This repository provides recipes to train (french) ASR models (either from scratch or by finetuning) using the NVIDIA NeMo toolkit.
The data used in these recipes is processed using the ssak NeMo pipeline. This pipeline takes Kaldi datasets and converts and cleans them into a single train_manifest.jsonl. Once generated, this manifest can be used to create a tokenizer or bucketed datasets to accelerate training.
You can analyze the number of hours per dataset and per bucket in your training, development, and test manifests by using the visualization tool.
Training checkpoints can be monitored using eval.py, which provides insights into training progress.
You can use the ASR Benchmark repository to evaluate the model's performance against other models.
The recipe linto_stt_fr_fastconformer finetunes the FastConformer-Hybrid Large FR model (base config file) on more than 9,000 hours of French speech to enhance its performance.
The list of datasets used during finetuning is available here.
The resulting model is available on Hugging Face: linto_stt_fr_fastconformer.