projecte-aina / stt-ca-citrinet-512

Aina Project's Catalan text-to-speech model

Model description

This model transcribes audio samples in Catalan to lowercase text without punctuation. The model was fine-tuned from a pre-trained Spanish stt-es-citrinet-512 model using the NeMo toolkit. It has around 36.5M parámeters and has been trained on Common Voice 11.0 .

Intended uses and limitations

You can use this model for Automatic Speech Recognition (ASR) in Catalan, to transcribe audio files in Catalan to plain text without punctuation.

How to use

Usage

Requiered libraries:

pip install nemo_toolkit['all']

Clone the repository to download the model:

git clone https://huggingface.co/projecte-aina/stt-ca-citrinet-512

Given that NEMO_PATH is the path that points to the downloaded stt-ca-citrinet-512.nemo file, to do inference over a set of .wav files you should:

# Load the model
model = nemo_asr.models.EncDecCTCModel.restore_from(NEMO_PATH)

# Create a list pointing to the audio files
paths2audio_files = ["audio_1.wav", ..., "audio_n.wav"]

# Fix the batch size to whatever number suits your purpose
batch_size = 8

# Transcribe the audio files
transcriptions = model.transcribe(paths2audio_files=paths2audio_files,
                                 batch_size=batch_size)
# Visualize the transcriptions
print(transcriptions)

Training data

This model has been trained on the training split of the Catalan version of Common Voice 11.0 .

Training

Data preparation

We have processed Common Voice 11.0 using the NeMo toolkit. We used get_commonvoice_data.py to process the manifests and made a subsequent data cleaning step.

After cleaning the dataset and normalizing the ñ character to ny , we have used the following charset to create the final NeMo manifests for training.

['c', ' ', 'ó', 'g', 'a', 'o', 'ü', 'v', 'p', 't', "'", '—', 'f', 'k', 'à', 'ï', 'í', 'ú', 'd', 'l', 'z', 'é', 'w', 'm', 'r', 'n', 'y', '-', 'u', 'i', 'h', 'ç', 'e', '·', 'q', 'è', 'ò', 'j', 'x', 's', 'b']

Training procedure

This model was trained starting from a pre-trained Spanish stt-es-citrinet-512 model. The initial learning rate was set to 0.005 and the minimum lr for weight decay was set to 1e-7.

The model was trained for 90 steps and then continued training for another 90 steps starting from a learning rate of 1e-4.

Evaluation

After evaluation on the test split of Common Voice 11.0 we have obtained a WER of 6.684.

Additional information

Author

Text Mining Unit (TeMU) at the Barcelona Supercomputing Center ( [email protected] )

Contact information

For further information, send an email to [email protected]

Copyright

Copyright (c) 2022 Text Mining Unit at Barcelona Supercomputing Center

Licensing Information

Apache License, Version 2.0

Funding

This work was funded by the Generalitat de Catalunya within the framework of Projecte AINA .

Disclaimer