Cantab Research is rewriting large vocabulary speech recognition to make it far faster and more accurate. We have completed the first stage of our language modelling phase where we have trained very large RNNLMs with many novel features which approximately halve perplexity over KN 5-grams and are far more compact (ideal for ASR, MT and predictive keyboards). We are now looking to make similar gains in acoustic modelling using recurrent networks and the decoder using our own novel design and to deploy this in many languages.

We have very many positions open.   Candidates should have good communication skills and one or more of:

• several years experience with Kaldi, HTK or similar
• several years experience with C/C++ particularly in writing fast efficient code
• have written or fully understand the code to implement Viterbi algorithm and it's variants
• data collection and/or text normalisation in very many languages
• have experience with real time APIs for cloud provisioned ASR
• have implemented ASR on Android and/or iOS
• have written significant CUDA
• have experience with neural net (RNN/DNN/LSTM) acoustic models and language models
• know WFST/lattice processing algorithms
• have worked with reverberant and noise robust speech recognition
• have implemented diarisation and/or speaker identification