HIWIRE MEETING Trento, January 11-12, 2007 José C. Segura, Javier Ramírez
2 HIWIRE Meeting – Trento, January, 2007 Schedule PEQ HAFE IS07 setup New improvements in robust VAD Revised multiple observation LRT (MO-LRT) Improve noise reduction and frame-dropping
3 HIWIRE Meeting – Trento, January, 2007 PEQ Evaluation AURORA2, AURORA3, AURORA4 Compared to HEQ PEQ shows better performance on all databases Results using Loquendo recognizer Improved results Slight degradation on clean conditions
4 HIWIRE Meeting – Trento, January, 2007 PEQ / HEQ comparative results
5 HIWIRE Meeting – Trento, January, 2007 HAFE In collaboration with TUC-NTUA Released two C modules, integrated in HAFE V1.0 Basic Analysis VAD (LTSD) Wiener filter (optional) Output: WAV / MFCC / FB Post-Processing PEQ (optional) Regression computation (optional) Frame-Dropping (optional) CMS /CMVN (optional)
6 HIWIRE Meeting – Trento, January, 2007 IS07 setup Prepared an HTK setup for evaluation on the HIWIRE database Training scripts based on LORIA ones Test scripts include MLLR adaptation with variable number of utterances Baseline results Only for clean data With and without adaptation
7 HIWIRE Meeting – Trento, January, 2007 IS07 setup (without adaptation)
8 HIWIRE Meeting – Trento, January, 2007 IS07 (with adaptation)
9 HIWIRE Meeting – Trento, January, 2007 A review of MO-LRT VAD Multiple observation likelihood ratio test: Given 2N+1 independent observations of the noisy speech Hypothesis test: G 0 : All the observations in the buffer are non-speech G 1 :“““noisy speech Gaussian model: where
10 HIWIRE Meeting – Trento, January, 2007 Hangover analysis
11 HIWIRE Meeting – Trento, January, 2007 Hangover analysis
12 HIWIRE Meeting – Trento, January, 2007 Revised MO-LRT Given 2N+1 independent observations of the noisy speech: All the possible hypothesis on the individual observations: h k = 0 :x k = n h k = 1 :x k = s + n Hypothesis subsets
13 HIWIRE Meeting – Trento, January, 2007 Revised MO-LRT We assume that just a single speech to non-speech or non- speech to speech transition can occur in h
14 HIWIRE Meeting – Trento, January, 2007 Compared to Sohn et al. VAD.
15 HIWIRE Meeting – Trento, January, 2007
16 HIWIRE Meeting – Trento, January, 2007 ROC curves in quiet noise conditions (stopped car and engine running) and close talking microphone.
17 HIWIRE Meeting – Trento, January, 2007 ROC curves in high noise conditions (high speed over a good road) and distant talking microphone.
18 HIWIRE Meeting – Trento, January, 2007 Presented at ICASSP 2007: Javier Ramirez, José C. Segura, Juan M. Górriz, “Revised contextual LRT for voice activity detection”, ICASSP Under review: Javier Ramírez, José C. Segura, Juan M. Górriz and Luz García, “Improved Voice Activity Detection Using Contextual Multiple Hypothesis Testing for Robust Speech Recognition”, IEEE Transactions on Audio, Speech and Language Processing.