Model Overview

Citrinet-512 model which has been trained on various Spanish corpus.

It utilizes a Google SentencePiece [1] tokenizer with vocabulary size 1024, and transcribes text in lower case Spanish alphabet alphabet characters.

Model Architecture

Citrinet is a deep residual convolutional neural network architecture that is optimized for Automatic Speech Recognition tasks. There are many variants of the Citrinet family of models, which are further discussed in the paper [2].

Training

We took an encoder from the English version of the CitriNet network trained on over 7,000 hours of English speech. Then we changed the model's decoder to output characters from the Spanish alphabet and fine-tuned this model using Spanish portions of the Mozilla CommonVoice (MCV) [3] and Multilingual LibriSpeech (MLS) [4] training sets. The NeMo toolkit [5] was used for training this model over several hundred epochs on multiple GPUs.

While finetuning this model, we used the following datasets:

• Mozilla Common Voice (Spanish) - 830 hours of training data after data cleaning
• Multilingual LibriSpeech (Spanish) - 235 hours of training data after data cleaning

Datasets

While training this model, we used the following datasets:

• Librispeech 960 hours of English speech (pre-training)
• Fisher Corpus (pre-training)
• Switchboard-1 Dataset (pre-training)
• WSJ-0 and WSJ-1 (pre-training)
• National Speech Corpus - 1 (pre-training)
• Mozilla Common Voice (pre-training)
• Mozilla Common Voice (Spanish) (fine-tuning)

Tokenizer Construction

The tokenizer for this model was built using the text that was in the training set.

We build a Google Sentencepiece Tokenizer [1] with the following script :

python [NEMO_GIT_FOLDER]/scripts/tokenizers/process_asr_text_tokenizer.py \
  --manifest="train_manifest.json" \
  --data_root="<OUTPUT DIRECTORY FOR TOKENIZER>" \
  --vocab_size=1024 \
  --tokenizer="spe" \
  --spe_type="unigram" \
  --spe_character_coverage=1.0 \
  --no_lower_case \
  --log

Performance

The performance of Automatic Speech Recognition models is measuring using Word Error Rate.

The model obtains the following scores on the following evaluation datasets -

• 9.1 % on MCV dev set
• 10.3 % on MCV test set
• 4.9 % on MLS dev set
• 5.2 % on MLS test set

Note that these scores above are not particularly indicative of the quality of transcriptions that models trained on ASR Set will achieve, but they are a useful proxy.

How to Use this Model

The model is available for use in the NeMo toolkit [5], and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.

Automatically load the model from NGC

import nemo.collections.asr as nemo_asr

asr_model = nemo_asr.models.EncDecCTCModelBPE.from_pretrained(model_name="stt_es_citrinet_512")

Transcribing text with this model

python [NEMO_GIT_FOLDER]/examples/asr/transcribe_speech.py \
  pretrained_name="stt_es_citrinet_512" \
  audio_dir="<DIRECTORY CONTAINING AUDIO FILES>"

Input

This model accepts 16000 KHz Mono-channel Audio (wav files) as input.

Output

This model provides transcribed speech as a string for a given audio sample.

Limitations

Since this model was trained on publically available speech datasets, the performance of this model might degrade for speech which includes technical terms, or vernacular that the model has not been trained on. The model might also perform worse for accented speech.

References

[1] Google Sentencepiece Tokenizer

[1] Citrinet: Closing the Gap between Non-Autoregressive and Autoregressive End-to-End Models for Automatic Speech Recognition

[3] Mozilla CommonVoice (MCV)

[4] Multilingual LibriSpeech (MLS)

[5] NVIDIA NeMo Toolkit

Licence

License to use this model is covered by the NGC TERMS OF USE unless another License/Terms Of Use/EULA is clearly specified. By downloading the public and release version of the model, you accept the terms and conditions of the NGC TERMS OF USE.