This paper introduces a common database, an evaluation framework, and its baseline recognition results for in-car speech recognition, CENSREC-3, as an outcome of the IPSJ-SIG SLP Noisy Speech Recognition Evaluation Working Group. CENSREC-3, which is a sequel to AURORA-2J, has been designed as the evaluation framework of isolated word recognition in real car-driving environments. Speech data were collected using two microphones, a close-talking microphone and a hands-free microphone, under 16 carefully controlled driving conditions, i.e., combinations of three car speeds and six car conditions. CENSREC-3 provides six evaluation environments designed using speech data collected in these conditions.

This paper addresses the problem of voice activity detection (VAD) in noisy environments. The VAD method proposed in this paper is based on a statistical model approach, and estimates statistical models sequentially without a priori knowledge of noise. Namely, the proposed method constructs a clean speech / silence state transition model beforehand, and sequentially adapts the model to the noisy environment by using a switching Kalman filter when a signal is observed. In this paper, we carried out two evaluations. In the first, we observed that the proposed method significantly outperforms conventional methods as regards voice activity detection accuracy in simulated noise environments. Second, we evaluated the proposed method on a VAD evaluation framework, CENSREC-1-C. The evaluation results revealed that the proposed method significantly outperforms the baseline results of CENSREC-1-C as regards VAD accuracy in real environments. In addition, we confirmed that the proposed method helps to improve the accuracy of concatenated speech recognition in real environments.

This paper addresses a speech recognition problem in non-stationary noise environments: the estimation of noise sequences. To solve this problem, we present a particle filter-based sequential noise estimation method for front-end processing of speech recognition in noise. In the proposed method, a noise sequence is estimated in three stages: a sequential importance sampling step, a residual resampling step, and finally a Markov chain Monte Carlo step with Metropolis-Hastings sampling. The estimated noise sequence is used in the MMSE-based clean speech estimation. We also introduce Polyak averaging and feedback into a state transition process for particle filtering. In the evaluation results, we observed that the proposed method improves speech recognition accuracy in the results of non-stationary noise environments a noise compensation method with stationary noise assumptions.

This paper introduces an evaluation framework for Japanese noisy speech recognition named AURORA-2J. Speech recognition systems must still be improved to be robust to noisy environments, but this improvement requires development of the standard evaluation corpus and assessment technologies. Recently, the Aurora 2, 3 and 4 corpora and their evaluation scenarios have had significant impact on noisy speech recognition research. The AURORA-2J is a Japanese connected digits corpus and its evaluation scripts are designed in the same way as Aurora 2 with the help of European Telecommunications Standards Institute (ETSI) AURORA group. This paper describes the data collection, baseline scripts, and its baseline performance. We also propose a new performance analysis method that considers differences in recognition performance among speakers. This method is based on the word accuracy per speaker, revealing the degree of the individual difference of the recognition performance. We also propose categorization of modifications, applied to the original HTK baseline system, which helps in comparing the systems and in recognizing technologies that improve the performance best within the same category.

We propose a microphone array speech enhancement method that integrates spatial-cue-based source power spectral density (PSD) estimation and statistical speech model-based PSD estimation. The goal of this research was to clearly pick up target speech even in noisy environments such as crowded places, factories, and cars running at high speed. Beamforming with post-Wiener filtering is commonly used in many conventional studies on microphone-array noise reduction. For calculating a Wiener filter, speech/noise PSDs are essential, and they are estimated using spatial cues obtained from microphone observations. Assuming that the sound sources are sparse in the temporal-spatial domain, speech/noise PSDs may be estimated accurately. However, PSD estimation errors increase under circumstances beyond this assumption. In this study, we integrated speech models and PSD-estimation-in-beamspace method to correct speech/noise PSD estimation errors. The roughly estimated noise PSD was obtained frame-by-frame by analyzing spatial cues from array observations. By combining noise PSD with the statistical model of clean-speech, the relationships between the PSD of the observed signal and that of the target speech, hereafter called the observation model, could be described without pre-training. By exploiting Bayes' theorem, a Wiener filter is statistically generated from observation models. Experiments conducted to evaluate the proposed method showed that the signal-to-noise ratio and naturalness of the output speech signal were significantly better than that with conventional methods.

We present a method for recognition of acoustic events in conversation scenarios where speech usually overlaps with other acoustic events. While speech is usually considered the most informative acoustic event in a conversation scene, it does not always contain all the information. Non-speech events, such as a door knock, steps, or a keyboard typing can reveal aspects of the scene that speakers miss or avoid to mention. Moreover, being able to robustly detect these events could further support speech enhancement and recognition systems by providing useful information cues about the surrounding scenarios and noise. In acoustic event detection, state-of-the-art techniques are typically based on derived features (e.g. MFCC, or Mel-filter-banks) which have successfully parameterized the spectrogram of speech but reduce resolution and detail when we are targeting other kinds of events. In this paper, we propose a method that learns features in an unsupervised manner from high-resolution spectrogram patches (considering a patch as a certain number of consecutive frame features stacked together), and integrates within the deep neural network framework to detect and classify acoustic events. Superiority over both previous works in the field, and similar approaches based on derived features, has been assessed by statical measures and evaluation with CHIL2007 corpus, an annotated database of seminar recordings.