Discriminant-function-based minimum recognition error rate pattern-recognition approach to speech recognition

Integrating Geometric Context for Text Alignment of Handwritten Chinese Documents

“…String-level training using the minimum classification error (MCE) criterion has shown superior performance in speech recognition and handwriting recognition [15] [16].…”

Minimization of Utterance Verification Error Rate as a Constrained Optimization Problem

“…In speech recognition, discriminative training using the minimum classification error (MCE) [4], [9] or the maximum mutual information (MMI) [10] criterion has shown significant improvement over ML training of acoustic model. The idea of MCE training later led to MVE discriminative training [5], [6] of acoustic models for speech or speaker verification.…”

“…can be optimized by adjusting the model parameters via the GPD algorithm [4]. This approach was first proposed in [6] with τ set to zero to simplify the formulation.…”

“…model means, variances, etc. can be found in [4] and will not be repeated here. Theoretically speaking, the threshold τ and the Lagrange multiplier β can be learned as well, using the approach as described above.…”

“…However, since the amount of training data is never unlimited in practice, and the true parametric form of these acoustic models is never known, it can be argued that these ML-estimated models may not give the best verification performance. Recently, discriminative training [4] that optimizes model parameters to achieve the minimum verification error (MVE) [3], [5], [6] or minimum verification error rate (MVER) [7] has shown to give better verification performance.…”