Auditorium — the whole record booth, on your own machine.
Auditorium is a free, Audition-class audio editor for Windows. Record a take, clean it with a vocal chain that derives every setting from the recording itself, split any song into stems, place your cover on the original's instrumental, align lyrics word by word and re-sing just the one that came out wrong — then mix it all on a multitrack timeline with automation, crossfades and sample-accurate editing. Every stage reports what it measured, and everything runs locally: no cloud, no account, nothing leaves the machine.
It's a no-subscription alternative to Adobe Audition with a distinct habit: numbers over promises. Settings are measured, not guessed; tools that can't measure something decline and say why; and the manual is honest about what each feature is worth, because we measured that too.
- Three editors, one cursor — destructive waveform editing, a spectral frequency display, and a multitrack session view with automation, clip fades, crossfades and group editing, sharing selection, zoom and playhead.
- 25 effects + a vocal chain that measures — the chain runs the corrections a rough take needs in one pass, with every threshold derived from the recording itself and a report saying what each stage did or why it declined.
- AI where it earns its keep, always local — stem separation, speech transcription with speaker labels, a voice changer, and word-level lyric alignment run on your CPU in isolated processes. Models download on first use, sha256-pinned; Cancel kills the process outright.
- The cover-chain journey — give it the original song and your vocal take: it separates, cleans, aligns, tone-matches and builds the session unattended, stating its measured caveats before you run it.
- Honesty as a feature — placement accuracy, speaker-separation limits, voice-change quality: measured and printed in the UI and the docs. Where a measurement says a feature shouldn't act, it doesn't.
- Free, MIT, no account — download the installer and start editing.
Just want to use Auditorium? Download the latest
Auditorium Setup X.Y.Z.exe from the
Releases page and run
it — no other software required. The installer adds a desktop shortcut and a
Start Menu entry; a plain-text README.txt ships beside it.
Prefer not to install anything? Grab Auditorium X.Y.Z portable.exe
instead — a single executable that runs from anywhere (Downloads folder, USB
stick) with no installation and no admin rights, and shares the same profile —
downloaded AI models and voice profiles — as the installed app.
Neither download is code-signed yet, so Windows SmartScreen may warn the first
time — choose More info → Run anyway. You can verify any download first
against SHA256SUMS.txt (published with every release):
CertUtil -hashfile "Auditorium Setup X.Y.Z.exe" SHA256 # compare with SHA256SUMS.txt- Windows 10 or 11 (64-bit)
- AI features download their models on first use (166 MB stems · ~323 MB transcription · 32.5 MB speaker separation · 161 MB voice · 378 MB lyric alignment), verified by pinned sha256 before every load.
Requires Node.js 20.19+ (or 22.12+) and Git, on Windows x64:
git clone https://github.com/Redrum624/Auditorium.git
cd auditorium
npm install
npm run dev # Vite dev server + Electron
npm run build:win # -> release/: Auditorium Setup <version>.exe + Auditorium <version> portable.exe
# (+ README.txt and SHA256SUMS.txt covering all three)The spectral frequency display — a log-axis spectrogram rendered off the main thread, sharing the waveform's cursor, selection and playhead.
The multitrack editor — clips with fades and a crossfade across the tracks, and a volume envelope open on the top lane, all at the same cursor the other two editors share.
The pipeline tools open beside the waveform, not over it. A multi-stage pass — Cover Chain, Vocal Chain, transcription — steps through its stages in the module column while you keep selecting audio and moving the playhead. Every stage row carries the note saying why it sits where it sits, and afterwards, what it measured.
 |
 |
| Vocal Chain, after a run — each stage reports the settings it derived, its measured delta, or the measurement that made it decline; the pass closes with a before/after table and lands as one undo entry. | Cover Chain, before you run it — the measured caveats are stated before anything runs, and every stage of the six-stage journey is written out on the card. |
 |
 |
| Align Lyrics — paste the words you know are in the recording; click a word to hear exactly it, and re-record just the one that came out wrong. | The Remix panel — every splice of a re-arrangement, inspectable, pinnable, re-rollable and undoable. |
The effects, from rack to card. Every registered effect lives in a categorized rack one click from the waveform; one click opens it as a card in the module column, and every card previews against the real document before Apply.
 |
 |
| The Effects rack — 25 effects in ten categories; the Pipeline tools have their own module. | Parametric EQ — five bands plus high- and low-pass, previewed on the real audio before Apply. |
 |
 |
| Reverb — room size, damping, mix and pre-delay; Preview plays the processed document, Apply is one undo entry. | The Spatial positioner — azimuth, elevation and distance as honest stereo projection, writing automation keys on release. |
The side panels — the module column's cards, always the same width, never over the audio.
 |
 |
| Properties — the selected clip's geometry and gain, with both fade editors and the crossfade's own controls. | Markers — named positions that persist sample-accurately in every container and survive destructive edits. |
- Waveform Editor — per-sample amplitude view with zoom, scroll, selection, cursor and playhead.
- Spectral Frequency Display — off-main-thread spectrogram, log or linear axis, HiDPI-rendered.
- Multitrack Editor — tracks and clips with volume/pan/mute/solo/arm, non-destructive fades and crossfades, multi-clip selection with cross-track group drag, ripple delete, merge selected clips into one (silence in the gaps), selectable gaps (double-click the empty stretch between two clips and
Delcloses it, moving only that track's later clips), per-track automation envelopes, and one-step-per-gesture session undo. - Spatial Panel — places a track's sound around the listener (azimuth/elevation/distance, honest stereo projection — not binaural), writing automation keys on release.
- Recorder — input-device selection, channel/rate choice, live level meter; multitrack punch-in on armed tracks.
- Effects Rack — categorized effects, each opening on one click as a card in the module column; the Pipeline tools have their own module.
- Pipeline Module & Menu — the twelve long-running tools in three groups (Tempo & Timing, Voice, Analysis); each opens as a wide card in the module column.
- Files / History / Markers / Properties Panels — open documents, browsable undo history, the marker list, and read-only facts about the active document or clip.
- Module Strip — the icon strip over the module column; the bar and the open module are always the same width.
- Edit Toolbar — floating Split/Merge/Copy/Paste/Delete, Trim/Silence, Undo/Redo pill, greyed by the same rules as the menu (Split at Cursor is
Ctrl+K; Merge Clips has no key and lights only in the Multitrack view; Cut keepsCtrl+Xin the Edit menu). - Transport & Level Meters — play/stop/record/loop, view toggle, zoom cluster, time readout and output meters, centred on the waveform rather than the window. Play starts at the position line, always — Pause moves the line to where playback stopped, so
Space·Spacestill resumes; the one exception is a selection the line sits outside, which plays from the selection's start. - Tempo Readout & Beat Grid — detected BPM with confidence everywhere it matters, and the tracked beats drawn under all three editors (dimmed and dashed when stale rather than presented as fact).
- Match Tempo Tool — one ratio or a per-beat tempo map from the confirmed grid; measured on accelerandi: worst interior beat 526 ms → 4.6 ms at 0.83 BPM/s drift.
- Align Vocal Timing Tool — warps syllables onto the grid between anchors you confirm; defaults to 25 % strength because a fully quantised vocal sounds machine-made.
- Align Lyrics Tool — places every word of known lyrics against the recording (median 20 ms word-start accuracy, measured); click a word to hear it, re-record just that word, and it splices in level- and pitch-matched, length-preserved.
- Vocal Chain Tool — eleven correction stages in a measured order, each reporting its derived settings, its delta, or the measurement that made it decline. The noise gate decides where, not how loud — see Features.
- Cover Chain Tool — the six-stage journey from original song + your take to a finished two-track session, with the caveats stated above the Run button.
- Podcast Chain Tool — ten speech-tuned stages ending in a delivery target: DC removal, noise reduction, de-hum, shortened pauses, gate, compression, de-essing, EQ, loudness to −16 LUFS (stereo) or −19 LUFS (mono) by ITU-R BS.1770-4, and a limiter at −1.0 dBFS sample peak. One undo entry; documents with more than two channels are refused rather than mis-measured.
- Auto-Remix Tool & Remix Panel — re-arranges a track's own bars to a target length; every splice inspectable, rejectable, pinnable (up to four pins guaranteed to survive every re-plan) and undoable.
- Separate Voice Tool — the same separation landed as two tracks instead of five:
— Voiceand— Backing(drums + bass + everything else + residual). The two add back up to the source within float32 rounding (measured worst error 4.32e-7) — the five-stem landing is the bit-exact one. When more than one person is talking it goes one step further and lands one track per speaker plus Backing: it counts the voices, shows you the count with each speaker's speech time before anything lands, and lets you overrule it. On the four test recordings (three with two speakers, one with four) the count was right every time, both from clean speech and through the separation the tool runs first — measured indocs/bench/diarize-bench-baseline.json. Speaker tracks carry short fades at each turn and share any overlap, so they do not add back sample for sample; Backing still does. - Separate into Stems Tool — Drums/Bass/Vocals/Other + Residual as five documents and a session that sums back to the source bit-exactly (measured: worst error 0).
- Transcribe Tool & Transcript Panel — timestamped speech-to-text with speaker labels, coloured timeline regions, instant speaker-count regrouping, SRT/WebVTT export.
- Voice Changer Tool — re-timbres a recording toward a saved voice profile; requires an explicit rights affirmation before any reference clip is used.
Editing & formats
- Sample-accurate copy/paste/trim/silence, a Delete and a
Ctrl+Xcut that keep the length (the span is zero-filled, markers stay put;Shift+Delripples), and Split at Cursor (Ctrl+K): every marker is a segment boundary and a double-click selects the segment under the pointer — with per-document undo (50 steps / 800 MB budget) and full session undo in the multitrack, where Split cuts every clip under the cursor on the selected tracks in one step. Merge Clips is Split's inverse there: on every track with two or more clips selected they become one clip spanning from the first start to the last end, silence in the gaps, over a newMerge Nfile that has the members' gain and fades rendered into it — one undo step for the whole gesture, no shortcut, Multitrack only. - Open WAV, MP3, FLAC and OGG (Opus) with native-rate import across container variants (including surround WAV with selectable ITU-R downmix); Save writes the project (
.audm, atomic) in every view and never overwrites an audio file — audio leaves through Export. - Export WAV 16/24/32-float, FLAC, MP3 CBR, OGG Opus — the active document in the editors, the session mixdown (muted tracks out, automation and fades in) in the multitrack.
- Markers persist sample-accurately in every container (WAV cue, ID3 chapters, FLAC/OGG chapter tags) and survive destructive edits by remapping.
- Snapping ("the magnet"): cursor, selections, clip drags and trims quantise to clip edges, beats, markers and the session cursor — placed geometry outranks derived; Alt suspends.
- The draggable red cursor handle rides all three views, snap-aware, transport-neutral.
- Zoom anchors on the position line, on every path and both surfaces —
Ctrl+wheel and the toolbar's−/+alike, in the waveform, spectral and multitrack views. The line keeps its place on screen; when it is off screen the new view is centred on it.Shift+wheel scrolls and Fit are unchanged. - Projects save as
.audm(binary v4, v3 still opens) carrying every open document plus the session's clips, fades, crossfades, automation and spatial lanes; Mix Down renders bit-identically to live playback.
The vocal chain, in one paragraph
Eleven stages in a measured order — DC offset, noise reduction, de-hum, optional silence removal, the gate, pitch correction, compression, de-essing, high-pass, limiting, optional reverb. Every setting is derived from the take: the compressor's threshold is the level the take exceeds half the time it is sounding, the noise print is learned from the quietest half-second of real material (never digital silence), the high-pass sits an octave under the lowest sung note. The gate asks WHERE, not how loud: aligned lyrics or a transcript protect word spans absolutely; every half-second is measured for vocal-tract resonance so whispers, breaths and held consonants survive; anything voiced is kept — and pause noise louder than your softest singing still goes, which no level threshold could ever do. When nothing qualifies it declines with the numbers and offers the manual threshold instead. Measured end to end on a real 142-second take: noise floor −61.3 → −67.4 dBFS, median pitch deviation 23.3 → 14.7 cents, programme level within 0.2 dB, length unchanged.
The cover chain, honestly
Separation's five stems sum back to the mix exactly, so the instrumental is the original minus its vocal to the last bit — but it still contains the original singer (measured 17.95 dB below the music), and the tool says so before you run it. Alignment is a placement, not a warp — cross-correlated against the separated vocal (deliberately not the song: refining against the mix was built, measured costing 3.3–12.9 ms, and withdrawn), recovered to within 10 ms at normal levels, and when the evidence is weak it still places the tracks at the measured lag with every rival lag one click away. Matching is EQ, loudness and a limiter measured against the separated original vocal; Match Reverb usually declines, having measured that the app's shortest decay exceeds what the record shows. The level trim after mixdown takes back its own overshoot on both faders equally — one undo entry.
Analysis & AI, with the measurements attached
- Tempo detection — tracked beats (not extrapolated), confidence reported, ×2/÷2 octave re-track.
- Stem separation — bit-exact recombination guaranteed; separation quality bounded by the model, and the UI says so. CPU at ~1.5× realtime.
- Speaker separation (Separate Voice) — the right count on 4 of 4 test recordings, in both conditions measured (clean speech and through the stem separation the tool runs first); ~162 s of material, count-only truth, no diarization error rate, near-zero overlap in it — the tool prints that limit beside the count and hands you the control.
- Transcription — 100 % speaker accuracy on clean two-voice material, 45–73 % at three voices, overlap not detected: printed in the tool, and the speaker count is a control you can change instantly.
- Voice changer — measured mean cosine 0.795 toward the target (vs 0.615 source) across five real voices; a change, not a forensic clone, and the tool says which.
- Lyric alignment — median 20 ms word starts, 88 % within 100 ms on sung material; a mismatch between lyrics and audio raises a warning, never a refusal.
- A per-phone pronunciation scorer was built, measured at AUC 0.642, and cut — your ear picks the word; the tool makes it reachable.
Quality of life
- A launch splash that reports the stages startup genuinely reached (measured: zero added latency).
- A crash card that names the error selectably instead of a frozen window; background failures raise a dismissible notice that blocks nothing.
- Keyboard shortcuts throughout — the full table in
KEYBOARD_SHORTCUTS.md.
- Desktop: Electron 43 —
contextIsolation,sandbox,nodeIntegration: false, typed IPC whitelist, fail-closed write-path policy; the renderer never touches the filesystem. - Frontend: React 19 + TypeScript (strict) + Vite 7 + Zustand; waveform, spectrogram and multitrack draw to canvas.
- DSP: pure synchronous TypeScript — every effect is a
process()overFloat32Arraychannels; heavy work runs in Web Workers. - Inference:
onnxruntime-node(CPU) inside one ElectronutilityProcessper feature — killable on Cancel, never loaded by the renderer. No GPU path: measured, DirectML exhausted 15.7 of 16 GB VRAM without finishing one segment, while the CPU runs stems at ~1.5× and transcription at ~9× realtime.
vocalChain.ts // the measured eleven-stage chain + its report
coverJourney.ts // the six-stage cover pipeline
stemPartition.ts // ratio masks + time-domain residual = exact recombination
MultitrackPlayer.ts // live engine, scheduled against one clock epoch
mixdown.ts // the offline ground truth the live engine must match
alignLyricsService // CTC forced alignment + the word-splice
npm run dev # dev server + Electron
npm test # jest — 3 projects, 6000+ tests
npm run typecheck # tsc --noEmit
npm run smoke # packaged end-to-end smoke (800+ assertions)
npm run navigate # packaged UI walker (50 surfaces)- docs/USER_GUIDE.md — the full walkthrough
- docs/KNOWN_LIMITATIONS.md — where Auditorium deliberately differs from Audition, with the measurements
- KEYBOARD_SHORTCUTS.md — every binding
- CHANGELOG.md — release notes, root causes included
- docs/bench/README.md — the measurement discipline behind the numbers
- THIRD_PARTY_NOTICES.md — every bundled component, model and ported algorithm
Issues and PRs welcome — see CONTRIBUTING.md. The short
version: open an issue first, one logical change per PR, keep npm test and
npm run typecheck green, and behavior changes carry tests that failed before
the change.
- Stem separation — HT-Demucs by Meta AI (MIT), via the
StemSplitio/htdemucs-onnxexport (MIT). - Speech recognition — Whisper base by OpenAI (Apache-2.0), via
onnx-community/whisper-base. - Speaker embeddings — CAM++ by the WeSpeaker project (Apache-2.0), from the sherpa-onnx model release.
- Speaker segmentation (Separate Voice) — pyannote-segmentation-3.0 by Hervé
Bredin / CNRS (MIT), via the ONNX conversion in
csukuangfj/sherpa-onnx-pyannote-segmentation-3-0— the upstream HuggingFace repository is gated; this mirror is not. - Speaker embeddings for Separate Voice — WeSpeaker ResNet34-LM (VoxCeleb, Apache-2.0), from the same sherpa-onnx model release.
- Voice conversion — OpenVoice V2 by MyShell.ai (MIT), via
Hinotsuba/OpenVoice-ONNX-v2(MIT). - Forced alignment — wav2vec2-base-960h by Meta AI (Apache-2.0 upstream); the ONNX
graph rides the
onnx-communitymirror, which declares no licence of its own — the derivation is stated rather than glossed, and the sha256 pins are load-bearing. - MP3 encoding — lamejs (LGPL-3.0),
statically bundled — full licence text and relink note in
THIRD_PARTY_NOTICES.md.
Every model is downloaded from its source on first use — nothing is rehosted, nothing is bundled.
Sampled daily from the GitHub Releases API — the curve builds from launch day forward.
PolyForm Noncommercial 1.0.0 © 2026 Redrum624 — free for any noncommercial purpose; commercial use requires a separate licence.
Auditorium — a free, Audition-class audio editor for Windows.