Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

A Common Phone Model Representation for Speech Recognition and Synthesis

Mats Blomberg

Dept. of Speech Communication and Music Acoustics, KTH, Stockholm, Sweden

A combined representation of context-dependent phones at the production parametric and the spectral level is described. The phones are trained in the production domain using analysis-by-synthesis and piece-wise linear approximation of parameter trajectories. For recognition, this representation is transformed to spectral subphones, using a cascade formant synthesis procedure. In a connected-digit recognition task, 99.1% average correct digit rate was achieved in a group of seven male speakers when, for each test speaker, training was done on the other six speakers. Simple rules for male-to-female transformation of the male phone library increased the performance for six female speakers from 88.9% without transformation to 96.3%. In informal listening tests of resynthesised digit strings, the speech has been judged as intelligible, however far from natural.

Full Paper

Bibliographic reference.  Blomberg, Mats (1994): "A common phone model representation for speech recognition and synthesis", In ICSLP-1994, 1875-1878.