PraxySante/stt_ca-es_conformer_transducer_large_fine_tuned

NVIDIA Conformer-Transducer Large (ca-es)

Table of Contents

Click to expand

• Model Description
• Intended Uses and Limitations
• How to Get Started with the Model
• Training Details
• Citation
• Additional Information
• Model Updates

Summary

The "stt_ca-es_conformer_transducer_large" is an acoustic model based on "NVIDIA/stt_es_conformer_transducer_large" suitable for Bilingual Catalan-Spanish Automatic Speech Recognition.

Model Description

This model transcribes speech, and was fine-tuned on a Bilingual ca-es dataset comprising of 4000 hours. It is a "large" variant of Conformer-Transducer, with around 120 million parameters. We expaneded it is tokenizer vocab sise to be 5.5k t oinclude lowercase, uppercase, and punctuation See the model architecture section and NeMo documentation 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 with punctuation.

Installation

To use this model, install 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:

import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.ASRModel.from_pretrained("PraxySante/stt_ca-es_conformer_transducer_large_fine_tuned")
transcriptions = asr_model.transcribe(["file.wav"])transcription = nemo_asr_model.transcribe([audio_path])[0].text
print(transcription)

Training Details

Training data

The model was fine-tuned on bilingual datasets in Catalan and Spanish, for a total of 4k hours. Including:

• Parlament-Parla-v1
• multilingual_librispeech
• basque_parliament_1
• Voxpopuli (The datasets will be made accessible shortly.)
• Coser
• tv3_parla
• common_voice_16_0

Training procedure

This model is the result of finetuning the model "projecte-aina/stt_ca-es_conformer_transducer_large"

Results

The results were calculated using the validation dataset, which is a held-out split from the same data used for training. This split contains 174 hours of audio for each language

Spanish WER: 0.08 Catalan WER: 0.10

Spanish CER: 0.04 Catalan CER: 0.05

Model Updates

We plan to provide regular updates to this model and its documentation. Future releases may include more performant versions as we continue to improve training strategies, expand dataset coverage, and incorporate community feedback.