ŽODŽIŲ ATPAŽINIMO, GRĮSTO PASLĖPTAISIAIS MARKOVO MODELIAIS, VIZUALIZAVIMO IR ANALIZĖS PROGRAMINĖ ĮRANGA
A TOOL FOR VISUALIZATION AND ANALYSIS OF ISOLATED WORD RE COGNITION BA SED ON THE HIDDEN MAR KOV MODELS

Summary/Abstract: The paper presents a prototype of the isolated word recognition system based on hidden Markov models. The developed prototype of the speakerindependent Lithuanian isolated word recognition system is handy for recognition experiments and the analysis of their results. The user is provided with numeric and visual recognition information on the results. The word recognition pivots on the precision of the determination of the word limits. The main window contains a recognized word and its logarithmic likelihood, a visible waveform of the speech signal, the depicted energy of the speech signal, the identified word boundaries and energy detection thresholds. If the system misrecognized the word, such visualization enables to identify easier whether it resulted from wrong end-point detection. The segmentation window provides with a list of words which acoustic models to the given speech signal are the best, the scores of their likelihood and a diagram of the most likely sequence of the phoneme models aligned with the speech signal. Such visualization helps to analyze recognition errors and the impact of each phoneme model on the recognition accuracy. Results of preliminary experiments have shown that by changing the transcription of some words the recognition accuracy can be increased.

• Details
• Contents