Zonos v0.1 Transformer — GGUF
GGUF conversion of Zyphra/Zonos-v0.1-transformer for use with CrispASR and CrisperWeaver.
Model
Zonos v0.1 is a ~500M-parameter text-to-speech model trained on 200k+ hours of multilingual speech. It delivers expressive, high-quality synthesis with fine-grained control over:
- Speaker cloning from a few seconds of reference audio
- Emotion control (happiness, sadness, disgust, fear, surprise, anger, neutral)
- Speaking rate (phonemes/second)
- Pitch variation (pitch standard deviation)
- Audio quality (fmax parameter)
Output is native 44.1 kHz mono PCM via a 9-codebook DAC decoder.
Architecture
- Text input → eSpeak phonemization → phoneme embeddings
- Prefix conditioner: concatenates conditioning embeddings (phonemes, speaker, emotion, fmax, pitch_std, rate, language) through projection + LayerNorm
- AR backbone: 26-layer GPT-style transformer (d=2048, GQA 16q/4kv, SwiGLU, LayerNorm with bias, RoPE) generates 9 interleaved DAC codebook streams with delay pattern
- Classifier-free guidance (CFG): both conditioned and unconditioned prefixes are decoded; logits are interpolated with configurable scale (default 2.0)
- DAC decoder: separate GGUF converts 9-codebook RVQ codes → 44.1 kHz PCM
Files
| File | Size | Description |
|---|---|---|
zonos-v0.1-transformer-f16.gguf | 3.1 GB | Full-precision (F16) — reference quality |
zonos-v0.1-transformer-q4_k.gguf | 872 MB | Q4_K quantised — ~3.5x smaller, good quality |
Required companion
Zonos needs the DAC 44.1 kHz decoder GGUF for audio output. Use dac-44khz.gguf from cstr/dia-1.6b-GGUF (same codec, shared between Zonos and Dia backends).
Usage
CrispASR CLI
# Download
crispasr -m auto --backend zonos
# Synthesize
crispasr --backend zonos \
-m zonos-v0.1-transformer-q4_k.gguf \
--tts-codec dac-44khz.gguf \
--tts "Hello, this is a test of the Zonos text to speech system." \
-o output.wav
# With emotion control
crispasr --backend zonos \
-m zonos-v0.1-transformer-q4_k.gguf \
--tts-codec dac-44khz.gguf \
--tts "I'm so happy to see you!" \
--emotion "0.8,0,0,0,0,0,0,0.2" \
-o happy.wav
# With voice cloning
crispasr --backend zonos \
-m zonos-v0.1-transformer-f16.gguf \
--tts-codec dac-44khz.gguf \
--voice reference.wav \
--tts "Cloned voice output." \
-o cloned.wav
CrisperWeaver (Flutter GUI)
Download from Model Management, select on the Synthesize screen, and set dac-44khz.gguf as the codec companion.
C API
#include "crispasr.h"
// Open session
crispasr_session* s = crispasr_session_open("zonos-v0.1-transformer-q4_k.gguf", "zonos");
crispasr_session_set_codec_path(s, "dac-44khz.gguf");
// Configure (optional)
crispasr_session_set_temperature(s, 0.7);
// Synthesize
int n_samples;
float* pcm = crispasr_session_synthesize(s, "Hello world", &n_samples);
// ... write pcm to WAV ...
free(pcm);
crispasr_session_free(s);
Conditioning Parameters
| Parameter | Range | Default | Description |
|---|---|---|---|
pitch_std | 0-400 | 45 | Pitch variation. 20-45 = normal, 60-150 = expressive |
speaking_rate | 0-40 | 15 | Phonemes per second. 10 = slow, 15 = normal, 30 = fast |
fmax | 0-24000 | 24000 | Maximum frequency. 22050 for voice cloning |
cfg_scale | 0-10 | 2.0 | Classifier-free guidance scale |
emotion | 8 floats | neutral | [happiness, sadness, disgust, fear, surprise, anger, other, neutral] |
temperature | 0-2 | 0.0 | Sampling temperature (0 = greedy) |
Conversion
Converted from the upstream PyTorch checkpoint using:
python models/convert-zonos-to-gguf.py \
--input Zyphra/Zonos-v0.1-transformer \
--output zonos-v0.1-transformer-f16.gguf
Quantised with:
crispasr-quantize zonos-v0.1-transformer-f16.gguf zonos-v0.1-transformer-q4_k.gguf q4_k
License
Apache 2.0 (same as Zyphra/Zonos-v0.1-transformer).
Links
- Upstream model: Zyphra/Zonos-v0.1-transformer
- Blog post: zyphra.com/post/beta-release-of-zonos-v0-1
- Engine: CrispASR
- App: CrisperWeaver