Model Overview

This collection contains large size versions of Conformer-Transducer (around 120M parameters) trained on ParlaSpeech-HR v1.0 dataset.

It utilizes a Google SentencePiece [2] tokenizer with vocabulary size 128, and transcribes text in lower case Croatian alphabet characters.

Model Architecture

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

Training

The NeMo toolkit [3] was used for training the models for over several hundred epochs. These model are trained with this example script and this base config.

The tokenizers for these models were built using the text transcripts of the train set with this script, and can be found inside the .nemo files.

The vocabulary we use is contains 27 characters:

[' ', 'a', 'b', 'c', 'č', 'ć', 'd', 'đ', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'r', 's', 'š', 't', 'u', 'v', 'z', 'ž']

Full config can be found inside the .nemo files.

Datasets

All the models in this collection are trained on ParlaSpeech-HR v1.0 Croatian dataset, which contains around 1665 hours of training data after data cleaning, 2.2 hours of developement and 2.3 hours of test data.

Tokenizer construction

python [NEMO_GIT_DIR]/scripts/tokenizers/process_asr_text_tokenizer.py \
  --manifest=train_manifest.json,dev_manifest.json \
  --data_root="[OUTPUT_DIR_FOR_TOKENIZER]" \
  --tokenizer="spe" \
  --spe_type="unigram" \
  --vocab_size=128 \
  --no_lower_case \
  --log

Performance

The list of the available models in this collection is shown in the following table. Performances of the ASR models are reported in terms of Word Error Rate (WER%) with greedy decoding.

You may use language models (LMs) and beam search to improve the accuracy of the models.

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.EncDecRNNTBPEModel.from_pretrained(model_name="stt_hr_conformer_transducer_large")

Transcribing text with this model

python [NEMO_GIT_DIR]/examples/asr/transcribe_speech.py \
    pretrained_name="stt_hr_conformer_transducer_large" \
    audio_dir="[AUDIO_DIR]"

Input

This model accepts 16 kHz single-channel audio as input.

Output

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

Limitations

Since the model is trained just on ParlaSpeech-HR v1.0 dataset, the performance of this model might degrade for speech which includes terms, or vernacular that the model has not been trained on. The model might also perform worse for accented speech.

Reference

[1] Conformer: Convolution-augmented Transformer for Speech Recognition

[2] Google Sentencepiece Tokenizer

[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. The user is responsible for checking if any datasets listed here are suitable for their intended use.