Speech Lab @ Queens College "Reading Group" @andrewmaxwellrosenberg - Tumblr Blog

Stuttered Speech Recognition For Robotic Control

by G. Manjula, Dr. M. Shiva Kumar

This is an uninformative paper -- there's no description of how stuttering is addressed. The ASR is based on DTW template matching to a very limited (5 word) vocabulary of robot commands.

The Voice and Eye Gaze Behavior of an Imposter: Automated Interviewing and Detection for Rapid Screening at the Border

By Aaron C. Elkins, Douglas C. Derrick and Monica Gariup in EACL 2012

Pros: Fairly 'real world' scenario based around an automated agent at customs. Found that vocal quality decreases rapidly in deceptive conditions. A minor effect that subjects fixated visually on fraudulent elements of a visa document.

Cons: Very small sample set: 37 honest users, only 4 instances of deception. No incentive for successful deception. Performance is overstated: no held out data was used for training a decision tree, testing on training data revealed no misses and 2 false positives. i.e. recall = 100% but precision is only 66%. (By selecting users who were most confident of their ability to lie, they may be detecting this personality trait rather than deceptive behavior.)

Using Classifier Features for Studying the Effect of Native Language on the Choice of Written Second Language Words

by Oren Tsur and Ari Rappoport in CACLA '07

this paper uses bi-gram features to see if language learners are biased towards or away from particular words in english based on the lexico-tactics or phonotactics of their native language. they are: "the use of L2 words is strongly influenced by the frequency of L1 syllables".

In general this paper seems more like a more in-depth error analysis of Koppel 2005. It's a helpful addition to understanding these phenomena -- it investigates questions of if there is a "suffix bias" here, or a bias due to content words. In both the answer seems to be no. not significantly anyway.

(also... i'm not sure what "classifier features" are. weird title.)

Parametric modeling of intonation using vector quantization

By Gregor Mohler, Alistair Conkie in 3RD ESCA WORKSHOP ON SPEECH SYNTHESIS (1998)

I'm fairly certain I read this paper before, but promptly forgot the details. Essentially the idea is to model prominence pitch contours with two sigmoid functions abutting each other at the peak. The work goes on to generate a codebook based on the discovered parameters of the two sigmoids for synthesis modeling.

There's really no reason that I haven't coded up these "PaIntE" (Parametric Intonational Event) parameters in AuToBI. soon enough.

iVector-based Acoustic Data Selection

by Olivier Siohan and Michiel Bacchiani at Interspeech 2013

I read this because it sounded like an idea for active learning that Ali and I had been kicking around. It turns out it is pretty similar, but more geared toward selecting segments for transcription in service of training a domain specific recognizer. The task we're working on is a little different, but its encouraging to know that other people are thinking in the same direction.

Automatic Prosodic Events Detection by Using Syllable-based Acoustic, Lexical and Syntactic Features

by Chong-Jia Ni, Wen-Ju Liu and Bo Xu in Interspeech 2011

This is a paper on a topic I've worked on quite a lot, so I tend to get picky particularly around results.

There's a pitfall when working on the Boston University Radio News Corpus, that I've written about before. There are 6 speakers, but they are all reading the same data. If care isn't taken in setting up train and test sets, you can end up training and testing on the same speaker (this is a problem, but easy to fix) or training and testing on the same lexical material. The latter is a much more insidious problem. It takes particular care to make sure that you don't end up with a speaker-dependent, or even more damaging, text-dependent results. This paper suffers from both of these pitfalls. While the results are very good -- low 90%s on a lot of tasks, i'm *very* skeptical of the performance gains.

That said, the ensemble approach they use makes sense, and there are some potentially useful features described.

Modeling Biological Processes for Reading Comprehension

by Jonathan Berant, Vivek Srikumar, Pei-Chun Chen, Brad Huang and Christopher D. Manning, Abby Vander Linden and Brittany Harding at EMNLP 2014

Really interesting paper on micro-understanding and machine reading. They describe a system that can perform well on biological process multiple choice reading comprehension questions -- like on a sat subject test. There are a lot of components to this work, and it's a very nice demonstration of leveraging these to perform a very challenging task.

A Latent Variable Model for Geographic Lexical Variation

by Jacob Eisenstein Brendan O’Connor Noah A. Smith Eric P. Xing in EMNLP 2010

A nice application of latent variable modeling to model both topic and geography in twitter data. Another bonus is that the data is available for comparison. Really thorough write up and some nice analysis of results too. very satisfying paper.

UNSUPERVISED LEXICAL CLUSTERING OF SPEECH SEGMENTS USING FIXED-DIMENSIONAL ACOUSTIC EMBEDDINGS

by Herman Kamper, Aren Jansen, Simon King, and Sharon Goldwater

This is a paper I wish I had written. It's effective, fits into a clear research program, and unbelievably clear. If I could clear the deck for a couple of months, and work only along the lines laid out here, I would in a second.

I’m a Belieber: Social Roles via Self-identification and Conceptual Attributes

by Charley Beller, Rebecca Knowles, Craig Harman, Shane Bergsma, Margaret Mitchell, and Benjamin Van Durme in ACL 2014

Really fun extension of previous work on fixed broad roles like gender, looking to see what you can extract from I'm a ____ or ___'s X constructions. It turns out a fair amount.

Modeling Latent Biographic Attributes in Conversational Genres

by Nikesh Garera and David Yarowsky in ACL 2009

This paper classifies gender, age and native language using lexical features from speech transcripts (switchboard and fisher). The approach is evaluated on arabic and email data (enron). One of the more interesting findings is the use of the partner's gender etc. in classification based on the finding that classifying same-gender conversation partners was easier than mixed-gender (by kind of a lot). This is a potentially important thing to recognize as our group looks at classifying demography

Streaming Analysis of Discourse Participants

by Ben Van Durme in EMNLP 2012

has two nice extensions: 1. extends static user classification to a streaming context, essentially through a relatively clever online accumulation and normalization of features and counts. 2. applies "Reservoir Counting" and "Reservoir Averaging" to this count accumulation problem to keep the storage very small. I still need to read about the approximate part of Reservoir Computing, but it's an interesting approach.

Automatic Recognition of Personality in Conversation

by François Mairesse and Marilyn Walker in NAACL 2006

Fairly straight paper on recognizing personality based on speech and transcripts. Typical lexical features (LIWC, etc.) and classification, but the data is interesting. One flaw is (possibly) that the personality measures are based on external observations rather than personality questionnaires of the speakers.

ICASSP Preprint -- title pending to protect the innocent

By friendly friends who share their submissions.

This is a clever paper about [an ICASSP relevant topic]. I think the abstract says it all.

[Important Task] systems require a mapping from [X], or [Y], into [Z] to enable [an important subtask], and to model [something challenging]. The standard approach to achieve this is to [do something that most people know about]. [Standard] models are then constructed using [a standard technique]. [Sometimes], it may not be practical to [do this]. An alternative approach is to use [a different approach]. This paper proposes a simple approach for [doing this].

... and the friendly authors deliver on the promise of the abstract.

DETECTING DELETIONS IN ASR OUTPUT

By M.S. Seigel and P.C. Woodland in ICASSP 2014

The remarkable bit of this paper is recognizing that estimations of ASR confidence ignore the likelihood of a deletion. The approach used to address this (a CRF with lattice based posterior features of words and subwords) is less wow-ing. The Deletion informed normalized cross entropy (DNCE) measure is pretty cool too.

Modulation Spectrum in Speech Processing

By Hynek Hermansky 1998

I probably should have known about the modulation spectrum earlier. I had seen some of the results reported in this (slightly old) survey before, but I hadn't tied them together as a common research thread. It was a talk during a speech synthesis session that it was brought closer to my attention.

The modulation spectrum is defined as the temporal dynamics of spectral components. Namely, the change over time (in frequency units) of spectral energy.

Happy to fill in some knowledge gaps and tie some previously known results together.

Defectors Cannot Be Detected during ‘‘Small Talk’’ with Strangers

By Joseph H. Manson*, Matthew M. Gervais, Michelle A. Kline in PLOS|One 2013

Continued analysis of the results of the previous paper on psychopathy. Here subjects are asked to guess the behaviors of others in a prisoner's dilemma game. Conversational participants are pretty bad at this. Naive eavesdroppers are pretty good -- and better than informed eavesdroppers.

Methodologically, I think the paper's use of the true rate of cooperation (~60%) as the chance level of prediction may be an overly challenging baseline. I'm not convinced that eavesdroppers (or participants or people in general) know that there is a prosocial bias (likelihood to cooperate) in these one-shot prisoner's dilemmas games. I would think that their prior might be closer to 0.5 rather than the true prior of ~0.6. You'd have to get a little clever to suss out what folks thought of an unknown participant's prisoner's dilemma behavior to really sort this out.

Trending Blogs

Recently Viewed Blogs

Speech Lab @ Queens College "Reading Group"