README.md

---
license: cc-by-4.0
language:
- ca
- es
base_model:
- nvidia/stt_es_conformer_transducer_large
tags:
- automatic-speech-recognition
- NeMo
model-index:
- name: stt_ca-es_conformer_transducer_large
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: CV Benchmark Catalan Accents
      type: projecte-aina/commonvoice_benchmark_catalan_accents
      config: ca
      split: test
      args:
        language: ca
    metrics:
    - name: Test WER
      type: wer
      value: 2.503
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Mozilla Common Voice 17.0
      type: mozilla-foundation/common_voice_17_0
      config: ca
      split: test
      args:
        language: es
    metrics:
    - name: Test WER
      type: wer
      value: 3.88
---
# NVIDIA Conformer-Transducer Large (ca-es)

## Table of Contents
<details>
<summary>Click to expand</summary>

- [Model Description](#model-description)
- [Intended Uses and Limitations](#intended-uses-and-limitations)
- [How to Get Started with the Model](#how-to-get-started-with-the-model)
- [Training Details](#training-details)
- [Citation](#citation)
- [Additional Information](#additional-information)

</details>

## Summary

The "stt_ca-es_conformer_transducer_large" is an acoustic model based on ["NVIDIA/stt_es_conformer_transducer_large"](https://huggingface.co/nvidia/stt_es_conformer_transducer_large/) suitable for Bilingual Catalan-Spanish Automatic Speech Recognition.

## Model Description

This model transcribes speech in lowercase Catalan and Spanish alphabet including spaces, and was fine-tuned on a Bilingual ca-es dataset comprising of 7426 hours. It is a "large" variant of Conformer-Transducer, with around 120 million parameters.
See the [model architecture](#model-architecture) section and [NeMo documentation](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/models.html#conformer-transducer) for complete architecture details.

## Intended Uses and Limitations

This model can be used for Automatic Speech Recognition (ASR) in Catalan and Spanish. It is intended to transcribe audio files in Catalan and Spanish to plain text without punctuation.

### Installation

To use this model, install [NVIDIA NeMo](https://github.com/NVIDIA/NeMo). We recommend you install it after you've installed the latest PyTorch version.
```
pip install nemo_toolkit['all']
``` 

### For Inference
To transcribe audio in Catalan or in Spanish using this model, you can follow this example:


```python
import nemo.collections.asr as nemo_asr

nemo_asr_model = nemo_asr.models.EncDecRNNTBPEModel.restore_from(model)
transcription = nemo_asr_model.transcribe([audio_path])[0].text
print(transcription)
```

## Training Details

### Training data

The model was trained on bilingual datasets in Catalan and Spanish, for a total of 7k hours. Including:
- [Parlament-Parla-v3](https://huggingface.co/datasets/projecte-aina/parlament_parla_v3)
- [Corts Valencianes](https://huggingface.co/datasets/projecte-aina/corts_valencianes_asr_a)
- [3cat](https://www.isca-archive.org/iberspeech_2024/hernandezmena24_iberspeech.pdf) 
- [IB3](https://huggingface.co/datasets/projecte-aina/ib3_ca_asr) (The datasets will be made accessible shortly.)
- [ciempiess light](https://huggingface.co/datasets/ciempiess/ciempiess_light)
- [ciempiess fem](https://huggingface.co/datasets/ciempiess/ciempiess_fem)
- [ciempiess complementary](https://huggingface.co/datasets/ciempiess/ciempiess_complementary)
- [ciempiess balance](https://huggingface.co/datasets/ciempiess/ciempiess_balance)
- [CHM150](https://huggingface.co/datasets/carlosdanielhernandezmena/chm150_asr)
- [Tedx spanish](https://huggingface.co/datasets/ciempiess/tedx_spanish)
- [librivox spanish](https://huggingface.co/datasets/ciempiess/librivox_spanish)
- [Wikipedia spanish](https://huggingface.co/datasets/ciempiess/wikipedia_spanish)
- [voxforge spanish](https://huggingface.co/datasets/ciempiess/voxforge_spanish)
- [Tele con ciencia](https://huggingface.co/datasets/ciempiess/tele_con_ciencia)
- [Argentinian Spanish Speech Dataset](https://openslr.org/61/)
- [Dimex100 light](https://huggingface.co/datasets/carlosdanielhernandezmena/dimex100_light)
- [Glissando Spanish](https://glissando.labfon.uned.es/es)
- [Herico](https://openslr.org/39/)
- [Latino40](https://catalog.ldc.upenn.edu/LDC95S28)
- [Common voice 17 es](https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0)

### Training procedure

This model is the result of finetuning the base model ["Nvidia/stt_es_conformer_transducer_large"](https://huggingface.co/nvidia/stt_es_conformer_transducer_large) by following this [tutorial](https://github.com/NVIDIA/NeMo/blob/main/tutorials/asr/Transducers_with_HF_Datasets.ipynb).

## Citation
If this model contributes to your research, please cite the work:
```bibtex
@misc{conformer-transducer-BSC-2024,
      title={Bilingual ca-es ASR Model: stt_ca-es_conformer_transducer_large.}, 
      author={Messaoudi, Abir; Külebi, Baybars},
      organization={Barcelona Supercomputing Center},
      url={https://huggingface.co/projecte-aina/stt_ca-es_conformer_transducer_large},
      year={2024}
}
```

## Additional Information

### Author

The fine-tuning process was performed during 2024 in the [Language Technologies Unit](https://huggingface.co/BSC-LT) of the [Barcelona Supercomputing Center](https://www.bsc.es/) by [Abir Messaoudi](https://huggingface.co/AbirMessaoudi).

For the Catalan Valencian data we had the collaboration of [CENID](https://cenid.es/) within the framework of the [ILENIA](https://proyectoilenia.es) project.

### Contact
For further information, please send an email to <[email protected]>.

### Copyright
Copyright(c) 2024 by Language Technologies Unit, Barcelona Supercomputing Center.

### License

[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)

### Funding
This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project ILENIA with reference 2022/TL22/00215337.

The training of the model was possible thanks to the computing time provided by [Barcelona Supercomputing Center](https://www.bsc.es/) through MareNostrum 5.