The 1998 HTK system for transcription of conversational telephone speech

Abstract

This paper describes the 1998 HTK large vocabulary speech recognition system for conversational telephone speech as used in the NIST 1998 HubSE evaluation. Front-end and language modelling experiments conducted using various training and test sets from both the Switchboard and Callhome English corpora are presented. Our complete system includes reduced bandwidth analysis, sidebased cepstral feature normalisation, vocal tract length normalisation (VTLN), triphone and quinphone hidden Markov models (HMMs) built using speaker adaptive training (SAT), maximum likelihood linear regression (MLLR) speaker adaptation and a confidence score based system combination. A detailed description of the complete system together with experimental results for each stage of our multi-pass decoding scheme is presented. The word error rate obtained is almost 20% better than our 1997 system on the development set.

Acoustic-phonetic speech parameters for speaker-independent speech recognition

“…The most popular VTLN technique performs speaker-specific piecewise linear frequency scaling of the Mel-Frequency Cepstral Coefficients (MFCCs) [3]. The overall improvement in the word error rate (WER) obtained with this technique is usually on the order of 0.6% as compared to results obtained without VTLN.…”

The Development of the AMI System for the Transcription of Speech in Meetings

“…Acoustic models are phonetic decision tree state clustered triphone models with standard left-to-right 3-state topology. They were obtained using standard HTKmaximum likelihood training procedures (see for example [11]). The system uses approximately 7000 states where each state is represented as a mixture of 16 Gaussians.…”

“…Speaker adaptive training is performed in the form of vocal tract length normalisation (VTLN) both in training and test. Warp factors are estimated using a parabolic search procedure, a piecewise linear warping function and a maximum likelihood criterion [11]. Speaker adaptation is perfermed using maximum likelihood linear regression (MLLR) of the means and variances [8].…”

“…Hence models are constructed from other sources and interpolated (as in e.g. [11]). This is true for both CTS and meeting data.…”

The synergy between bounded-distance HMM and spectral subtraction for robust speech recognition

“…Within this category we find techniques such as RASTA-PLP (Hermansky and Morgan (1994)), CMN (Cepstral Mean Normalisation) (Furui (1981)), SCMN (Segmental Cepstral Mean Normalisation) (Viikki and Laurila (1998)), VTLN (Vocal Tract Length Normalisation) (Hain et al (1999)) or histogram equalization (de la Torre et al (2005)). …”