As one essential part of Automatic Speech Recognition, statistical Language Models (LMs) learn which word sequences are likely and which are not. Hence, LMs implicitly learn which words are likely to occur in the beginning of an utterance and where an utterance is likely to end. Naturally, LMs perform better if they are provided with meaningful chunks of speech during training.

However, in conversational speech, chunking (i.e., separating the speech signal into meaningful segments) is challenging since in spoken language, we (humans) allow for much more flexibility – e.g., in terms of grammar – than in written language. For instance, when having a casual conversation with a friend, we often refrain from rephrasing broken sentences or correcting mispronounced words, and we are usually still able to communicate efficiently even when we produce highly disfluent or incomplete sentences.

Feeding such ‘broken’ sentences (in terms of grammar) into LMs is likely to result in bad performance. Hence, for obtaining meaningful chunks, we can use additional properties of speech, such as communicative functions, that humans include in their interpretation of conversational speech (e.g., does the same speaker continue talking or are they done and the other speaker can talk now). Since it is very effortful to annotate communicative functions by hand, it is desirable to make use of prosodic features (e.g., fundamental frequency, intensity, speech rate) to automatically classify different communicative functions.

The aim of this project is to build a semi-supervised learning based classifier for automatic labelling of communicative functions.

Your Profile

• interest in speech phenomena and machine learning
• good experience in Python
• background in machine learning is appreciated

Your tasks

• literature research, esp. on suitable algorithms
• data preprocessing
• implementation of a classifier
• evaluation of classification results

Groups are welcome!

Contact:

Saskia Wepner (wepner@tugraz.at)