Zonos TTS Voice Cloning

Run Zonos TTS by Zyphra for voice cloning with emotion and pitch control on Clore.ai GPUs.

Zonos by Zyphra is a 0.4B-parameter open-weight text-to-speech model trained on 200K+ hours of multilingual speech. It performs zero-shot voice cloning from just 2–30 seconds of reference audio and offers fine-grained control over emotion, speaking rate, pitch variation, and audio quality. Output is high-fidelity 44 kHz audio. Two model variants are available: Transformer (best quality) and Hybrid/Mamba (faster inference).

GitHub: Zyphra/Zonos HuggingFace: Zyphra/Zonos-v0.1-transformer License: Apache 2.0

Key Features

• Voice cloning from 2–30 seconds — no fine-tuning required

• 44 kHz high-fidelity output — studio-grade audio quality

• Emotion control — happiness, sadness, anger, fear, surprise, disgust via 8D vector

• Speaking rate & pitch — independent fine-grained control

• Audio prefix inputs — enables whispering and other hard-to-clone behaviors

• Multilingual — English, Japanese, Chinese, French, German

• Two architectures — Transformer (quality) and Hybrid/Mamba (speed, ~2× real-time on RTX 4090)

• Apache 2.0 — free for personal and commercial use

Requirements

Component	Minimum	Recommended
GPU	RTX 3080 10 GB	RTX 4090 24 GB
VRAM	6 GB (Transformer)	10 GB+
RAM	16 GB	32 GB
Disk	10 GB	20 GB
Python	3.10+	3.11
CUDA	11.8+	12.4
System	espeak-ng	—

Clore.ai recommendation: RTX 3090 ($0.30–1.00/day) for comfortable headroom. RTX 4090 ($0.50–2.00/day) for the Hybrid model and fastest inference.

Installation

# Install system dependency
apt-get install -y espeak-ng

# Clone and install
git clone https://github.com/Zyphra/Zonos.git
cd Zonos
pip install -e .

# For the Hybrid model (requires Ampere+ GPU, i.e., RTX 3000 series or newer)
pip install -e ".[compile]"

# Verify
python -c "from zonos.model import Zonos; print('Zonos ready')"

Quick Start

import torch
import torchaudio
from zonos.model import Zonos
from zonos.conditioning import make_cond_dict

# Load model (downloads weights on first run)
model = Zonos.from_pretrained("Zyphra/Zonos-v0.1-transformer", device="cuda")

# Load reference audio for voice cloning
wav, sr = torchaudio.load("reference_speaker.wav")
speaker = model.make_speaker_embedding(wav, sr)

# Build conditioning
cond_dict = make_cond_dict(
    text="Hello from Clore.ai! This is a voice cloning demonstration.",
    speaker=speaker,
    language="en-us",
)
conditioning = model.prepare_conditioning(cond_dict)

# Generate
torch.manual_seed(42)
codes = model.generate(conditioning)
wavs = model.autoencoder.decode(codes).cpu()

torchaudio.save("output.wav", wavs[0], model.autoencoder.sampling_rate)
print(f"Saved output.wav at {model.autoencoder.sampling_rate} Hz")

Usage Examples

Emotion Control

Zonos accepts an 8-dimensional emotion vector: [happiness, sadness, disgust, fear, surprise, anger, other, neutral].

import torch
import torchaudio
from zonos.model import Zonos
from zonos.conditioning import make_cond_dict

model = Zonos.from_pretrained("Zyphra/Zonos-v0.1-transformer", device="cuda")

wav, sr = torchaudio.load("speaker_ref.wav")
speaker = model.make_speaker_embedding(wav, sr)

text = "I can't believe what just happened today!"

emotions = {
    "happy":   [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
    "sad":     [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
    "angry":   [0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0],
    "fearful": [0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0],
    "neutral": [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0],
}

for name, emo_vec in emotions.items():
    cond_dict = make_cond_dict(
        text=text,
        speaker=speaker,
        language="en-us",
        emotion=torch.tensor(emo_vec).unsqueeze(0),
    )
    conditioning = model.prepare_conditioning(cond_dict)
    codes = model.generate(conditioning)
    audio = model.autoencoder.decode(codes).cpu()
    torchaudio.save(f"emotion_{name}.wav", audio[0], model.autoencoder.sampling_rate)
    print(f"Generated: {name}")

Speaking Rate and Pitch Control

import torch
import torchaudio
from zonos.model import Zonos
from zonos.conditioning import make_cond_dict

model = Zonos.from_pretrained("Zyphra/Zonos-v0.1-transformer", device="cuda")

wav, sr = torchaudio.load("speaker_ref.wav")
speaker = model.make_speaker_embedding(wav, sr)

# Slow and calm
cond_slow = make_cond_dict(
    text="Take your time. There is no rush at all.",
    speaker=speaker,
    language="en-us",
    speaking_rate=torch.tensor([8.0]),   # lower = slower
    pitch_std=torch.tensor([20.0]),      # lower = more monotone
)
codes = model.generate(model.prepare_conditioning(cond_slow))
audio = model.autoencoder.decode(codes).cpu()
torchaudio.save("slow_calm.wav", audio[0], model.autoencoder.sampling_rate)

# Fast and energetic
cond_fast = make_cond_dict(
    text="Hurry up! We need to go right now!",
    speaker=speaker,
    language="en-us",
    speaking_rate=torch.tensor([22.0]),  # higher = faster
    pitch_std=torch.tensor([80.0]),      # higher = more expressive
)
codes = model.generate(model.prepare_conditioning(cond_fast))
audio = model.autoencoder.decode(codes).cpu()
torchaudio.save("fast_energetic.wav", audio[0], model.autoencoder.sampling_rate)

Gradio Web Interface

cd Zonos
python gradio_interface.py
# Or with uv:
# uv run gradio_interface.py

Expose port 7860/http in your Clore.ai order and open the http_pub URL to access the UI.

Tips for Clore.ai Users

• Model choice — Transformer for best quality, Hybrid for ~2× faster inference (requires RTX 3000+ GPU)

• Reference audio — 10–30 seconds of clean speech gives best results; shorter clips (2–5s) work but with lower fidelity

• Docker setup — use pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime, add apt-get install -y espeak-ng to startup

• Port mapping — expose 7860/http for Gradio UI, 8000/http for API server

• Seed control — set torch.manual_seed() before generation for reproducible output

• Audio quality parameter — experiment with the audio_quality conditioning field for cleaner output

Troubleshooting

Problem	Solution
espeak-ng not found	Run apt-get install -y espeak-ng (required for phonemization)
CUDA out of memory	Use the Transformer model (smaller than Hybrid); reduce text length per call
Hybrid model fails	Requires Ampere+ GPU (RTX 3000 series or newer) and pip install -e ".[compile]"
Cloned voice sounds off	Use a longer reference clip (15–30s) with clear speech and minimal background noise
Slow generation	Normal for Transformer (~0.5× real-time); Hybrid achieves ~2× real-time on RTX 4090
ModuleNotFoundError: zonos	Ensure you installed from source: cd Zonos && pip install -e .