Task #4213
Updated by Tihelka Dan over 6 years ago
Since the experiment with syllable-based experiments with positional features, described on [[Positional_features|wiki]], show quite good results, it can be used as the baseline for ARTIC features modification. The aim is that instead of re-implementing positional features (which will *require data image changes*), we will tweak the current computation scheme (using the current set of features). In this way, we can achieve fast improvement (not perfect, though!) with low cost of coding. The original cost computation is * pos_cost = _beg(w)_ * abs(_beg(t) - beg(u)_) + _mid(w)_ * abs(_mid(t) - mid(u)_) + _end(w)_ * abs(_end(t) - end(u)_) ** _weight_ = 7 for all positions The first tweaked versions are: * pos_cost_1 = pos_cost + 150 * abs(_end(t) - end(u)_) ** _weight_ = 7 * pos_cost_2 = pos_cost + 999 * _end(u)_ * abs(_end(t) - mid(u)_) ** _weight_ = 9 ** for match on both unit and target being prosodic word transitional, the tweaked addition is set to 0 * pos_cost_3 = pos_cost + 999 * _end(u)_ * abs(_end(t) - end(u)_) ** _weight_ = 9 where: * _beg(u)_, _mid(u)_ _beg(u)_,_mid(u)_ and _end(u)_ are the position weights of the candidate unit _u_ related to the beginning. middle and end of its prosodic word * _beg(t)_, _mid(t)_ _beg(t)_,_mid(t)_ and _end(t)_ are the position weights of the target unit _t_ related to the beginning. middle and end of its prosodic word * _beg(w)_, _mid(w)_ _beg(w)_,_mid(w)_ and _end(w)_ are the corresponding weights, unit and target independent *Results:* * pos_cost_3: lowers the number of position failures from 134k to approx. 11k (*winner*). * pos_cost_2: still displays more than 70k position failures. * pos_cost_1: lowers the number of position failures from 134k to approx. 17k. *The key tak is how to further improve* pos_cost_3 *scheme*.