Building Speech Recognition System from Untranscribed Data

• Research Group of the 2016 Third Frederick Jelinek Memorial Summer Workshop

Summary of Goals and Objectives

Plenary Lectures

• July 15
  Brian Roark (Google Research)
  Good-Turing Estimation from Uncertain Data for Semi-Supervised Language Model Adaptation

Slides from Weekly Update Presentations (coming soon)

Findings and Accomplishments

• Lukáš Burget – Project leader. Introduction to unsupervised ASR framework and Bayesian generative modeling approach as a potential solution
• Lucas Ondel – Acoustic Unit Discovery.  A summary of the Jackie Lee Bayesian non-parametric model for acoustic unit discovery and improvements due to variational inference and phonotactic constraints
• Yibo Yang – Alternative Inference Method. Collapsed Variational Bayes for GMM training
• Leda Sari – Multilingual acoustic unit discovery. Using cross lingual AUDs as priors and multilingual bottleneck features
• Alena Rott – Bootstrapped acoustic feature transformations using linear discriminant analysis for improved acoustic unit discovery.linear discriminant analysis for improved acoustic unit discovery
• Jinyi Yang – Iterative training of bottleneck features for improved acoustic unit discovery.bottleneck features for improved acoustic unit discovery
• Chunxi Liu – Weak supervision through word classes for bootstrapped acoustic modeling
• Ming Sun – Same different task for evaluating acoustic unit discovery
• Santosh Kesiraju – Topic ID and topic clustering for evaluation of acoustic unit discovery
• Lukáš Burget – Preview of unsupervised speech recognition architectures
• Thomas Glarner – Nested Hierarchical Pitman Yor Language models for joint acoustic unit and word discovery
• Mirko Hanneman – Bayesian pronunciation lexicon learning based on Graphone Hierarchical Pitman Yor language model ; Ondrej Cífka – Supervised letter-to-phone training Preliminary unsupervised ASR results ; Mirko Hanneman – Unsupervised letter-to-phone training
• Takahiro Shinozaki – Pitman Yor lexicon model for word induction from unsegmented phone sequences and language model
• Daichi Mochihashi – A WFST-based Nested Pitman Yor language model on phones with stochastic prunining for word discovery from acoustic observations
• Matthew Wiesner – Knowledge transfer for unsupervised speech recognition through the use of cross-lingual graphemic acoustic models
• All – Group Question And Answer

Team Final Report

• Final report

Publications (TBA)