thuongvv/MOSS-Audio-Tokenizer-Nano-GGUF

MOSS Audio Tokenizer Nano — GGUF for ZeroTTS Swift

GGUF conversions of OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano for native, on-device waveform decoding with ZeroTTS-swift.

These are converted and selectively quantized weights, not a newly trained model. They are intended for the ggml MOSS decoder shipped by ZeroTTS Swift and are not a standalone Hugging Face Transformers model.

Files

File	Size	Recommendation
moss-audio-tokenizer-nano-f32.gguf	87,872,192 bytes (83.80 MiB)	Reference-quality baseline
moss-audio-tokenizer-nano-f16.gguf	44,307,136 bytes (42.25 MiB)	Recommended for iOS
moss-audio-tokenizer-nano-q8_0.gguf	24,259,008 bytes (23.14 MiB)	Smaller; verify perceived audio quality

F32 is the unquantized reference build. F16 is the default recommendation because it measured 49.93 dB SNR against F32 on an eight-frame deterministic random-code fixture. Q8_0 measured 24.36 dB on the same fixture. These figures measure numerical drift, not perceived speech quality.

Codec format

• Sample rate: 48,000 Hz
• Native output: stereo, interleaved Float32 PCM
• Residual codebooks: 16
• Expected input layout: frame-major, 16 codec codes per frame
• Container: GGUF

The quantization process keeps codebooks, normalization weights, scales and biases in F32. Only compatible matrix-multiplication weights are reduced.

ZeroTTS Swift usage

Download one GGUF variant and pass its local file URL as codecModelURL:

import ZeroTTS

let voice = try ZeroTTSVoice(contentsOf: voiceURL)
let synthesizer = try ZeroTTSSynthesizer(
    modelURL: generationModelURL,
    codecModelURL: mossCodecModelURL,
    threads: 4
)

let pcm = try await synthesizer.synthesize(
    "Xin chào Việt Nam.",
    voice: voice
)

For production applications, download from a release tag or immutable commit instead of main, and verify the SHA-256 value in manifest.json or checksums.sha256 before loading the model.

Example URL:

https://huggingface.co/OWNER/MOSS-Audio-Tokenizer-Nano-GGUF/resolve/v1.0/moss-audio-tokenizer-nano-f16.gguf

Conversion provenance

• Base checkpoint: OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano
• Runtime/conversion implementation: thuongvovan/ZeroTTS-swift
• Runtime backend: ggml/GGUF only; ONNX Runtime is not required
• Packaged: 2026-09-15

Conversion and quantization scripts are available under Native/MOSS/scripts/ in the ZeroTTS Swift repository.

License

The base checkpoint declares the Apache License 2.0. This repository preserves that license for the converted weights. See LICENSE and the upstream model card for details.