LABORATORIO DE COMUNICACIÓN ORAL
ROBERT WAYNE NEWCOMB
Díaz, F., Rubio, M., Gómez, P., Nieto, V., Rodellar, V.
- "Using Hidden Markov Models in segmentation of speaker-independent connected-digit corpus"
- Proc. of the WSEAS 2002 Multiconference (ICOSSIP’02) (ISBN: 960-8052-68-8)
- Skiathos-Greece, September 25-28, 2002.
The first task to be accomplished in speech recognition is the segmentation and labeling of records.
Regarding speech, this is a very complicated and costly procedure, although of most importance because at the present
time many available speech corpora are not segmented. This paper proposes a semi-automatic segmentation
method in order to reduce the manual segmentation burden of a very large corpus. First, Hidden Markov
Models are created with a reduced set of records. Afterwards they are used to perform an automatic
segmentations on the rest. Recursively, new more roust models are created and used to create new segmentations.
The method consists in three main steps: (1) Initial Reduced Segmentation, (2) Recursive-Extended
Segmentation and (3) Post-processing of the lebels. This method was evaluated in the segmentation of the TIDIGITS
corpus with two independent initial manual segmentations. Finally the method was able to label correctly 96.18%
and 95.72% od the corpus records, respectively,
Pulse aquí para bajarse el artículo
Formato PDF (1.196 Kb.)
Comprimido en zip (167 Kb.)
Madrid a 17 de junio de 2004