The ACLP research team has been awarded a number of research grants and has been involved in several projects with leading companies in the industry, in addition to initiating self-funded projects.
Magneton 2013-2015Funded by the Israeli Ministry of Commerce Chief Scientist as part of a Magneton project encouraging the transfer of technology from academia to the industry. The research is being carried out in collaboration with NICE Systems.Intent Insight ImpactThe amount of recorded data collected by intelligence systems is overwhelming, making it impossible for human examiners to handle alone. This big data problem is often overcome by utilizing keyword spotting systems that are capable of pinpointing pre-designated terms among thousands of hours of recordings. This project focuses on combining LVCSR and Phonetic Search KWS engines in order to maximize keyword spotting capabilities and provide keyword flexibility.
Funded by the Israeli Ministry of Commerce Chief Scientist as part of a Magneton project encouraging the transfer of technology from academia to the industry. The research is being carried out in collaboration with NICE Systems.Intent Insight ImpactSpeaker diarization estimates the number of speakers in a conversation and produces a time-stamped conversational "diary" of participating speakers and is becoming an increasingly important component of speech and speaker recognition technologies. The focus of this research is on minimizing speaker diarization errors in commercial applications used by live call-centers and trading floor arenas.
Funded by Maf'at (The Israeli Ministry of Defense research administration for development of arms and technological infrastructure).2013-2014: Speech Driven Question AnsweringAdapting the Kaldi speech recognition engine to domain specific QA system with a focus on acoustical modeling and discriminative training.2012-2013: Incorporation of Textual Entailment
Incorporation of textual entailment algorithm as an extension of a textual distance measure.
2011-2012: Domain Specific Speech RecognitionAdapting the HTK speech recognition engine to domain specific QA system while improving the textual similarity measure for locating the response to a question or information request posed by the user.
2010-2011: Textual Distance MeasureProof-of-concept for speech-driven question answering in a specific domain. The focus was on developing the framework for a Question-Answering (QA) system that integrates speech recognition, text-based analysis and text to speech.
TeleMessage 2013- 2014The main high-level objective of the project is to develop a robust algorithm for training language profiles and identifying the language of short text messages. The Language Identification (LI) algorithm will be able to deal with a finite set of 4 a-priori known languages, but will enable support of additional languages by training profiles for each new language and adding them to the LI engine.
The objective of the Keyword Spotting project was to develop a phoneme decoder and a phonetic search keyword spotting engine that locates designated keywords within a sequence of phonemes representing a speech signal.
The objective of the Speaker Verification project was to develop a text and language-independent speaker verification engine including a speaker training engine.
The focus of the project was to provide the company with the information needed to serve as a basis for the definition, assessment, integration and construction process of a voice-driven demo.
All evaluation participants are provided the same data from a “surprise” language and are allotted an initial development phase. Each research center is allowed to use whatever KWS technologies and methods are available to them, but no human intervention is allowed. Evaluations are submitted according to strict guidelines and scored by NIST using a Term Weighted Value (TWV) function.
The 2013 surprise language was Vietnamese, a South-East Asian tonal language with a vocabulary that consists only of monosyllabic words. The ACLP approach to the OpenKWS13 evaluation was to perform phonetic search KWS, using a maximum TWV based thresholding mechanism to improve the resulting keyword detections.
The 2014 surprise language was Tamil, a morphologically rich Indian language spoken mainly in Southern India and Sri Lanka. The ACLP approach to the OpenKWS2014 evaluation was to perform KWS on multiple LVCSR and phonetic-search based engines while applying maximum TWV based thresholding and score normalization mechanisms which enabled merging the KWS results of the various engines.
ACLP Hebrew Speech Corpus is a speech database that is being collected by the researchers from the Afeka Center for Language Processing.
The corpus will contain a large number of recordings of phonetically rich sentences, as well as, spontaneous speech. The main goal of the project is to model Modern Hebrew phonemes for Automatic Speech Recognition (ASR) and Text-to-Speech (TTS). The corpus includes a large lexicon with phonetic transcriptions.
2012The main purpose of the project was to evaluate both commercial and open-source speech recognition engines according to criteria that will enable the company to assess the potential contribution of their technology to speech recognition results, and to determine what data is needed from the engine in order optimize this contribution.
Funded by the Israeli Ministry of Commerce Chief Scientist as part of a Magneton project encouraging the transfer of technology from academia to the industry. The research is being carried out in collaboration with NICE Systems.
One of the more efficient methods for analyzing recorded dialogue is by detecting specific keywords within the speech (Keywords Spotting). This is done using automated systems that are able to scan large quantities of data quickly and efficiently and then recognize when any given word occurs. To date, such systems have been developed mainly for more common languages, since compiling the linguistic resources needed to develop such a system is both time and resource consuming. Keyword spotting solutions for exotic languages with high security interest rarely exist, if at all. The goal of this research is to develop a keyword spotting methodology that utilizes existing acoustic models and other language resources from a source language for detecting keywords in speech in an under-resourced target language, without training new acoustic models, etc.
This project was carried out with VocalZoom, developers of a unique optoelectronic microphone.
For this project, the ACLP designed an American English speech database to be compiled for testing a laser optic microphone and its influence on Automatic Speech Recognition (ASR) performance. The database was designed to test improvement of ASR performance in various noisy environments such as inside running and driving vehicles and in public locations. Environments such as home and office were also included to verify that there are no adverse effects in quieter surroundings. During the project, the algorithms and the speech recognition engine were tuned to work in the new environments.
This project was carried out with Athena Security Implementations Ltd., a security solutions company.
This project entailed the design and construction of a demo and marketing tool for KeyWord Spotting (KWS) in several languages. The ACLP designed the flow of the demo in cooperation with the company. For each language, a list of keywords and sentences containing the keywords were composed; evaluation and tuning databases of native speakers were collected; and tuning experiments were conducted.
Funded by the Israeli Ministry of Commerce Chief Scientist as part of a Magneton project encouraging the transfer of technology from academia to the industry. The research was carried out in collaboration with SpeechModules Ltd.
The goal of the research was to efficiently automatically transcribe voicemail messages using a speech recognition engine and a very large lexicon of 100K words, and then to forward the textual results via SMS to the user. The results of the research showed that the computational complexity of a speech recognition engine used for transcribing spontaneous speech can be significantly reduced to accommodate real-time processing by using various methods for limiting the search space, while at the same time maintaining recognition performance.
Speech Modules 2009
The project was carried out in collaboration with SpeechModules Ltd, a company specializing in advanced speech recognition technologies.
In this project the performance of a Large Vocabulary Speech Recognition Engine (LVCSR) was tested at various working points in order to generate a sensitivity model to phoneme recognition. The LVCSR engine was a three-stage engine used by SpeechModules (phonemes, words, sentences) and the work agenda was to determine and tune the phoneme recognition engine to the optimal working point with regards to overall LVCSR performance.
The ACLP provides research, development and consulting services for industry clients that are either already in or are looking to enter the Speech Processing arena.