Automatic detection of emotions has been evaluated using standard Mel-frequency Cepstral Coefficients, MFCCs, and a variant, MFCC-low, calculated between 20 300 Hz, in order to model pitch. Also plain pitch features have used. These acoustic all modeled by Gaussian mixture models, GMMs, on the frame level. The method tested two different corpora languages; Swedish voice controlled telephone services English meetings. results indicate that GMMs level is feasible technique for emotion...

An important task in automatic conversation understanding is the inference of social structure governing participant behavior. We explore dependence between several dimensions, including assigned role, gender, and seniority, a set low-level features descriptive talkspurt deployment multiparticipant context. Experiments conducted on two large, publicly available meeting corpora suggest that our are quite useful predicting these excepting gender. The classification experiments we present...

Automatic speech understanding in natural multiparty conversation settings stands to gain from parsing not only verbal but also non-verbal vocal communicative behaviors. In this work, we study the most frequently annotated behavior, laughter, whose detection has clear implications for tasks, and automatic recognition of affect particular. To complement existing acoustic descriptions phenomenon, explore temporal patterning laughter over course conversation, with a view towards its...

This work explores the timing of very short utterances in conversations, as well effects excluding intervals adjacent to such from distributions betweenspeaker interval durations. The results show that are more precisely timed preceding utterance than longer terms a smaller variance and larger proportion no-gap-no-overlaps. Excluding furthermore measures central tendency closer zero (i.e. no-gap-no-overlaps) relatively gaps overlaps). Index Terms: Human speech production, Prosody

The taking of turns to speak is an intrinsic property conversation. It expected that models turns, providing a prior distribution over conversational form, can reduce the perplexity what attended and processed by spoken dialogue systems. We propose single-port model multi-party turn-taking which allows conversants behave independently but condition their behavior on past entire group. performs at least as well existing multi-port subsequent speech activity. quantify effect longer histories...

As spoken dialogue systems become deployed in increasingly complex domains, they face rising demands on the naturalness of interaction. We focus system responsiveness, aiming to mimic human-like flow control by predicting speaker changes as observed real human-human conversations. derive an instantaneous vector representation pitch variation and show that it is amenable standard acoustic modeling techniques. Using a small amount automatically labeled data, we train models which significantly...

The detection of laughter in conversational interaction presents an important challenge meeting understanding, primarily because is predictive the emotional state participants. We present evidence which suggests that ignoring unvoiced improves prediction involvement collocated speech, making a case for distinction between voiced and during detection. Our experiments show exclusion model training as well its explicit modeling lead to scores are much higher than those otherwise obtained all...

We propose an algorithm for segmenting multispeaker meeting audio, recorded with personal channel microphones, into speech and non-speech intervals each microphone’s wearer. An of this type turns out to be necessary prior subsequent audio processing because, in spite close-talking the channels exhibit a high degree crosstalk due unbalanced calibration small inter-speaker distance. The proposed is based on short-time crosscorrelation all pairs. It requires no training executes one fifth real...

The study of meetings, and multi-party conversation in general, is currently the focus much attention, calling for more robust accurate speech activity detection systems. We present a novel multichannel algorithm, which explicitly models overlap incurred by participants taking turns at speaking. Parameters overlapped states are estimated during decoding using combining knowledge from other observed same meeting, an unsupervised manner. demonstrate on NIST Rich Transcription Spring 2004 data...

In recent years, the field of automatic speaker identification has begun to exploit high-level sources speaker-discriminative information, in addition traditional models spectral shape. These include pronunciation models, prosodic dynamics, pitch, pause, and duration features, phone streams, conversational interaction. As part this broader thrust, we explore a new frame-level vector representation instantaneous change fundamental frequency, known as frequency variation (FFV). The FFV...

This paper describes the 2006 lecture recognition system developed at Interactive Systems Laboratories (ISL), for individual head-microphone (IHM), single distant microphone (SDM), and multiple microphones (MDM) conditions.It was evaluated in RT-06S rich transcription meeting evaluation sponsored by US National Institute of Standards Technologies (NIST).We describe principal differences between our current those submitted previous years, namely, improved acoustic language models, cross...

This paper describes the Interactive Systems Lab’s Meeting transcription system, which performs segmentation, speaker clustering as well transcriptions of conversational meeting speech. The system described here was evaluated in NIST’s RT-04S “Meeting” speech evaluation. compares performance our Broadcast News and most recent Switchboard on data both with a newly-trained recognizer. Furthermore we investigate effects automatic segmentation adaptation. Our best achieves 44.5% MDM condition

The field of speaker identification has recently seen significant advancement, but improvements have tended to be benchmarked on near-field speech, ignoring the more realistic setting far-field-instrumented speakers. In this work we present several findings far-field speech from MIXER5 Corpus, in areas feature extraction, modeling, and multichannel score combination. First, observe that minimum-variance distortionless response (MVDR) features outperform Mel-frequency cepstral coefficient...