Utterance-based Selective Training for Cost-effective Task-adaptation of Acoustic Models

The construction of acoustic models for speech recognition systems is a very costly and time-consuming process, since their robust training requires large amounts of transcribed speech data, which have to be collected and labeled by humans. This paper describes an approach for costeffective construction of task-adapted acoustic models. Existing speech data(bases) are employed to set up a large training data pool. Apart from that, only a small amount of taskspecific speech data is required. Based on an algorithm for utterance-based selective training of acoustic models, training utterances are selected from the training data pool so that the likelihood of the acoustic model given the task-specific speech data is maximized. The proposed method is evaluated for acoustic models with context-independent and contextdependent phonetic units. Results are reported for building an infant (preschool children) acoustic model with speech from elementary school children and an elderly acoustic model with adult speech. The proposed approach is already effective if there are only 20 task-specific utterances available. A relative improvement in word accuracy of up to 10% is achieved over conventional acoustic model construction and up to 2.8% over MAP and MLLR adaptation with the task-specific data. The gap in performance to an acoustic model trained on large amounts of task-specific data was reduced up to 76%.