The CUED Speech Group Dr Mark Gales Machine Intelligence Laboratory Cambridge University Engineering Department
Computational and Biological Learning Lab 1. CUED Organisation 130 1100 450 Academic Staff Undergrads Postgrads CUED: 6 Divisions A. ThermoFluids B. Electrical Eng C. Mechanics D. Structures E. Management Control Lab Signal Processing Lab Computational and Biological Learning Lab Machine Intelligence Lab F. Information Engineering Division Speech Group Vision Medical Imaging 4 Staff Bill Byrne Mark Gales Phil Woodland Steve Young 9 RA’s 12 PhD’s 2
2. Speech Group Overview Funded Projects in Recognition/Translation/Synthesis (5-10 RAs) MPhil in Computer Speech, Text and Internet Technology Computer Laboratory NLIP Group PhD Projects in Fundamental Speech Technology Development (10-15 students) Computer Speech and Language HTK Software Tools Development International Community Primary research interests in speech processing 4 members of Academic Staff 9 Research Assistants/Associates 12 PhD students 3
Principal Staff and Research Interests Dr Bill Byrne Statistical machine translation Automatic speech recognition Cross-lingual adaptation and synthesis Dr Mark Gales Large vocabulary speech recognition Speaker and environment adaptation Kernel methods for speech processing Professor Phil Woodland Large vocabulary speech recognition/meta-data extraction Information retrieval from audio ASR and SMT integration Professor Steve Young Statistical dialogue modelling Voice conversion 4
Synthesis Translation Dialogue Recognition Machine Learning Research Interests data driven techniques voice transformation HMM-based techniques Synthesis statistical machine translation finite state transducer framework Translation data driven semantic processing statistical modelling Dialogue large vocabulary systems [Eng, Chinese, Arabic ] acoustic model training and adaptation language model training and adaptation rich text transcription & spoken document retrieval Recognition fundamental theory of statistical modelling and pattern processing Machine Learning 5
Example Current and Recent Projects Global Autonomous Language Exploitation DARPA GALE funded (collab with BBN, LIMSI, ISI …) HTK Rich Audio Trancription Project (finished 2004) DARPA EARS funded CLASSIC: Computational Learning in Adaptive Systems for Spoken Conversation EU (collab with Edinburgh, France Telecom,,…) EMIME: Effective Multilingual Interaction in Mobile Environments EU (collab with Edinburgh, IDIAP, Nagoya Institute of Technology … ) R2EAP: Rapid and Reliable Environment Aware Processing TREL funded Also active collaborations with IBM, Google, Microsoft, … 6
3. Rich Audio Transcription Project New algorithms Rich Transcript Natural Speech English/Mandarin DARPA-funded project Effective Affordable Reusable Speech-to-text (EARS) program Transform natural speech into human readable form Need to add meta-data to the ASR output For example speaker-terms/handle disfluencies http://mi.eng.cam.ac.uk/research/projects/EARS/index.html See 7
Rich Text Transcription ASR Output okay carl uh do you exercise yeah actually um i belong to a gym down here gold’s gym and uh i try to exercise five days a week um and now and then i’ll i’ll get it interrupted by work or just full of crazy hours you know Meta-Data Extraction (MDE) Markup Speaker1: / okay carl {F uh} do you exercise / Speaker2: / {DM yeah actually} {F um} i belong to a gym down here / / gold’s gym / / and {F uh} i try to exercise five days a week {F um} / / and now and then [REP i’ll + i’ll] get it interrupted by work or just full of crazy hours {DM you know } / Final Text Speaker1: Okay Carl do you exercise? Speaker2: I belong to a gym down here, Gold’s Gym, and I try to exercise five days a week and now and then I’ll get it interrupted by work or just full of crazy hours. 8
4. Statistical Machine Translation Aim is to translate from one language to another For example translate text from Chinese to English Process involves collecting parallel (bitext) corpora Align at document/sentence/word level Use statistical approaches to obtain most probable translation 9
GALE: Integrated ASR and SMT Member of the AGILE team (lead by BBN) The DARPA Global Autonomous Language Exploitation (GALE) program has the aim of developing speech and language processing technologies to recognise, analyse, and translate speech and text into readable English. Primary languages for STT/SMT: Chinese and Arabic http://mi.eng.cam.ac.uk/research/projects/AGILE/index.html See 10
5. Statistical Dialogue Modelling Speech Understanding Generation Dialogue Manager System Waveforms Words/Concepts Dialogue Acts Use a statistical framework for all stages 11
1-Best Signal Selection CLASSiC: Project Architecture st Speech Input ASR NLU DM NLG TTS Context t-1 ut ht at wt rt 1-Best Signal Selection x Speech output Legend: ASR: Automatic Speech recognition NLU: Natural Language Understanding DM: Dialogue Management NLG: Natural Language Generation TTS: Text To Speech st: Input Sound Signal ut: Utterance Hypotheses ht: Conceptual Interpretation Hypotheses at: Action Hypotheses wt: Word String Hypotheses rt: Speech Synthesis Hypotheses X: possible elimination of hypotheses http://classic-project.org See
6. EMIME: Speech-to-Speech Translation Personalised speech-to-speech translation Learn characteristics of a users speech Reproduce users speech in synthesis Cross-lingual capability Map speaker characteristics across languages Unified approach for recognition and synthesis Common statistical model; hidden Markov models Simplifies adaptation (common to both synthesis and recognition) Improve understanding of recognition/synthesis http://emime.org See 13
7. R2EAP: Robust Speech Recognition Current ASR performance degrades with changing noise Major limitation on deploying speech recognition systems 14
Project Overview Aims of the project To develop techniques that allow ASR system to rapidly respond to changing acoustic conditions; While maintaining high levels of recognition accuracy over a wide range of conditions; And be flexible so they are applicable to a wide range of tasks and computational requirements. Project started in January 2008 – 3 year duration Close collaboration with TREL Cambridge Lab. Common development code-base – extended HTK Common evaluation sets Builds on current (and previous) PhD studentships Monthly joint meetings http://mi.eng.cam.ac.uk/~mjfg/REAP/index.html See 15
Approach – Model Compensation Model compensation schemes highly effective BUT Slow compared to feature compensation scheme Need schemes to improve speed while maintaining performance Also automatically detect/track changing noise conditions 16
8. Toshiba-CUED PhD Collaborations To date 5 Research studentships (partly) funded by Toshiba Shared software - code transfer both directions Shared data sets - both (emotional) synthesis and ASR 6 monthly reports and review meetings Students and topics Hank Liao (2003-2007): Uncertainty decoding for Noise Robust ASR Catherine Breslin (2004-2008): Complementary System Generation and Combination Zeynep Inanoglu (2004-2008): Recognition and Synthesis of Emotion Rogier van Dalen (2007-2010): Noise Robust ASR Stuart Moore (2007-2010): Number Sense Disambiguation Very useful and successful collaboration 17
9. HTK Version 3.0 Development HTK is a free software toolkit for developing HMM-based systems 1000’s of users worldwide widely used for research by universities and industry 1989 – 1992 1993 – 1999 2000 – date V1.0 – 1.4 V1.5 – 2.3 V3.0 – V3.4 Initial development at CUED Commercial development by Entropic Academic development at CUED Development partly funded by Microsoft and DARPA EARS Project Primary dissemination route for CU research output 2004 - date: the ATK Real-time HTK-based recognition system http://htk.eng.cam.ac.uk See 18
10. Summary Speech Group works on many aspects of speech processing Large vocabulary speech recognition Statistical machine translation Statistical dialogue systems Speech synthesis and voice conversion Statistical machine learning approach to all applications World-wide reputation for research CUED systems have defined state-of-the-art for the past decade Developed a number of techniques widely used by industry Hidden Markov Model Toolkit (HTK) Freely-available software, 1000’s of users worldwide State-of-the –art features (discriminative training, adaptation …) HMM Synthesis extension (HTS) from Nagoya Institute of Technology http://mi.eng.cam.ac.uk/research/speech See 19