The Marchex 2018 English Conversational Telephone Speech Recognition System

In this paper, we describe recent improvements to the production Marchex speech recognition system for our spontaneous customer-to-business telephone conversations. We outline our semi-supervised lattice-free maximum mutual information (LF-MMI) training process which can supervise over full lattices from unlabeled audio. We also elaborate on production-scale text selection techniques for constructing very large conversational language models (LMs). On Marchex English (ME), a modern evaluation set of conversational North American English, for acoustic modeling we report a 3.3% ({agent, caller}:{3.2%, 3.6%}) reduction in absolute word error rate (WER). For language modeling, we observe a separate {1.3%, 1.2%} point reduction on {agent, caller} utterances respectively over the performance of the 2017 production system.

Semi-Supervised Model Training for Unbounded Conversational Speech Recognition

For conversational large-vocabulary continuous speech recognition (LVCSR) tasks, up to about two thousand hours of audio is commonly used to train state of the art models. Collection of labeled conversational audio however, is prohibitively expensive, laborious and error-prone. Furthermore, academic corpora like Fisher English (2004) or Switchboard (1992) are inadequate to train models with sufficient accuracy in the unbounded space of conversational speech. These corpora are also timeworn due to dated acoustic telephony features and the rapid advancement of colloquial vocabulary and idiomatic speech over the last decades. Utilizing the colossal scale of our unlabeled telephony dataset, we propose a technique to construct a modern, high quality conversational speech training corpus on the order of hundreds of millions of utterances (or tens of thousands of hours) for both acoustic and language model training. We describe the data collection, selection and training, evaluating the results of our updated speech recognition system on a test corpus of 7K manually transcribed utterances. We show relative word error rate (WER) reductions of {35%, 19%} on {agent, caller} utterances over our seed model and 5% absolute WER improvements over IBM Watson STT on this conversational speech task.