Model Overview

This collection contains large size versions of Conformer-Transducer (around 120M parameters) trained on NeMo ASRSet with over 1500 hours of French speech. The models was initialized with parameters from an English model and finetuned for French. Parameters from the three checkpoints with lowest Word Error Rate were then averaged to produce the final model.

Utilizes a Google SentencePiece [1] tokenizer with vocabulary size 128, and transcribes text in lower case French alphabet along with spaces, apostrophes, and hyphens.

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. 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.

The three highest performing checkpoint of the model were averaged with this script.

Datasets

All the models in this collection are trained on a composite dataset (NeMo ASRSET) comprised of over a thousand hours of French speech:

• MozillaCommonVoice7.0: 356 hours
• MultilingualLibreSpeech: 1036 hours
• VoxPopuli: 182 hours

(Note: The MultiliingualLibreSpeech set includes orthography of both pre- and post- 1990 reform, which may affect predictability of hyphenation.)

Tokenizer Construction

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

We build a token set with the following script:

python [NEMO_GIT_FOLDER]/scripts/tokenizers/process_asr_text_tokenizer.py \
 --manifest="train_manifest.json" \
 --data_root="" \
 --vocab_size=128 \
 --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. Since this dataset is trained on multiple domains and a much larger corpus, it will generally perform better at transcribing audio in general.

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

6.85 % on MCV7.0 dev 7.95 % on MCV7.0 test 5.05 % on MLS dev 4.10 % on MLS test

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_fr_conformer_transducer_large")

Transcribing text with this model

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

Further, since portions of the training set contain text from both pre- and post- 1990 orthographic reform, regularity of punctuation may vary between the two styles. For downstream tasks requiring more consistency, finetuning or downstream processing may be required.

References [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.