Design and Evaluation of Acoustic and Language Models for Large Scale Telephone Services

This paper describes the specification, design and development phases of two widely used telephone services based on automatic speech recognition. The effort spent for evaluating and tuning these services will be discussed in detail. In developing the first service, mainly based on the recognition of ``alphanumeric`` sequences, a significant part of the work consisted in refining the acoustic models. To increase recognition accuracy we adopted algorithms and methods consolidated in the past over broadcast news transcription tasks. A significant result shows that the use of task specific context dependent phone models reduces the word error rate by about 40% relative to using context independent phone models. Note that the latter result was achieved over a small vocabulary task, significantly different from those generally used in broadcast news transcription. We also investigated both unsupervised and supervised training procedures. Moreover, we studied a novel partly supervised technique that allows us to select in some ``optimal`` way the speech material to manually transcribe and use for acoustic model training. A significant result shows that the proposed procedure gives performance close to that obtained with a completely supervised training method. In the second service, mainly based on phrase spotting, a wide effort was devoted to language model refinement. In particular, several types of rejection networks were studied to detect out of vocabulary words for the given task; a major result demonstrates that using rejection networks based on a class trigram language model reduces the word error rate from 36.7% to 11.1% with respect to using a phone loop network. For the latter service, the benefits and related costs brought by regular grammars, stochastic language models and mixed language models will be also reported and discussed