TUDelft/inclusive-asr-moe

Inclusive ASR MoE

This repository contains the final NVIDIA NeMo automatic speech recognition checkpoints developed by Amelia Sasin for the MSc thesis:

Mixture-of-Experts for Age-Inclusive ASR: Reducing the Adult-Child Performance Gap in Multilingual Speech Recognition Delft University of Technology, 2026.

The models investigate inclusive and robust speech recognition for adult and child speech using dense FastConformer and sparse Mixture-of-Experts architectures.

Source code, training configurations, preprocessing scripts, evaluation scripts, and reproducibility instructions are available in the GitHub repository:

a-sasin/inclusive-asr-moe

Base model

All models were initialized directly or indirectly from:

nvidia/stt_en_fastconformer_ctc_large

The released checkpoints include further training, child-speech adaptation, multilingual training, and Mixture-of-Experts modifications performed by Amelia Sasin.

These models are independent research derivatives and are not endorsed by NVIDIA.

Available checkpoints

Checkpoint	Language scope	Training population	Architecture
en_adult_fastconformer.nemo	English	Adult	Dense FastConformer
en_adult_moe.nemo	English	Adult	Mixture of Experts
en_child_fastconformer.nemo	English	Child	Dense FastConformer
en_child_moe_lb_off.nemo	English	Child	MoE, load balancing disabled
en_child_moe_lb_on.nemo	English	Child	MoE, load balancing enabled
multilingual_adult_fastconformer.nemo	English, Dutch, German, Polish	Adult	Dense FastConformer
multilingual_adult_moe.nemo	English, Dutch, German, Polish	Adult	Mixture of Experts
multilingual_child_fastconformer.nemo	English, Dutch, German, Polish	Child	Dense FastConformer
multilingual_child_moe_lb_off.nemo	English, Dutch, German, Polish	Child	MoE, load balancing disabled
multilingual_child_moe_lb_on.nemo	English, Dutch, German, Polish	Child	MoE, load balancing enabled

Downloading a checkpoint

from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="TUDelft/inclusive-asr-moe",
    filename="en_adult_fastconformer.nemo",
)

Loading a checkpoint

from nemo.collections.asr.models import ASRModel

model = ASRModel.restore_from(
    restore_path=checkpoint_path,
)

model.eval()
transcriptions = model.transcribe(["example.wav"])
print(transcriptions)

Required NeMo fork

The Mixture-of-Experts checkpoints require the custom NeMo fork developed for this work:

a-sasin/NeMo

This fork contains the sparse Mixture-of-Experts implementation used to train and restore the released MoE checkpoints.

Install it from source:

git clone https://github.com/a-sasin/NeMo.git
cd NeMo
pip install -e ".[asr]"

The dense FastConformer checkpoints may work with standard NVIDIA NeMo, but using this fork is recommended for consistency with the training environment.

Reproducible installation

For exact reproducibility, check out the same NeMo commit used during training:

git clone https://github.com/a-sasin/NeMo.git
cd NeMo
git checkout <NEMO_COMMIT_HASH>
pip install -e ".[asr]"

The Mixture-of-Experts checkpoints may require the custom NeMo implementation documented in the accompanying GitHub repository.

Data

The training and evaluation datasets, preprocessing procedures, manifest formats, and access instructions are documented in the accompanying GitHub repository and thesis.

Raw speech datasets are not included in this model repository. Users are responsible for obtaining the datasets from their original providers and complying with their respective licenses and access conditions.

Intended use

The checkpoints are intended for research involving:

• Automatic speech recognition
• Child and adult speech recognition
• Multilingual ASR
• Inclusive speech technology
• Mixture-of-Experts architectures
• Expert routing and load-balancing analysis

Limitations

Recognition performance may vary depending on:

• Speaker age
• Accent and dialect
• Language
• Speech characteristics
• Recording conditions
• Background noise
• Microphone quality
• Speaking style
• Domain-specific vocabulary

The terms “adult” and “child” describe the training and evaluation data. These models are not age-classification systems.

The models should not be treated as error-free or used without additional validation in medical, legal, emergency, educational-assessment, surveillance, or other high-stakes applications.

License

Model checkpoints

The model checkpoints are released under the Creative Commons Attribution 4.0 International license.

The checkpoints derive from NVIDIA’s stt_en_fastconformer_ctc_large, which is also distributed under CC BY 4.0.

When redistributing or adapting these checkpoints, retain attribution to:

• NVIDIA Corporation for the original checkpoint
• Amelia Sasin for the subsequent training, architecture modifications, experiments, and released checkpoints

Source code

Original source code authored for the thesis is licensed separately under the MIT License in the accompanying GitHub repository.

Copyright © 2025 Amelia Sasin.

Citation

@mastersthesis{sasin2026inclusiveasr,
  author = {Sasin, Amelia},
  title = {Mixture-of-Experts for Age-Inclusive ASR: Reducing the Adult-Child Performance Gap in Multilingual Speech Recognition},
  school = {Delft University of Technology},
  year = {2026}
}

---