Are you ready to shape the future of open-source voice technology? As part of the upcoming Dutch Speech Tech Day 2027, we are thrilled to launch an exciting Hackathon focused on creating high-quality, efficient text-to-speech (TTS) for the Dutch language.
Whether you are an AI researcher, developer, audio enthusiast, or hobbyist, this is your chance to build a voice model that makes a real-world impact.
The Challenge Your mission is to train a Dutch TTS model capable of fast, accurate inference directly on a CPU.
Baseline Model: Piper TTS
Core Dataset: A curated subset of the CML-TTS Dutch Dataset
Customization: You can supplement the core dataset with up to 10 GB of additional data—including public datasets, custom recordings, or synthetic audio. Note: Any additional data used must be made publicly available.
Why Participate?
Open to Everyone: Anyone, anywhere in the world, can join the challenge and submit a model.
National Recognition: Submissions from Dutch residents/organizations will be eligible for the Audience Choice Award during the Dutch Speech Tech Day!
Community Impact: Help advance open and reproducible AI for the Dutch language.
Student Advantage: Student participants/teams will get one exclusive Q&A opportunity with an industry speech tech expert during the hackathon to seek advice, troubleshoot, or refine their approach. Feel free to reach out to the organizers for the support via email!
Submission & Evaluation Participants must submit audio samples generated by their system, the model checkpoints, and a brief system overview.
Registration: Registration is not required; submitting your work before the deadline automatically counts as participation.
Submission Deadline: January 18, 23:59 (CET), 2027
Submit all links and the required team information through Submission form.
Evaluation: Audio samples will undergo a listening test to evaluate naturalness and quality, alongside processing speed.
Reproducibility Commitment: To foster an open community, all code, checkpoints, and added training datasets must be publicly released by Dutch Speech Tech Day 2027.
The Dutch Speech Challenge invites participants to develop high-quality Dutch text-to-speech (TTS) systems that produce natural, clear speech and run efficiently on a CPU. Starting from the shared baseline and core dataset, participants will complete the following three tasks:
(Optional) Data Augmentation & Curation: Expand and refine Dutch speech training data through approaches such as custom recordings, web scraping, voice conversion, or synthetic data generation. Select and clean the data to prioritize audio clarity, accurate text–audio alignment, and balanced phonetic coverage. A carefully curated dataset can be more valuable than a larger, noisier one.
Model Training & Code Management: Train or fine-tune the TTS model using the augmented and curated dataset. Organize data preparation scripts, training pipelines, and dataset links in a structured repository so that others can reproduce the work.
System Reporting & Open Sharing: Submit a maximum one-page A4 report describing the dataset strategy, cleaning methods, augmentation choices, and training insights. Commit to publicly releasing the source code and augmented datasets following Dutch Speech Tech Day 2027 to support the broader speech research community.
Participation is open worldwide, individually or in teams. Submissions from Dutch residents or organizations are eligible for the Audience Choice Award.
2. Training data
The organizers will provide the core dataset. Participants may supplement the core dataset with up to 10 GB of additional data, including public datasets, custom recordings, and synthetic audio.
The following rules apply:
All additional training data must be documented and made publicly available by Dutch Speech Tech Day 2027.
Participants must have permission to use and redistribute their additional data. Public availability alone does not establish permission to redistribute it.
For new recordings, participants must obtain permission from the speakers for the intended training use and public release.
Synthetic data must be identified as such, with the generating model and generation procedure documented.
Data cleaning, filtering, resampling, transcription correction, and augmentation are permitted. Participants must describe these steps.
Piper is the shared baseline. Participants must use the organizer-provided implementation and recipe as their starting point.
Proposed rule: changes to preprocessing, training settings, and the Piper architecture are permitted, provided they are fully documented and the final system remains reproducible and capable of CPU-only inference, without requiring a GPU, an internet connection, or an external inference service. Replacing Piper with an unrelated TTS system is outside the main challenge.
External pretrained TTS models, vocoders, speaker encoders, and other learned components may only be used under the published pretrained-model policy. Any permitted use must be disclosed, including use solely to generate synthetic training data.
Participants must report their training hardware, training duration, approximate GPU-hours, and main training settings.
All eligible submissions will be evaluated using the same held-out Dutch texts and a standardized inference procedure. The evaluation will assess three main dimensions:
Naturalness
A blinded listening test will assess how natural the speech sounds, including rhythm, intonation, stress, phrasing, and pauses.
Audio quality
A separate listening rating will assess clarity and audible defects, such as distortion, clicks, noise, metallic artifacts, or unstable volume.
Naturalness and audio quality will be reported separately. Intelligibility may also be assessed using human review and an automatic speech-recognition-based word error rate. Automatic metrics will be treated as supporting evidence and will not replace listening evaluations.
For the system developed by a Netherlands-based team, an Audience Choice Award may be determined through a separate listening vote during Dutch Speech Tech Day. Its results will be distinguished from the standardized evaluation.
Each submission must include the following:
Model files — hosted on Hugging Face
The final trained checkpoint.
(Optional) a CPU-ready inference model, such as an ONNX export.
Required configuration files and supporting assets, including speaker mappings where applicable.
A model card describing the system, training data, intended use, limitations, and license.
Source code — hosted on GitHub
Training, preprocessing, augmentation or export (if needed), and inference code.
Installation instructions and a dependency file with pinned versions.
Commands to reproduce training and generate speech from the submitted checkpoint.
A versioned release or commit identifying the exact submitted code.
Additional dataset
Provide a public download location, dataset description, license information, source references, and the scripts used to prepare the data. Include a manifest documenting its size, duration, speaker composition, and synthetic-data content.
Generated audio
Submit one WAV file for each organizer-provided submission text, following the prescribed naming and audio-format requirements. Samples must come from the submitted model and inference pipeline. Manual editing, cherry-picking, or sample-specific processing is not permitted. Upload one ZIP file containing all evaluation audio files to the model repository (preferred Hugging Face).
System overview
Provide an at most one A4 page report describing the approach, data preparation, model changes, training resources, and known limitations. Participants are encouraged to include a comparison with the unmodified baseline to show which changes improved performance.
Finally, submit all links and the required team information through Submission form. Repository links alone do not constitute a complete submission.
The organizers must be able to download and run the submitted system at the deadline. Reproducing inference is required; retraining the model from scratch is not required for routine evaluation.
Code, checkpoints, and additional training datasets must be publicly released no later than the official date of Dutch Speech Tech Day 2027. Each artifact must have a clear license permitting its intended use and redistribution.
If repositories remain private before public release, participants must provide organizer access at submission.
The submission must identify fixed code revisions and model files. Changes after the deadline require organizer approval and must be disclosed. Any accepted technical correction must not change the system’s evaluated behavior.
Participants retain ownership of their contributions and agree that submitted evaluation audio may be used for challenge evaluation and presentation of the results.
The submission deadline is Monday January 18, 2027, at 23:59 CET.
Participants are responsible for ensuring that all links work and all required files are accessible before the deadline.
Incomplete submissions, undisclosed training data or pretrained components, evaluation-data leakage, or systems that cannot run under the published CPU requirements may be excluded from ranking. The organizers may request clarification or evidence to verify compliance.
Challenge resources, evaluation specifications, and rule clarifications will be published through the current official challenge page. Clarifications affecting eligibility or evaluation will be shared with all participants.
Student participants or teams will receive one Q&A opportunity with an industry speech technology expert, as described in the announcement. You can send a request to chuhan.wang@readspeaker.com.
For questions about participation, data eligibility, or technical requirements, contact chuhan.wang@readspeaker.com.
You can access the data via CML-TTS Dutch Dataset.
We will release the scores from the listening tests after the evaluation. For full details on our process, please refer to section 4. Evaluation in the Challenge Regulations above.
The submission deadline is Monday, January 18, 2027, at 23:59 (CET).