Model Overview

Conformer-CTC model which was fine-tuned from English language to Spanish. We took a model trained on over 7,000 hours of English speech. Then we changed the model's decoder to output characters from the Spanish alphabet and fine-tuned this model using 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]. This amounted to 1,340 hours of Spanish training data.

It utilizes a Google SentencePiece [7] tokenizer with vocabulary size 1024, and transcribes text in lower case Spanish alphabet characters.

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 [8] was used for training this model over several hundred epochs on multiple GPUs.

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

Tokenizer Construction

The tokenizer for this model was built using the text that was in the training set.

We build a Google Sentencepiece Tokenizer with the following script:

python [NEMO_GIT_FOLDER]/scripts/process_asr_text_tokenizer.py \
  --manifest="train_manifest.json" \
  --data_root="<OUTPUT DIRECTORY FOR TOKENIZER>" \
  --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 measuring using Word Error Rate.

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

• 6.3 % WER on MCV7.0 dev set
• 4.3 % WER on MLS dev set
• 6.1 % WER on Voxpopuli dev set
• 18.3 % WER on Fisher dev set (manually created)

• 6.9 % WER on MCV7.0 test set
• 4.2 % WER on MLS test set
• 7.5 % WER on Voxpopuli test set
• 18.5 % WER on Fisher test set (manually created)

• 22.4 % WER on Spanish CALLHOME train set [9,10] (model was not trained on this data)
• 23.7 % WER on Spanish CALLHOME devtest set [9,10] (model was not trained on this data)
• 25.3 % WER on Spanish CALLHOME evltest set [9,10] (model was not trained on this data)

We also evaluated the model performance with beam search with a 4-gram KenLM model (included with this NGC model). The KenLM model was trained on the transcriptions of the Spanish speech data used to finetune the ASR model, and also a dataset of Spanish news text (about 50 million sentences) from the WMT competition's News Crawl corpus [11].

The following results were obtained with beam search with the 4-gram KenLM model, using hyperparameters alpha = 0.8, beta = 1.5, beam width = 128.

• 5.0 % WER on MCV7.0 dev set
• 3.6 % WER on MLS dev set
• 5.5 % WER on Voxpopuli dev set
• 17.4 % WER on Fisher dev set (manually created)

• 5.5 % WER on MCV7.0 test set
• 3.6 % WER on MLS test set
• 6.7 % WER on Voxpopuli test set
• 17.5 % WER on Fisher test set (manually created)

• 20.7 % WER on Spanish CALLHOME train set [9,10] (model was not trained on this data)
• 21.6 % WER on Spanish CALLHOME devtest set [9,10] (model was not trained on this data)
• 23.3 % WER on Spanish CALLHOME evltest set [9,10] (model was not trained on this data)

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.EncDecCTCModelBPE.from_pretrained(model_name="stt_es_conformer_ctc_large")

Transcribing text with this model

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

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] 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] Google Sentencepiece Tokenizer

[8] NVIDIA NeMo Toolkit

[9] CALLHOME Spanish Transcripts

[10] CALLHOME Spanish Speech

[11] WMT competition text datasets

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.