Modelpublic

cagyirey/ZONOS2-GGUF

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

2likes45downloads

Model Card

ZONOS2-GGUF

GGUF builds of Zyphra/ZONOS2 — an ~8B-parameter (SonicMoE, ~900M active) real-time text-to-speech model with voice cloning — for mistral.rs.

Files

File	Size	Use
`ZONOS2-Q8_0.gguf`	8.27 GB	Recommended. Near-lossless and the smallest coherent quantization.
`ZONOS2-F16.gguf`	15.3 GB	Full precision, for maximum fidelity / reference.

Quantization: Q80 is the floor for this architecture — smaller quants (Q6K and below) are incoherent, because ZONOS2's SonicMoE residual structure amplifies low-bit error. Use Q8_0 or F16.

Usage (mistral.rs)

Serve the model:

bash

mistralrs-server -p 8080 speech --arch zonos2 --model-id ZONOS2-Q8_0.gguf

Generate speech via the OpenAI-compatible /v1/audio/speech endpoint. For voice cloning, pass a speaker embedding:

bash

curl localhost:8080/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{
        "model": "default",
        "input": "Your text here.",
        "speaker_embedding": [ /* 2048-dim speaker vector */ ],
        "response_format": "wav"
      }' --output out.wav

—speaker_embedding — the [2048]-dim speaker vector (e.g. from an ECAPA speaker encoder). Omit it for unconditional generation.
—Tip: punctuate naturally — short clauses help prosody and clean utterance termination.

Model

—Architecture: ZONOS2 (Zyphra) — SonicMoE decoder + DAC neural codec.
—Output: 16-bit mono PCM, 44.1 kHz.
—Voice cloning: via a per-request speaker embedding.

Provenance

Converted from the upstream Zyphra/ZONOS2 checkpoint with mistralrs-modeltool zonos2-gguf.