Model Overview

Conformer-Transducer Large model, which was initially pre-trained on the English NeMo ASRSet with around 16000 hours of english speech and then finetuned on a combination of Spanish and English labeled speech data.

It can transcribe audio samples into English or Spanish. The language is detected automatically.

We started with the english Conformer-Transducer large model and finetuned it using labelled Spanish portions of the Mozilla CommonVoice (MCV7.0) [2], Multilingual LibriSpeech (MLS) [3], and Voxpopuli [4] training sets, as well as a subset of the Spanish Fisher dataset [5, 6], plus English Librispeech [7]. This amounted to 1,340 hours of Spanish training data and 960 hours of English data.

Model Architecture

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

Training

The NeMo toolkit [8] was used for training this model over several hundred epochs on multiple GPUs using bfloat16.

The English and Spanish tokenizers for these models were built separately using the English and Spanish text transcripts respectively of the train set with this script, idential to the procedure for large mono-lingual models (e.g. 1024 tokens). Both tokenizers were then used during training via the AggregateTokenizer construct. The total number of tokens in the model is 2048 (1024 for english and 1024 for spanish).

Datasets

While finetuning this model, we used the following datasets:

• Mozilla Common Voice 7.0 (Spanish) - 289 hours of training data after data cleaning
• Multilingual LibriSpeech (Spanish) - 801 hours of training data after data cleaning
• Voxpopuli transcribed subset (Spanish) - 110 hours of training data after data cleaning
• Fisher dataset (Spanish) - 140 hours of training data after data cleaning
• Librispeech (English) - 960 hours of training data

Performance

The performance of Automatic Speech Recognition models is measured using Word Error Rate (WER).

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

• 5.0 % WER on MCV7.0 dev set (es)
• 3.3 % WER on MLS dev set (es)
• 5.3 % WER on Voxpopuli dev set (es)
• 16.2 % WER on Fisher dev set (es)
• 2.0 % WER on LibriSpeech dev-clean (en)
• 2.2 % WER on LibriSpeech test-clean (en)
• 4.6 % WER on LibriSpeech dev-other (en)
• 4.6 % WER on LibriSpeech test-other (en)

The model was not trained on the above datasets.

How to Use this Model

The model is available for use in the NeMo toolkit [5], 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_enes_conformer_transducer_large")

Transcribing text with this model

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

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. The output string may contain English or Spanish characters, depending on the detected language in the audio sample.

Limitations

Since this model was trained on publicly 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.

Finally, the model may not peform well on true "code-switching" samples, where each sample contains more than one language. This is because the model was trained only on samples where each sample had only one language. We will address this in the future releases.

References

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

[2] Mozilla CommonVoice (MCV7.0)

[3] Multilingual LibriSpeech (MLS)

[4] VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation

[5] Fisher Spanish - Transcripts

[6] Fisher Spanish Speech

[7] LibriSpeech ASR Corpus

[8] NVIDIA NeMo Toolkit

[9] CALLHOME Spanish Transcripts

[10] CALLHOME Spanish Speech

License

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.