while the top-end lift is applied around it.
Split-band, driven by the sibilant band's share of the signal. Measured against the print, the render's sibilant band sits about 3 dB under his at 5.3 kHz and about 2 dB under at 8.5 kHz.
A light harmonic layer for the lift that remains in the print after the EQ's own contribution is subtracted.
The chain's most tentative step. The decode marks that finding estimated, at low to medium confidence, and the chain treats it that way.
A generated dark stereo tail, band-shaped so its energy lives in the low mids, the way the print's tail does in the break.
Tail check: the render's tail enters at -40.7 dB, the engineer's at -42.2 dB. Its width measures -12.5 dB side-to-mid; his measures -13.6 dB.
A short repeat tucked under the voice. The artist's ear caught it before any measurement could: an estimated eighth note at the song's tempo, most of it routed into the reverb, so it reads as part of the space rather than as an echo.
Estimated, and labeled that way: 268 ms, return 17.5 dB under the dry vocal. Presence is artist-confirmed. The print itself never gives the setting up.
One gain stage to the engineer's own print loudness. No limiter on it, like his print.
Printed at -16.40 LUFS. His acapella measures -16.35 LUFS.
The tuning, measured: in version 1.5, pitch correction runs per stem. Each vocal chain is corrected on its own before the stems are summed, the way the engineer worked in his session. The corrected vocal lands at a median of 8.7 cents off the note grid. The engineer's print measures 8.9 cents; the raw take measured 21.9. Stem by stem: the top stem lands at 11.2 cents, the bottom stem at 8.2 cents, and the double-layer harmony at 10.5 cents. Across the whole vocal, 54.7% of frames sit within 10 cents of the note; his print measures 54.2%. Version 1.4 pushed the correction harder and measured tighter, 6.1 cents at the median, and the founder's ear sent it back: the tuning sounded scratchy, and the vibrato had been ironed down to 3.8 cents of depth. His print holds 6.3. Version 1.5 backs the correction off to his numbers and leaves the performance's vibrato standing at 11.0 cents, nearly the raw take's own depth. The founder's verdict on each version is recorded with the render notes.
Same yardstick as the decode. Before any comparison was trusted, the verification pass re-measured both prints across the full files. That is why the figures here sit a hair off the decode's: the decode measured matched sections and read the raw take at 22.2 cents and the engineer's print at 8.4. Measured whole, they read 21.9 and 8.9. Both prints sit at -16.4 LUFS, so loudness is not doing any quiet persuading.
What moved in version 1.5, measured: the tuning was backed off to the engineer's numbers, and the fullness work from 1.4 held. The render's density lands at 77 to 92 percent of the engineer's print at the scales measured, where version 1.3 sat near 70, and the harmony layer still sits where he keeps it, well under the leads. The tuning's fast passages trail again, as they did in version 1.3: 80.8% of frames within 25 cents of the note, against 86.1% in the engineer's print. That gap is the price of the gentler correction, and it is named rather than tuned away. What does not match yet, measured and named: his print carries a transient ceiling this chain has not reproduced: crest factor 19.35 in his print, 15.60 in the render. The gentler tuning did not buy the crest back, because the fullness is density added under a fixed peak. His stereo width also moves over time in a way the render's does not yet: the width varies with a spread of 3.56 dB in his print against 1.89 in the render, and the mechanism behind his movement has not been identified from the print. And his sibilance control behaves like dynamic EQ, which this chain's split-band de-esser does not replicate: the render's upper harmonics still sit 2 to 3 dB lighter than his. Each miss is a number, and each number is what the next pass of the chain works against.
The prototype above ran one song through one decoded chain, assembled by hand from one decode. The product is that loop, learned: a system that reads a raw vocal, chooses the chain for it, and revises in plain words. Three pieces have to exist for that. Each one is already started.
Songs where the raw vocal and the engineer's finished mix of the same take both survive. The founder's catalog seeds this layer: six songs, 16 stem files, verified on disk. New pairs get commissioned with the session file named as a deliverable, so their settings arrive documented instead of estimated.
Released records decoded from the finished print alone: what a record's vocal measures like, genre by genre. This layer teaches targets and vocabulary. It is never described as ground truth, because the raw side of those songs sits in label vaults.
The system learns to set the numbers of an editable processing chain, the same chain shape the Lab writes down, chosen fresh for the vocal in front of it. Because the output is a chain of named settings, every render can be inspected, and plain words from the artist map onto real controls.
Before any wider build, the same raw vocal and the same beat go to a working genre engineer and to TopLine, loudness matched, judged blind by working producers and engineers. The bar is win or tie. The results publish either way. Full protocol: Step 04 of the plan.
Nothing makes it sound professional and Hollywood. It always comes back muddy, or messed up, or okay. And okay still needs human work.
The founder, on every tool he has tried
The song is written and recorded, then it sits. The founder's engineer has held songs for nine months to a year at a stretch, and gone silent for weeks at a time. The record moves when he moves. Until then, nothing moves.
Mixing is priced per song, and professional engineers charge hundreds of dollars for a single one. The founder's engineer sat at the cheap end of that market, and the cheap end still cost him a year of waiting. Price was never the only problem.
The artist becomes the project manager of his own bottleneck. Checking in. Following up. Asking where his songs are. Spending the relationship, message by message, to get his own record back.
The engineer holds the only copies of the project files. And the quality of the finished record rides on one person's week: his health, his mood, his other clients. When he stops, every song he holds stops with him.
Hear something wrong and the fix is another round of messages, another wait, and a little more goodwill spent. Artists learn to stop asking. They release records they can still hear problems in.
The tools that exist make a file louder or cleaner overall. The vocal itself still sounds like a bedroom: muddy, buried, or polished in the wrong places. Close still means hiring a human to finish it, which saves nothing.
Underneath all of it sits one fact. The judgment that makes a vocal sound professional has never been written down. It lives in engineers' heads, built up one reference record at a time, and nobody has turned it into a system. It has never been written down because the evidence for it was never collected: the pairs that would teach it, a raw vocal beside the finished mix of that same vocal and the settings that got from one to the other, exist only inside private studio sessions that share nothing. TopLine exists to write it down.
Mastering is math: loudness and frequencies you can measure. Vocal mixing is taste. Hundreds of small judgment calls. How dry. How upfront. How much tuning before the character dies. That judgment is what engineers charge for, and nobody has automated it.
Two more reasons the lane stayed open. The training data this job needs, raw vocals paired with the finished mixes of those same vocals, sits locked inside studios that share nothing. And the big audio companies built for professionals, never for bedroom artists. The tech to close the gap only just arrived.
A professional vocal mix is a chain of decisions, and every link is a judgment call. This is the whole chain, in the order an engineer runs it on a commercial hip hop or R&B record. Read what the machine has to learn, stage by stage.
The engineer inherits whatever the room and the microphone actually did. The first decision is how much of the job is repair. A bedroom take needs the repair half of the craft before the taste half can start.
One lead gets built from many takes, phrase by phrase, sometimes syllable by syllable. The call is continuity: a rougher take that holds the feeling beats a cleaner take that breaks it.
Phrases get nudged toward the groove, edits get faded clean, and breaths get lowered instead of deleted. How tight is a genre call. Breaths kept low read close. Breaths removed read edited.
How much correction, at what speed, on which notes, before the character dies. In hip hop and R&B the amount is an identity decision before it is a repair decision.
Clicks, plosives, hum, room, noise. The professional discipline is restraint: over-cleaning strips consonants, breath detail, and life along with the damage.
Mud and boxiness come out, harsh resonances get hunted down, and the voice gets shaped against a target. The reference record supplies that target, section by section.
Stages of compression with different jobs: one catches peaks and hard consonants, one levels phrases smooth. Too much, and the performance goes flat and the room comes up with it.
Only the sibilant energy gets compressed, by an amount set per voice and per microphone. Bright and intelligible on one side of the line, dull on the other.
Harmonics added as an insert or blended underneath on a parallel path. Done right, a vocal reads louder and closer on small speakers. Done heavy, it reads fuzzy and harsh.
Every line's level against the beat, written section by section: hooks larger than verses, quiet lines kept intimate, every lyric intelligible. Engineering on the surface, arrangement underneath.
Plate, hall, and tempo-synced delays on separate returns, ducked under the dry lead. At low level it is glue. A delay throw on the last word of a hook is a production event, and somebody decides where that line sits.
Doubles are second performances, never copies, aligned tight and tuned tighter than the lead. Ad-libs get placed in the gaps and treated harder. Verses stay single so hooks have somewhere to lift.
The vocal gets seated against a beat that was finished before the vocal existed, usually a two-track the artist bought or leased. Where the beat masks the words, it dips, dynamically and by a little. There is a ceiling here and it gets named: no system can rebalance instruments inside a finished stereo file. The honest claim is the best seat a vocal can get, and stems, when the artist has them, raise the ceiling.
Final loudness, true peak, and translation checks on earbuds, phone speakers, and cars. Loudness itself is a genre posture. This is also the one stage machines already solved, years ago, which is why mastering tools are everywhere and vocal judgment is nowhere.
Pitch in cents. Onsets in milliseconds. Resonance peaks, sibilance energy, loudness, true peak. Every one of these can be extracted from audio today. This part of the craft is close to solved, and TopLine treats it that way.
How bright a current hip hop lead sits. How wet a hook runs. How tight the grid gets. Nobody writes these down either, but they repeat across records, so they can be learned from enough of them, with the reference record as the target.
The emotion inside a comp. The size of a hook. How much of the singer's character correction should keep. This kind belongs to the artist. The product asks about it and learns it from that artist's revisions. It never gets hard coded.
TopLine is designed around that split. The first two kinds get automated. The third kind gets a conversation: the direction interface below, where the artist says what they hear and the system does the engineering underneath.
Artists already do the creative half of this job every day. You listen, you hear what is wrong, you say it. TopLine is built so that saying it is the whole interface.
Upload the raw vocal and the beat. TopLine is built first for the bedroom studio: a mic, a laptop, a treated-enough room. That is the setup independent artists already record in, and the exact point where their records stall.
Give it a reference: a released song whose vocal sound you want. TopLine aims at real records, matched section by section, not at a preset's guess of one.
More upfront. Drier. Less tuning, keep the character. Direction, not settings. TopLine translates taste into the hundreds of small calls underneath.
The record comes back, you listen, you answer, it revises. The revision loop is the product. It ends when you say the record is done.
Tap the notes you would give an engineer. This is a working sketch of the product's core interface, the direction language TopLine is built around. It renders no audio and changes no file. It shows the artist's whole job inside TopLine: listen, and say what you hear.
No notes yet. Give it direction.
The first artists TopLine serves record at home: a dry vocal take, a beat, and one reference record they want to sound like. Hip hop and R&B vocals carry the heaviest conventions and the most reference-driven sound in music, and the market is small enough to own. One market, proven properly, then the next.
A subscription for artists who need finished records every week, plus TopLine plugins of our own for the engineers and producers who want the engine inside their own sessions.
The hardest input there is: phone mic, room echo, compression baked in. The day a voice memo comes back as a record, every phone on earth is a studio.
TopLine gets built in this order, in the open. Every step names who owns it. This page gets updated as each one lands.
This site goes live, and the project runs in the open from here: the problem, the test bed, and every step below, published as they land. Razi owns this step. He built the site, and he answers for the project.
Razi preps the test bed himself: his own catalog, more than 100 songs, as raw vocal and beat pairs. The build team takes shape alongside it: an audio ML engineer leads the technical build, with contract DSP and ML engineers carrying the prototype in the meantime.
The first working chain already exists: it decoded an engineer's mix of Confide and rendered the song from the decode, and it is at the top of this page. This step turns that chain into the genre engine. The engine takes hip hop and R&B vocals only, and learns the judgment calls: tuning with taste, balance against the beat section by section, doubles and ad-libs placed as arrangement, space used on purpose. The audio ML engineer builds it, with contract support. Razi directs every pass against his own catalog and judges what comes back.
This is the project's proof, and it is designed to be believed. The same raw vocal and the same beat go to both sides. The human side is a working engineer with credits in the genre, paid his normal rate, working in his normal tools, given the same reference record the system gets. The machine side is the prototype with the reference loaded and no hand-tuning. Both mixes are loudness-matched before anyone hears them, because louder wins lazy comparisons and every professional in the room knows it. Working producers and engineers judge blind, order randomized, and answer three questions on each pair: which do you prefer, which do you believe is the human mix, and would you release yours. Win or tie on preference is the bar. The protocol, the audio, and the results are published either way. A test published only when it wins would be marketing. A test published on a fixed protocol is evidence. Razi runs the test.
Inside the first genre, the bar gets harder: the system has to match the vocal sound of two or three named commercial reference records, section by section, on real songs from the test bed. Razi judges every pass against the records he pointed at.
Ten to twenty independent artists, recruited personally from the founder's own community, mix real songs they actually release. One working mix engineer or producer joins as an advisor. Their feedback trains the revision loop.
Genre by genre outward, and the recording source gets rougher: phone vocals, room echo and all. The end state is a voice memo coming back as a record.
Three doors, one engine. Per-song pricing for the artist who releases a few tracks a year, priced at a clear fraction of the human engineer, who runs from about 25 dollars a song at the budget end to several hundred a song at the professional end. Subscriptions for the prolific artists, camps, and indie labels who mix every week and carry the biggest engineer bills. And TopLine plugins, the same vocal judgment packaged for the engineers and producers who do mix and want the engine as their starting point.
Razi Ismail is a recording artist from Wayne, New Jersey, in the New York metro. He started in his teens. Lyricist first. Then a vocalist, self-taught, who learned to sing by doing it badly until it was not bad anymore. He has written and recorded more than 100 of his own songs, and most of them are still waiting on the step this company exists to replace. His engineer has held songs of his for nine months to a year at a time, and he holds the only copies of the project files.
He does not produce. He directs AI production. His working days run the same way this product is designed to run: directing AI systems to a quality standard, listening hard, and iterating until the output is right. In the last year, solo, with no investors and no employees, he has shipped five live products: a daily news product, a social gaming platform, shift coverage software a real restaurant operator runs his business on, a complete web game, and a storefront business. A creator marketplace is in build now.
He left college in 2022 and has been building ever since. TopLine is the company his own catalog has been asking for, and his catalog is the test bed it gets proven on. The archive behind that test bed has been recovered and verified file by file: six of his songs now survive complete, raw vocal chains beside the engineer's finished work, and 27 titles carry an engineer mix or master on disk. The three files at the top of this page are his at every stage: his voice in the raw take, his voice in the mix, his voice on the released record. Nobody else building in this lane has a catalog of 100-plus songs of their own, recorded and waiting.
He is the user, and he is the channel. TopLine's first distribution is Razi's own artist channels and his own catalog: his releases are the proof, and the artists who follow his work are the first door. The first outside users are the 10 to 20 artists in the test group, recruited by him personally.
“The world's changing and AI vocals are taking over. This is a form of humans competing whilst aiding the AI revolution.”
Razi, on why this company, in his own words. Part of the labels' power has always been that professional sound costs money and access. Remove that cost, and artists stop renting a finished sound and start owning their output.
The prototype works on the founder's own catalog today. The first outside artists come in through a test group: ten to twenty artists mixing real songs they actually release, with their feedback training the revision loop. The waitlist is the door to that group.
One email when the test group opens, and one if your slot comes up. Nothing else. The list is never sold.
TopLine is in development, and the work runs in the open. If you record at home and want to hear when the first test group opens, join the waitlist above or write to the founder. Artists in that group will mix real songs they actually release, and their feedback shapes the revision loop.
If you mix records for a living, the blind test needs judges: working engineers and producers willing to hear two mixes of the same vocal and say which one is human. That is product testing. Your ears help decide whether the prototype passes.
officialrazibusiness@gmail.com