🎙️ Vaniq-Edge: Ultra-Low Latency Conversational TTS of 8.5M parameters.
🎧 Audio Samples
Listen to the model's native 24 kHz generation across different speech contexts:
1. Conversational
"Wait, are you actually being for real? I thought the train left twenty minutes ago!"
2. Technical / Academic
"The asynchronous gradient descent algorithm converges rapidly when paired with adaptive learning rates."
3. Narrative / Audiobook
"Night fell over the quiet valley, and as the cold wind whispered through the ancient pines, a distant light flickered in the dark."
Benchmark Metrics
Vaniq-Edge was benchmarked against popular open-source TTS models (standard VITS, Piper, and Supertonic 3) across four key metrics: quality, intelligibility, storage footprint, and computational efficiency. Tests were conducted using standard 50-word English prompts.
🎭 Audio Quality & Intelligibility
This dashboard compares the predicted naturalness (UTMOS22) and pronunciation accuracy (Whisper WER%). Vaniq-Edge achieves top-tier quality while maintaining competitive intelligibility.
⚡ Inference Speed (RTF)
Measured in Real-Time Factor (RTF) — seconds required to generate 1.0s of audio. Lower is faster. Vaniq-Edge runs significantly faster than real-time on a standard CPU and offers lightning-fast performance on GPU.
💾 Model Footprint & Resource Efficiency
Vaniq-Edge is explicitly optimized for constrained environments, offering a tiny disk footprint compared to high-fidelity alternatives, making it ideal for on-device and mobile deployment.
Quickstart
1. Install Dependencies
pip install torch torchaudio phonemizer espeakng-loader==0.2.4 num2words scipy huggingface_hub Unidecode
2. Python API
import sys
import torch
import scipy.io.wavfile as wavfile
from huggingface_hub import snapshot_download
# 1. Download model weights & inference script from Hugging Face
model_path = snapshot_download(repo_id="Abiray/Vaniq-Edge")
if model_path not in sys.path:
sys.path.insert(0, model_path)
# 2. Import VaniqTTS and set device automatically (GPU or CPU)
from inference import VaniqTTS
device = "cuda" if torch.cuda.is_available() else "cpu"
tts = VaniqTTS(model_path, device=device)
# 3. Generate 24kHz Audio
sample_rate, wav_array = tts.synthesize("Hello world! Vaniq-Edge is running locally.")
# 4. Save output
wavfile.write("output.wav", sample_rate, wav_array)
ONNX Runtime
The official verified dynamic export is published directly in this repository. It supports dynamic sequence lengths, CPU/CUDA provider selection, and runs the complete text-to-waveform path without importing PyTorch:
git clone [https://huggingface.co/Abiray/Vaniq-Edge](https://huggingface.co/Abiray/Vaniq-Edge)
cd Vaniq-Edge
python -m pip install -r onnx/requirements.txt
python onnx/inference_onnx.py \
--text "The complete model now runs through ONNX Runtime." \
--output sample-onnx.wav \
--provider cpu \
--speed 1.0