Model Overview

Citrinet-1024 model which has been trained on over 2300 hours of German speech.

It utilizes a Google SentencePiece [1] tokenizer with vocabulary size 1024, and transcribes text in lower case German alphabet with spaces but without punctuation.

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

This model was pre-trained for 100 epochs on unlabeled German speech data, and then fine-tuned on a composite dataset comprising of several thousand labeled hours of speech, compiled from various publicly available sources. The NeMo toolkit [3] was used for training this model over several hundred epochs on multiple GPUs.

Datasets

For pre-training, we used the German unlabeled speech from the Vox Populi dataset (~20k hours).

While fine-tuning this model, we used the following cleaned datasets:

• MCV7.0 567 hours
• MLS 1524 hours
• VoxPopuli 214 hours

Tokenizer Construction

The tokenizer for this model was built using text corpus provided with the train dataset.

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="" \
 --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 measured using Word Error Rate.

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

• 6.63 % on MCV7.0 dev
• 7.59 % on MCV7.0 test
• 4.06 % on MLS dev
• 5.07 % on MLS test
• 12.33 % on Vox Populi dev
• 10.02 % on Vox Populi test

After beam search with a beam size of 128 and 4-gram KenLM model, the Word Error Rates are:

• 5.00 % on MCV7.0 dev
• 5.84 % on MCV7.0 test
• 3.76 % on MLS dev
• 4.50 % on MLS test
• 11.79 % on Vox Populi dev
• 9.42 % on Vox Populi test

Note that these evaluation datasets have been filtered and preprocessed to only contain German alphabet characters and are removed of punctuation.

How to Use this Model

The model is available for use in the NeMo toolkit [3], 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_de_citrinet_1024")

Transcribing text with this model

python [NEMO_GIT_FOLDER]/examples/asr/transcribe_speech.py \
 pretrained_name="stt_de_citrinet_1024" \
 audio_dir=""

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

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

[3] 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.