Error-driven Fixed-Budget ASR Personalization for Accented Speakers (ICASSP 2021)

This repository provides an implementation of experiments in our ICASSP-2021 paper

@inproceedings{awasthi2021error,
  title={Error-Driven Fixed-Budget ASR Personalization for Accented Speakers},
  author={Awasthi, Abhijeet and Kansal, Aman and Sarawagi, Sunita and Jyothi, Preethi},
  booktitle={ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={7033--7037},
  year={2021},
  organization={IEEE}
}

Requirements

This code was developed with python 3.8.9.
Create a new virtual environment and install the dependencies by running bash install_requirements.sh

Dataset

IndicTTS dataset used in our experiments can be obtained by contacting IndicTTS team at smtiitm@gmail.com. Please mention that you require data for all the accents in Indian English.

The dataset consists of utterances from various accented speakers. The speaker utterances are in the format <accent>_<gender>_english.zip. The data can be unzipped in appropriate format using the data/indicTTS_audio/unzip_indic scripts.

cd data/indicTTS_audio
python3 unzip_indic.py --input_path <path_to_speaker_zip_file>

Usage

• Generate transcripts for the seed+dev set using the pre-trainded ASR (Transcripts are used while training error models)

  cd models/quartznet_asr
  bash scripts/infer_transcriptions_on_seed_set.sh
  

• Train error model by aligning the references and generated transcripts for seed+dev set

  cd models/error_model
  bash train_error_model.sh
  

• Infer the trained error model on the set of sentences from which we wish to do the selection

  cd models/error_model
  bash infer_error_model.sh
  

• Select the sentences using the error model as proposed in our paper (Equation-2 and Algorithm-1)

  cd models/error_model
  bash error_model_sampling.sh
  

• Finetune and Test the ASR on the sentences selected via error model

  cd models/quartznet_asr
  bash scripts/finetune_on_error_model_seleced_samples.sh
  bash scripts/test_ASR_finetuned_on_error_model_sents.sh
  

• Finetune and Test the ASR on the randomly selected sentences

  cd models/quartznet_asr
  bash scripts/finetune_on_randomly_seleced_samples.sh
  bash scripts/test_ASR_finetuned_on_random_sents.sh
  

Contents

• data/$accent/manifests:
  • all.json: All the samples for a given accented speaker
  • seed.json: Randomly selected seed set for learning error model. Also used with selected samples while training ASR.
  • dev.json: Dev set used while training ASR. Also used with seed set while training error models
  • seed_plus_dev.json: Used for training error models (Concatenation of seed.json and dev.json)
  • selection.json: Sentences are selected from this file (either randomly or through error model, depending on selection startegy)
  • test.json: Used for evaluating the error model.
  • train/error_model/$size/seed_"$seed"/train.json: Contains size number of sentences selected via error model from selection.json, appended with seed.json for training the ASR model. seed represents an independent run.
  • train/random/$size/seed_"$seed"/train.json: Contains size number of sentences selected randomly from selection.json, appended with seed.json for training the ASR model. seed represents an independent run.
• data/indicTTS_audio:
  • Audio files for each speaker are placed here.
  • After obtaining .zip files from IndicTTS team, place them in this folder.
  • Run unzip_indic.sh to unzip the speaker data of your choice.
• models/pretrained_checkpoints/quartznet: Pretrained Quartznet models:
  • librispeech/quartznet.pt: Quartznet ckpt trained on 960hrs librispeech
  • librispeech_mcv_others/quartznet.pt: Quartznet ckpt trained on 960hrs librispeech + Mozilla Common Voice + Few other corpus
  • The above checkpoints were originally provided by NVIDIA to work with NeMO toolkit. We modified these slightly to make them work with Jasper's pytorch code.
  • Link to original checkpoints: Librispeech, Librispeech+MCV+Others
• models/pretrained_checkpoints/error_models
  • $accent/seed_"$seed"/best/ErrorClassifierPhoneBiLSTM_V2.pt : Pre-trained error model on seed_plus_dev.json outputs.
  • librispeech/seed_"$seed"/best/ErrorClassifierPhoneBiLSTM_V2.pt: Pre-trained error model on dev+test sets of librispeech.
• models/error_model : Code related to error model
  • train_error_model.sh: Trains the error model using the ASR's transcripts on the seed+dev set and gold references
  • infer_error_model.sh: Dumps aggregated error probabilities from the error model for sentence scoring
  • error_model_sampling.sh: Samples the sentences using error model outputs as per Equation-2 and Algorithm-1 in our paper.
• models/quartznet_asr: Code related to training and inference of Quartznet model, adapted from Nvidia's implementation of Jasper in PyTorch
  • scripts/finetune_on_error_model_seleced_samples.sh: Finetunes the ASR on sentences utterances selected by error model
  • scripts/finetune_on_randomly_seleced_samples.sh: Finetunes the ASR on randomly selected sentences
  • scripts/test_ASR_finetuned_on_error_model_sents.sh: Infers the ASR finetuned on error model selected sentences
  • scripts/test_ASR_finetuned_on_random_sents.sh: Infers the ASR on randomly selected sentences
  • scripts/infer_transcriptions_on_seed_set.sh: Infers the ASR on seed+dev set. Inferred outputs used for training the error model.

Acknowledgement

• Dataset for a wide variety of Indian accents was provided by IndicTTS team at IIT Madras. Dataset for other accents was obtained from L2-Arctic Corpus
• Code for Quartznet ASR model is adapted from Nvidia's implementation of Jasper in PyTorch. Pretrained quartznet checkpoints were downloaded from here and modified slightly to make them compatible with Jasper's PyTorch implementation.
• Power-ASR is used for aligning ASR-generated transcripts and references for obtaining error labels.
• Park et al's Grapheme-to-Phoneme model is used for converting graphemes to phonemes as an input to the error model .