Wavelength

How an agent makes a song

No DAW and no screen. The agent writes the music as data, Wavelength performs it with real instruments, and the report tells the agent what it can't hear.

1. Write a job

   Notes in beats, sounds by preset name, effects, sends and a master chain, all in one JSON file. Orchestral parts carry their own articulations.

   {
     "tempo": 138,
     "tracks": [
       {"name": "Lead", "plugin": "Altitude", "preset": "LD Trance Spectrum",
        "transpose": 12,
        "notes": [{"beat": 0, "dur": 1, "key": "D5", "vel": 0.9}]},
       {"name": "Violins", "plugin": "BBC Symphony Orchestra",
        "preset": "Violins 1", "articulations": {"long": 0, "spiccato": 1},
        "notes": [{"beat": 0, "dur": 4, "key": "A5", "art": "long"}]}
     ],
     "master": {"loudness": -14, "fx": [{"type": "limiter", "ceiling": -1}]}
   }

2. Render through real plugins

   CLAP and VST3 synths and sample libraries play offline: Vital, Serum 2, Surge XT, Dexed, BBC Symphony Orchestra and about 17,000 presets by name.

   $ wavelength render job.json --out out --json
   rendered 31 tracks in 67 s: out/mix.wav, out/report.json

3. Measure what it can't hear

   Loudness per track and per section, true peak, pitch, brightness and stereo width. The agent balances the mix from the numbers and asks you to listen when it's a matter of taste.

4. Master it

   EQ, multiband compression, limiting and a loudness target such as -14 LUFS for streaming, with a second of silence up front for upload.

   $ wavelength master out/mix.wav --chain mastering.json --lead-in 1
   input    -9.4 LUFS    7.4 dBTP
   output  -14.0 LUFS   -1.0 dBTP