Phone duration models for fast broadcast news transcriptions

Phone duration modeling in HMM based LVCSR systems has been largely studied during last few years. In this paper, we address the problem of duration modeling in the particular context of fast decoding on LVCSR task. Discrete distributions are integrated in the LIA's broadcast news transcription system, and influence of duration modeling is studied using various pruning schemes. Experimental results show that duration modeling improve significantly the pruning efficiency. In a second time, we show that durations are intrinsecly acoustic-context dependent. Crossed experiments are conducted combining context independent acoustic models and context dependent duration models. We show how durations are affected by acoustic context. At last, we propose a rate dependent modeling of phone durations. This method outperforms significantly our rate independent duration model based system. Globally, integration of rate dependent models allows a absolute WER gain of 3.3%.