Search

HMM recognizers using either a single Gaussian or a Gaussian mixture per state have been shown to work fairly well for 1000-word vocabulary continuous speech recognition. However, the large number of Gaussians required to cover the entire English language makes these systems unwieldy for large vocabulary tasks. Tied mixtures offer a more compact way of representing the observation pdf's. We have converted our independent mixture systems to tied mixtures and have obtained mixed results: a 13% improvement in speaker-dependent recognition without cross-word triphone models, but no improvement in our speaker-dependent system with cross-word boundary triphone models or in our speaker-independent system. There is also a reduction in CPU requirements during recognition--but this is counter-balanced by an increase during training. This paper also includes a comment on the validity of the DARPA program's evaluation test system comparisons.

A design study has been carried out for an easy-to-program high-speed signal processor. The machine achieves a throughput of about 25 million arithmetic operations per second by incorporating parallel addressing and data access into a general-purpose architecture. The study indicates that, for typical signal processing applications, nearly all of the instruction cycles can be used to perform primary arithmetic operations rather than data access or addressing. The proposed machine would require about 1165 ECL 100K chips, occupy 2-3 cubic feet, and consume about 700 watts.

This note describes a homomorphic pitch detector which yields good performance on clear speech and moderate robustness to additive broadband noise, narrowband noise, and to degradation by a telephone simulator. It achieves the performance by the use of an adaptive time window, log-spectral windowing, an adaptive voicing threshold, and pitch track smoothing. It has been implemented in a real-time LPC vocoder for testing. Finally, a pilot study of a preprocessor to improve the performance of any coherence seeking pitch detector is presented.

The following is a description of a low data rate voice terminal. The unit uses an LPC-10 algorithm which provides a 2400 bps NATO standard compatible mode and a 1200 bps frame-fill mode. The unit has an RS-449 digital interface with MIL-STD-188-114 levels. The user operates the unit from a telephone instrument with a PTT handset. The user interface models an augmented combination of an FM transceiver and a telephone to inform the user of the state of the data communications equipment in a natural manner. These protocols cover the full and half duplex channels. The unit exhibits a DRT (intelligibility) score of 91.1% at 2400 bps and 88.7% at 1200 bps in benign acoustic environments. (Author)