Citrinet-1024  model with kernel scaling factor (gamma) of 25% which was fine-tuned from English language to Spanish. We took a model 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 (MCV7.0) , Multilingual LibriSpeech (MLS) , and Voxpopuli  training sets, as well as a subset of the Spanish Fisher dataset [5, 6]. This amounted to 1,340 hours of Spanish training data.
It utilizes a Google SentencePiece  tokenizer with vocabulary size 1024, and transcribes text in lower case Spanish alphabet characters.
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 .
The NeMo toolkit  was used for training this model over several hundred epochs on multiple GPUs.
While finetuning this model, we used the following datasets:
The tokenizer for this model was built using the text that was in the training set.
We build a Google Sentencepiece Tokenizer with the following script:
python [NEMO_GIT_FOLDER]/scripts/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
The performance of Automatic Speech Recognition models is measuring using Word Error Rate.
The model obtains the following greedy scores on the following evaluation datasets:
We also evaluated the model performance with beam search with a 4-gram KenLM model (included with this NGC model). The KenLM model was trained on the transcriptions of the Spanish speech data used to finetune the ASR model, and also a dataset of Spanish news text (about 50 million sentences) from the WMT competition's News Crawl corpus .
The following results were obtained with beam search with the 4-gram KenLM model, using hyperparameters alpha = 0.5, beta = 1.3, beam width = 128.
The model is available for use in the NeMo toolkit , and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.
import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.EncDecCTCModelBPE.from_pretrained(model_name="stt_es_citrinet_1024_gamma_0_25")
python [NEMO_GIT_FOLDER]/examples/asr/transcribe_speech.py \ pretrained_name="stt_es_citrinet_1024_gamma_0_25" \ audio_dir=""
This model accepts 16000 Hz Mono-channel Audio (wav files) as input.
This model provides transcribed speech as a string for a given audio sample.
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.
 Majumdar, Somshubra, et al. "Citrinet: Closing the Gap between Non-Autoregressive and Autoregressive End-to-End Models for Automatic Speech Recognition." arXiv preprint arXiv:2104.01721 (2021). link