Paste a YouTube URL or upload an audio file. We'll separate instruments, remove noise, correct EQ for the recording era, and master the output for modern sound systems.
49 restored recordings from the golden age of tango
We restore 1920s–1940s Argentine tango recordings for use by tango DJs (musicalizadores) at milongas. The goal is to take noisy, degraded vintage recordings and make them sound clear, warm, and ready for modern PA systems — while preserving the musical character of the original era.
Each recording goes through a multi-stage restoration process:
Every restored track is evaluated across 7 dimensions of audio quality. The scoring system models the noise-distortion tradeoff: Clarity and Transients (top of radar chart) reward noise removal, while Fidelity and Naturalness (bottom) penalize over-processing. This creates a natural sweet spot where the score peaks — removing too much is penalized just like removing too little.
| Dimension | Weight | What It Measures |
|---|---|---|
| Clarity | 20% | How clean the audio is — absence of hiss, crackle, and background noise. Bell-curved: penalizes both noise AND over-cleaning. |
| Transients | 20% | Instrument attack clarity — whether you can hear the bandoneon staccato, violin phrasing, and piano accents that dancers respond to. |
| Consistency | 5% | Volume stability over time — steady loudness makes it easier for DJs to mix tracks in a tanda without riding the fader. |
| Fidelity | 15% | Timbre preservation — how faithfully the restored version preserves the original instrument colors and tonal character. |
| Naturalness | 15% | Preservation of natural instrument texture — detects when denoising strips the aperiodic energy (breath, bow noise) that gives instruments their life. |
| Warmth | 15% | Low-mid frequency body — the tonal richness from strings, bandoneon, and bass that gives tango its emotional weight. |
| Balance | 10% | Overall tonal quality — not muffled (too dark) and not harsh (too bright). Tuned for the natural spectral character of 1920s–40s recordings. |
Advanced users can fine-tune 17 individual parameters across five categories — source detection, noise reduction (with 5 fine-tuning controls), EQ & tone, harmonic enhancement, and dynamics & output — for complete control over the restoration process.
Start with a preset, listen, then adjust. Here are the most impactful parameters and when to use them:
| Problem | What to Adjust |
|---|---|
| Too much background noise remains | Increase Denoise strength, add more Denoise Passes, or switch Noise Profile to Auto |
| Metallic or tinkly artifacts | Increase Gain Smooth to 3–5, or switch Noise Estimation to Conservative |
| Sound is muffled or "in a cave" | Switch Noise Estimation to Average, or reduce Denoise Passes |
| Instruments sound thin or lifeless | Reduce EQ Intensity to 50–80%, increase Exciter to 1.5–2.0, or set Harmonic Restore to High |
| Too loud / distorted at peaks | Lower Loudness to −18 or −20 LUFS, reduce Compression |
| Clicks and pops remain | Increase Declick sensitivity |
| Stereo sounds unnatural | Reduce Stereo Width to 80% or lower |
| Parameter | What It Does | Typical Range |
|---|---|---|
| Noise Reduction | ||
| Denoise | Overall noise reduction intensity | Auto (recommended) or 0.2–0.8 |
| Denoise Passes | Number of denoising iterations. More passes = cleaner but risks stripping instrument texture. Auto adapts per stem based on noise level | Auto or 2–3 |
| Noise Profile | Auto measures each recording's noise. Shellac/Vinyl use generic profiles | Auto for most tracks |
| Noise Estimation | How the noise profile is computed. Average is strongest, Conservative protects instruments, Adaptive balances both | Average for orchestras, Conservative for solos |
| Declick | Removes clicks and pops from scratched records | Auto or 0.3–0.7 |
| Fine-tune Denoising (click to expand) | ||
| Gain Smooth | Smooths denoising across frequencies. Reduces metallic artifacts at the cost of some clarity. All presets now include some smoothing by default | 1–3 (default), 5+ for artifact-prone tracks |
| Residual Floor | How much noise to leave behind. Lower = cleaner, higher = more natural | −42 to −34 dB |
| Noise Floor | Threshold for what counts as noise. Lower = more sensitive | −50 to −40 dB |
| Band Width | Precision of noise band analysis. Lower = more targeted, higher = smoother | 0.8–2.0 |
| Adaptivity | How fast denoising reacts to changes. Lower = faster response | 0.3–0.7 |
| EQ & Tone | ||
| EQ Intensity | How much spectral correction to apply. 0% = bypass, 100% = full correction | 50–120% |
| Per-Stem EQ | Applies separate EQ to each instrument layer. Auto tries both and picks the best | Auto |
| Harmonic Enhancement | ||
| Harmonic Restore | Pitch-tracked restoration of missing overtones. Off in Light, Low in Standard (very noisy only), High in Full (most noisy sources) | Off / Low / Medium / High |
| Exciter | Broadband brightness via nonlinear harmonic generation. Runs after harmonic restoration | 0 (off) – 2.0 |
| Dynamics & Output | ||
| Compression | Reduces volume range for consistent playback. Higher = more even | 1.3:1 to 2.0:1 |
| Loudness | Target output level in LUFS (broadcast standard) | −18 to −14 LUFS |
| Stereo Width | Spatial widening for mono recordings. 100% = standard, higher = wider | 80–120% |
For questions, feedback, or feature requests, please reach out via email or GitHub.
If you believe content processed through this service infringes your copyright, please contact us with details of the original work and the specific content in question. We will respond promptly.
This is a non-commercial research project for restoring historical tango recordings. The pipeline uses open-source tools (Demucs, ffmpeg) to separate instruments, remove noise, and correct spectral balance for recordings from the 1920s–1950s.
About 20–25 minutes for a new track (AI stem separation takes most of the time on our CPU-only server). Reprocessing the same track with different settings takes about 5 minutes because the stems are cached. Jobs are queued and processed one at a time.
Yes. Once a track is processed, the AI stem separation result is cached. Reprocessing the same track with different presets or settings takes only about 5 minutes instead of 25. Use the “Reprocess” button on a completed job to try different settings quickly.
Light removes noise and corrects EQ while preserving the original character — no compression, stereo, or harmonic enhancement. Standard adds gentle compression, stereo widening, and exciter for dance floor readiness. Full applies maximum processing including harmonic restoration for very noisy sources. Start with Standard.
Some noise on 78rpm recordings is physically inseparable from the music — it's groove modulation noise that follows the signal's amplitude. Our pipeline removes stationary hiss, clicks, crackle, and hum, but signal-correlated noise is a fundamental limitation of the medium, not a processing failure.
You can paste a YouTube URL or upload MP3, WAV, FLAC, M4A, OGG, or OPUS files (max 50 MB, 15 minutes). Output is 320kbps MP3.
Processed files are automatically deleted after 3 days. Download your restored tracks promptly.
The denoising and stem separation work on any historical recording (jazz, blues, classical). The spectral EQ correction is tuned for tango's frequency balance, so results on other genres may have slightly different tonal character. Try it — it often works well.
The 7-dimension Playback Readiness score measures Clarity (noise level), Transients (attack clarity), Consistency (volume stability), Fidelity (timbre preservation), Naturalness (texture preservation), Warmth (low-mid body), and Balance (tonal quality). On the radar chart, noise removal metrics (top) are opposite preservation metrics (bottom). Higher is better, but trust your ears over scores.