Model Overview

This collection contains large size versions of Conformer-CTC (around 120M parameters) trained on Mozilla Common Voice 10.0 (Russian), Golos (Russian), Russian LibriSpeech (RuLS) and SOVA (RuAudiobooksDevices, RuDevices) datasets with around 1636 hours of Russian speech.

It utilizes a Google SentencePiece [1] tokenizer with vocabulary size 128, and transcribes speech in lowercase russian alphabet along with spaces.

Model Architecture

Conformer-CTC model is a non-autoregressive variant of Conformer model [1] for Automatic Speech Recognition which uses CTC loss/decoding instead of Transducer. You may find more info on the detail of this model here: Conformer-CTC Model.

Training

The NeMo toolkit [2] was used for training the models for over several hundred epochs. These models are trained with this example script and this base config.

Datasets

All the models in this collection are trained on a composite dataset (NeMo ASRSET) comprising of more than a thousand hours of Russian speech:

• Mozilla Common Voice 10.0 (Russian) - train subset [28 hours]
• Golos - crowd [1070 hours] and fairfield [111 hours] subsets
• Russian LibriSpeech (RuLS) [92 hours]
• SOVA - RuAudiobooksDevices [260 hours] and RuDevices [75 hours] subsets

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="<OUTPUT DIRECTORY FOR TOKENIZER>" \
  --vocab_size=128 \
  --tokenizer="spe" \
  --spe_type="unigram" \
  --spe_character_coverage=1.0 \
  --no_lower_case \
  --log

Performance

The list of the available models in this collection is shown in the following table. The performance of Automatic Speech Recognition models are reported in terms of Word Error Rate (WER%) with greedy decoding. Since this dataset is trained on multiple domains and a much larger corpus, it will generally perform better at transcribing audio in general.

How to Use this Model

The model is available for use in the NeMo toolkit [2], 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_ru_conformer_ctc_large")

Transcribing text with this model

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

Input

This model accepts 16000 Hz 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] Conformer: Convolution-augmented Transformer for Speech Recognition

[2] NVIDIA NeMo Toolkit

[3] Google Sentencepiece Tokenizer

Licence

License to use this model is covered by the CC-BY-4 License 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 CC-BY-4 License.