Music Tech3 min read

Neural Audio Synthesis, Explained

How neural networks generate sound directly — from autoregressive waveform models to neural codecs and DDSP, and why each approach sounds the way it does.

Generating audio is harder than generating images, and the approaches taken to make it tractable explain everything about how AI-generated music sounds.

Why it's hard

An image is a fixed grid of pixels. Three minutes of stereo audio at 44.1kHz is over 15 million samples, each depending on those around it, with structure at every timescale from microseconds (timbre) to minutes (song form).

Generating that directly, sample by sample, is computationally brutal. Every practical approach is a strategy for avoiding it.

The approaches

Autoregressive waveform models

Predict each sample from the previous ones. WaveNet was the landmark example.

Quality: excellent — it models audio directly with no compression loss. Problem: generating 44,100 sequential predictions per second is extremely slow. Impractical for long-form music without heavy optimisation.

Neural codecs plus sequence models

The dominant approach. A neural codec compresses audio into a few hundred discrete tokens per second. A transformer generates token sequences. The codec's decoder turns them back into audio.

Quality: good, bounded by the codec. The compression is why generated music often has a slightly glassy quality in cymbals and sibilance. Advantage: hundreds of tokens per second instead of 44,100 — tractable for full songs.

Diffusion on spectrograms

Generate a spectrogram by iteratively denoising, then convert to audio with a vocoder.

Quality: excellent texture. Problem: the spectrogram-to-audio step must reconstruct phase information, which is lossy. And global refinement handles long-range structure poorly.

DDSP

Rather than generating audio, the network predicts parameters for traditional synthesis — harmonic oscillator amplitudes, filtered noise levels, envelope shapes. Real DSP does the sound generation.

Quality: very clean, because the output is synthesised by well-understood algorithms rather than reconstructed. Advantage: tiny models, small training sets, real-time capable, and inherently controllable. Limitation: works best on harmonic, monophonic sounds. Less suited to dense polyphonic mixes.

DDSP is what powers most good timbre transfer.

Why this shows up in what you hear

SoundCause
Glassy cymbals, metallic sibilanceNeural codec compression
Blurry stereo imageCodec or vocoder phase reconstruction
Beautiful texture, wandering structureDiffusion without a structure model
Clean, expressive single instrumentDDSP

The practical point

All of these produce rendered audio. However sophisticated the synthesis, the output is a fixed file.

That's why the more useful question for a musician isn't which synthesis approach a tool uses — it's whether the tool gives you something editable afterwards: MIDI you can change, or stems you can rework.

Related reading: how AI generates music, timbre transfer explained, and diffusion vs transformer music models.

Frequently asked questions

What is neural audio synthesis?

Generating sound using neural networks rather than traditional oscillators and filters. The network learns the statistical structure of real audio from training data and produces new audio matching those patterns.

Why is generating raw audio so hard?

Because audio has 44,100 samples per second per channel, and each one depends on those before it. Generating three minutes of stereo audio means producing over 15 million correlated values, which is far harder than generating an image.

What is DDSP?

Differentiable Digital Signal Processing combines neural networks with traditional synthesis. The network predicts parameters for classic oscillators and filters rather than generating raw samples, which gives cleaner results with much less training data.

Start making music in Veena

Free, browser-based, no downloads required.

Try Veena Free