Loquendo Acoustic Model Adaptation (AMA) is able to adapt acoustic models using data collected in the field, i.e. when the application is running.
The general scheme for Loquendo learning and adaptation from data collected in the field is outlined in Fig. 1. The diagram also features the Phonetic Learning tool and LogCollector tool. The Loquendo AMA uses the LogCollector tool to gather the adaptation data, and can be used in conjunction with the Phonetic Learning tool to improve the performance of the general acoustic model.

The acoustic model adaptation (and phonetic learning) architecture

As depicted in Fig.1, the global scheme is based on a three-step process:
(1) Data acquisition,
(2) Data analysis, and
(3) Adapted model activation.

For data acquisition, the user must acquire new data from the recognizers running on some recognition servers. With Loquendo ASR, it is possible to do this by activating the dumping of the Log data. In this way, the information necessary for adaptation is saved during the running of the application, e.g. the speech signal, phonetic transcription, etc.

Subsequently, data analysis must be carried out using Loquendo ASR installed on an off-line workstation. The LogCollector tool allows the selection and gathering of the Log data and makes it available to Loquendo AMA, which produces the adapted acoustic model. The adaptation takes place in unsupervised mode, where no human input is required to transcribe the collected material.

The material needed for adaptation requires some further comment:
1. Quantity of vocal material required may vary from a few minutes (for speaker adaptation) to a few hours. In general, the larger the amount of material, the better the results will be. It must be noted however, that adaptation time will increase accordingly.

2. Kind of vocal material required. Users are asked to bear in mind that Loquendo AMA adapts to everything that is present in the adaptation material, and only to what is present. Thus, if you want to preserve speaker independence, collect material from many different speakers.
However, Loquendo AMA incorporates an innovative, patented technology that avoids damaging the general performances of the adapted models even when the adaptation material is too limited and polarized in terms of lexicon.

The final adapted model activation can be carried out by exporting the adapted model from the off-line workstation and importing it into the recognition servers; the application is then configured/re-run to switch from the general acoustic model to the adapted model. The final model shows increased recognition performance, without any additional computational charge.

Following the adaptation phase, the application developer should validate the proposed adapted model using the Loquendo ASR Evaluation Tool Kit. This tool is an application-based software console developed for Microsoft Windows 2000/XP and Linux operating systems.

In difficult cases, when the performance of the general models is not high enough (less than 70% WA), you may wish to try using the "Supervised Loquendo AMA", an auxiliary tool not released in the standard distribution but provided only on request. To use the "Supervised Loquendo AMA" some vocal material has to be collected and transcribed manually by a human operator. Please contact Loquendo for more details.