Skip to content
Audio

Canora Dutch

high-quality speech synthesis for the Dutch language

2 min readupdated

Canora Dutch is an open text-to-speech model (TTS (text-to-speech, text-to-speech technology)) for the Dutch language, offering two voices: female (noor) and male (daan). Architecturally, it is based on the lightweight and fast Kokoro-82M base (an 82 million parameter model) and is distributed under the Apache 2.0 license, allowing commercial use. The model correctly pronounces numbers, dates, times, and prices, surpassing many existing alternatives in quality and intelligibility. It is well-suited for developers of voice assistants, audiobook creators, and text-to-speech applications in Dutch.

Key features

  • Two voices: natural-sounding female noor and male daan.
  • Correct reading of numbers, dates, times, and prices according to Dutch language rules.
  • Compact size: 82 million parameters, high operating speed, and quantized int8 ONNX weights of just 115 MB.
  • Standard ecosystem: full compatibility with the kokoro library and support for ONNX Runtime with transformers.js.
  • Permissive Apache 2.0 license for commercial and personal use.

What you can do

  • Integrate the model into bots or services for text-to-speech generation in Dutch.
  • Create tools for the automated generation of audiobooks and educational materials.
  • Build voice output into apps and accessibility tools for visually impaired users.
  • Run the model in the browser or on mobile devices using ONNX Runtime and transformers.js.

Example

import importlib, sys
from pathlib import Path
import soundfile as sf
import torch
from huggingface_hub import hf_hub_download
from kokoro import KModel

REPO = "empiriolabsai/canora-dutch"
sys.path.insert(0, str(Path(hf_hub_download(REPO, "g2p_nl.py")).parent))
g2p = importlib.import_module("g2p_nl").DutchG2P()
model = KModel(repo_id=REPO, config=hf_hub_download(REPO, "config.json"),
               model=hf_hub_download(REPO, "canora-dutch.pth")).eval()
voice = torch.load(hf_hub_download(REPO, "voices/noor.pt"), weights_only=True)

phonemes, _ = g2p("Op 12 maart om 14:30 kost een kaartje € 12,50.")
with torch.no_grad():
    audio = model(phonemes, voice[len(phonemes) - 1], 1.0)
sf.write("speech.wav", audio.numpy(), 24000)
Read moreDetails
Benchmarks on 200 test sentences (WER: word error rate, lower is better; UTMOS: naturalness score from 1 to 5, higher is better): noor voice — 0.93% WER, 4.21 UTMOS; daan voice — 0.98% WER, 4.15 UTMOS.
Hardware requirements: lightweight model (82M parameters), runs faster than real-time on CPU (3.4–3.8x faster than real-time on 4 threads of AMD EPYC).
Limitations: supports only Dutch (words from other languages are read using Dutch rules), fixed speech style without emotions or intonation control.
Training details: synthetic voices created by the authors, not clones of real people.
Examples
Discuss Telegram 0 43

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.3 GB · ready files0.8 GB
  • FP8 / INT8recommendedweights 0.2 GB0.7 GB
  • 4-bitweights 0.1 GB0.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # empiriolabsai/canora-dutch
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download empiriolabsai/canora-dutch --local-dir ~/models/canora-dutch
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to canora-dutch
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. StyleTTS2-LJSpeechfamily root
  2. Kokoro-82M
  3. canora-dutch

New posts about versions of StyleTTS2-LJSpeech (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 3All on Hugging Face
Fine-tunes 12All on Hugging Face

About the model

empiriolabsai/canora-dutch
Author
empiriolabsai
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

430

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Audio

pingo-ptbr

pingo-ptbr