NAC audio tools — Easy demo listening page

Source — 1_rock_87_beat_4-4.wav

DAC EnCodec 32 kHz EnCodec 24 kHz SNAC 44 kHz SNAC 32 kHz Stable Audio Open VAE SAME-L Music2Latent DACVAE ACE-Step 1.5 VAE ACE-Step v1 music DCAE εar-VAE v2 48 kHz εar-VAE 44.1 kHz

DAC RVQ codec·44.1 kHz mono· 4 representations

decodes from codes· details

em.DACEncoder()

Representation	Shape	fps	Decodable
`encoder_z`	[1, 1024, 517]	86.13	✓ converted first
`latents`	[1, 72, 517]	86.13	✓ converted first
`codes` (default)	[1, 9, 517]	86.13	✓
`quantized_z`	[1, 1024, 517]	86.13	✓

DAC representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (mono)

EnCodec 32 kHz RVQ codec·32 kHz mono· 3 representations

decodes from codes· details

em.EnCodecEncoder()

Representation	Shape	fps	Decodable
`encoder_z`	[1, 128, 300]	50.00	✓ converted first
`codes` (default)	[1, 4, 300]	50.00	✓
`quantized_z`	[1, 128, 300]	50.00	✓

EnCodec 32 kHz representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (mono)

EnCodec 24 kHz RVQ codec·24 kHz mono· 3 representations

decodes from codes· details

em.EnCodecEncoder("facebook/encodec_24khz")

Representation	Shape	fps	Decodable
`encoder_z`	[1, 128, 450]	75.00	✓ converted first
`codes` (default)	[1, 32, 450]	75.00	✓
`quantized_z`	[1, 128, 450]	75.00	✓

EnCodec 24 kHz representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (mono)

SNAC 44 kHz multi-scale RVQ codec·44.1 kHz mono· 3 representations

decodes from codes· details

em.SNACEncoder()

Representation	Shape	fps	Decodable
`encoder_z`	[1, 1024, 704]	114.84	✓ converted first
`codes` (default)	[1, 4, 704]	114.84	✓
`quantized_z`	[1, 1024, 704]	114.84	✓

SNAC 44 kHz representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (mono)

SNAC 32 kHz multi-scale RVQ codec·32 kHz mono· 3 representations

decodes from codes· details

em.SNACEncoder("hubertsiuzdak/snac_32khz")

Representation	Shape	fps	Decodable
`encoder_z`	[1, 1024, 512]	83.33	✓ converted first
`codes` (default)	[1, 4, 512]	83.33	✓
`quantized_z`	[1, 1024, 512]	83.33	✓

SNAC 32 kHz representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (mono)

Stable Audio Open VAE VAE·44.1 kHz stereo· 3 representations

decodes from mu· details

em.SAOEncoder()

Representation	Shape	fps	Decodable
`mu` (default)	[1, 64, 130]	21.53	✓
`std`	[1, 64, 130]	21.53	—
`sample`	[1, 64, 130]	21.53	✓

Stable Audio Open VAE representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (stereo)

SAME-L autoencoder·44.1 kHz stereo· 2 representations

decodes from latent· details

em.SAMLEncoder()

Representation	Shape	fps	Decodable
`pre_softnorm`	[1, 256, 65]	10.77	✓ converted first
`latent` (default)	[1, 256, 65]	10.77	✓

SAME-L representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (stereo)

Music2Latent consistency autoencoder·44.1 kHz mono· 2 representations

decodes from latent· details

em.Music2LatentEncoder()

Representation	Shape	fps	Decodable
`features`	[1, 8192, 65]	10.77	—
`latent` (default)	[1, 64, 65]	10.77	✓

Music2Latent representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (mono)

DACVAE VAE·48 kHz mono· 3 representations

decodes from mu· details

em.DACVAEEncoder()

Representation	Shape	fps	Decodable
`mu` (default)	[1, 128, 150]	25.00	✓
`std`	[1, 128, 150]	25.00	—
`sample`	[1, 128, 150]	25.00	✓

DACVAE representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (mono)

ACE-Step 1.5 VAE VAE·48 kHz stereo· 3 representations

decodes from mu· details

em.ACEStep15Encoder()

Representation	Shape	fps	Decodable
`mu` (default)	[1, 64, 150]	25.00	✓
`std`	[1, 64, 150]	25.00	—
`sample`	[1, 64, 150]	25.00	✓

ACE-Step 1.5 VAE representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (stereo)

ACE-Step v1 music DCAE mel autoencoder + vocoder·44.1 kHz stereo· 1 representations

decodes from latent· details

em.ACEStepDCAEEncoder()

Representation	Shape	fps	Decodable
`latent` (default)	[1, 128, 65]	10.77	✓

ACE-Step v1 music DCAE representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (stereo)

εar-VAE v2 48 kHz VAE·48 kHz stereo· 3 representations

decodes from mu· details

em.EARVAEEncoder()

Representation	Shape	fps	Decodable
`mu` (default)	[1, 64, 300]	50.00	✓
`std`	[1, 64, 300]	50.00	—
`sample`	[1, 64, 300]	50.00	✓

εar-VAE v2 48 kHz representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (stereo)

εar-VAE 44.1 kHz VAE·44.1 kHz stereo· 3 representations

decodes from mu· details

em.EARVAEEncoder("ear_vae_44k")

Representation	Shape	fps	Decodable
`mu` (default)	[1, 64, 259]	43.07	✓
`std`	[1, 64, 259]	43.07	—
`sample`	[1, 64, 259]	43.07	✓

εar-VAE 44.1 kHz representations — Top: the decoded audio as a log-frequency spectrogram, mono mix, on the same dB scale for every model. Below it, each representation of batch item 0, dimensions on the y axis, time on the x axis.

Decoded (stereo)

Built 2026-09-29 from this repository's utils/encoder_models.py. Plots are EncodedFeature.plot() heatmaps of batch item 0; every clip is decoded from the representation marked default. Source audio is assets/source.wav, a 6.0-second excerpt of 1_rock_87_beat_4-4.wav from the Groove MIDI Dataset (CC BY 4.0); reconstructions are MP3 to keep the page small.