Scott Otterson

Graceful degradation of speech recognition performance over packet-erasure networks

This paper explores packet loss recovery for automatic speech recognition (ASR) in spoken dialog systems, assuming an architecture in which a lightweight client communicates with a remote ASR server. Speech is transmitted with source and channel codes optimized for the ASR application, i.e., to minimize word error rate. Unequal amounts of forward error correction, depending on the datas effect on ASR performance, are assigned to protect against packet loss. Experiments with simulated packet loss in a range of loss conditions are conducted on the DARPA Communicator (air travel information) task. Results show that the approach provides robust ASR performance which degrades gracefully as packet loss rates increase. Transmitting at 5.2 Kbps with up to 200 ms added delay, leads to only a 7% relative degradation in word error rate even under extremely adverse network conditions.

Efficient use of overlap information in speaker diarization

Speaker overlap in meetings is thought to be a significant contributor to error in speaker diarization, but it is not clear if overlaps are problematic for speaker clustering and/or if errors could be addressed by assigning multiple labels in overlap regions. In this paper, we look at these issues experimentally, assuming perfect detection of overlaps, to assess the relative importance of these problems and the potential impact of overlap detection. With our best features, we find that detecting overlaps could potentially improve diarization accuracy by 15% relative, using a simple strategy of assigning speaker labels in overlap regions according to the labels of the neighboring segments. In addition, the use of cross-correlation features with MFCCs reduces the performance gap due to overlaps, so that there is little gain from removing overlapped regions before clustering.

The 2004 ICSI-SRI-UW meeting recognition system

The paper describes our system devised for recognizing speech in meetings, which was an entry in the NIST Spring 2004 Meeting Recognition Evaluation. This system was developed as a collaborative effort between ICSI, SRI, and UW and was based on SRIs 5xRT Conversational Telephone Speech (CTS) recognizer. The CTS system was adapted to the Meetings domain by adapting the CTS acoustic and language models to the Meeting domain, adding noise reduction and delay-sum array processing for far-field recognition, and adding postprocessing for cross-talk suppression for close-talking microphones. A modified MAP adaptation procedure was developed to make best use of discriminatively trained (MMIE) prior models. These meeting-specific changes yielded an overall 9% and 22% relative improvement as compared to the original CTS system, and 16% and 29% relative improvement as compared to our 2002 Meeting Evaluation system, for the individual-headset and multiple-distant microphones conditions, respectively.