ZAC-1 (Zyphra Audio Codec 1)

The 44.1 kHz speech codec of ZONOS3-Mini: 15 Hz frames of 10 codebooks of 4096 entries. Codebook 0 is semantic (distilled from a frozen Wav2Vec2-BERT 2.0 teacher), codebooks 1-9 are an acoustic residual RVQ, and a causal multi-resolution iSTFT decoder synthesizes the waveform. Under similar data constraints, a TTS model trained with ZAC-1 achieves a 10x reduction in word error rate compared with one trained with DAC, while achieving better acoustic similarity. Compared with S1-DAC and MOSS v2, it achieves comparable results at better compression (ZONOS3-Mini post).

Code, usage, streaming and the C++/ggml CLI: github.com/Zyphra/ZAC-1.

Files

File	Size	Use
zac-1-decoder-keep2_plain.safetensors	0.58 GB	PyTorch decoder, default: pruned stages (decoder.0 keeps 2 of 32 layers, stages 2/4/6/8 keep 2 of 12), about 3× faster
zac-1-decoder-full_mr.safetensors	4.07 GB	PyTorch decoder, full-size
zac-1-encoder.safetensors	4.81 GB	PyTorch encoder (CUDA): acoustic encoder, semantic branch, quantizers (fp32) and the Wav2Vec2-BERT 2.0 teacher (bf16)
gguf/zac-1-decoder-keep2_plain-f16.gguf	0.45 GB	ggml decoder (zac-cli), keep2_plain
gguf/zac-1-decoder-full_mr-f16.gguf	2.20 GB	ggml decoder (zac-cli), full_mr
gguf/zac-1-encoder-f32.gguf	4.03 GB	ggml encoder (zac-encoder-cli)
gguf/zac-1-ssl-teacher-f32.gguf	1.55 GB	ggml semantic teacher (zac-encoder-cli)

sha256 of the safetensors files: zac-1-encoder.safetensors 89716a1f928a4de051577e664d316508ab7eac11decb5510d1bf284801fe49ee, zac-1-decoder-keep2_plain.safetensors 851ee0748dbcbf0b257944d8380c5c412418fea1628c8b2e142ae600c3c7ef40, zac-1-decoder-full_mr.safetensors 91a6d98e94e7f347606decb8444082ceec0f1fc2d971c1cef25893be6a949583.

The weights are those of ZONOS3-Mini's codec (autoencoder/ of Zyphra/ZONOS3-Mini: codec step 600,000, decoder heads step 50,000), repackaged; the GGUFs' tensors are byte-identical to ZONOS3-Mini's.

Usage

import soundfile as sf
import torch

decoder = torch.hub.load("Zyphra/ZAC-1", "zac_decoder", trust_repo=True)  # or variant="full_mr"
audio = decoder.decode(codes) * decoder.output_gain()                     # codes [frames, 10] -> [frames * 2940]
sf.write("decoded.wav", audio.numpy(), 44100)

encoder = torch.hub.load("Zyphra/ZAC-1", "zac_encoder", trust_repo=True)  # CUDA only
wav, sr = sf.read("speech.wav", dtype="float32", always_2d=True)
codes = encoder.encode(torch.from_numpy(wav.T), sr)                        # int32 [ceil(N / 2940), 10]

License

Apache-2.0. zac-1-encoder.safetensors (teacher.*) and gguf/zac-1-ssl-teacher-f32.gguf contain the first 16 layers of facebook/w2v-bert-2.0 (MIT).

Citation

@misc{zyphra2026zonos3mini,
  title        = {{ZONOS3-Mini Technical Report}},
  author       = {Gabriel Clark and Sofian Mejjoute and George Close and Beren Millidge},
  year         = {2026},
  month        = nov,
  howpublished = {\url{https://www.zyphra.com/our-work/zonos3-mini}},
}