Reishi Kondo

Speech synthesis apparatus

The invention provides a speech synthesis apparatus which can produce synthetic speech of a high quality with reduced distortion. To this end, upon production of synthetic speech based on prosodic information and phonological unit information, the prosodic information is modified using the phonological unit information, and duration length information and pitch pattern information of phonological units of the prosodic information and the phonological unit information are modified with each other. The speech synthesis apparatus includes a prosodic pattern production section for receiving utterance contents as an input thereto and producing a prosodic pattern, a phonological unit selection section for selecting phonological units based on the prosodic pattern, a prosody modification control section for searching the phonological unit information selected by the phonological unit selection section for a location for which modification to the prosodic pattern is required and outputting information of the location for the modification and contents of the modification, a prosody modification section for modifying the prosodic pattern based on the information of the location for the modification and the contents of the modification outputted from the prosody modification control section, and a waveform production section for producing synthetic speech based on the phonological unit information and the prosodic information modified by the prosody modification section using a phonological unit database.

Acoustic event detection based on non-negative matrix factorization with mixtures of local dictionaries and activation aggregation

This paper proposes a new non-negative matrix factorization (NMF) based acoustic event detection (AED) method with mixtures of local dictionaries (MLD) and activation aggregation. One of the key problems of conventional NMF-based methods is instability of activations due to redundancy of a region spanned by the bases of dictionaries. Sounds inside the redundant region are often decomposed into undesired combinations of bases and activations that cause failure of detection. The proposed method employs MLD for allocating sub-groups of basis dictionaries to acoustic elements to minimize redundancy in the region and obtain controlled activations. In order to make activations more stable, the proposed method also introduces activation aggregation which combines basis-wise activations into acoustic-element-wise activations. Much more stable activations by the proposed method lead to significant improvement in F-measure by up to 60% compared to an ordinary convolutive-NMF-based method. The proposed method also outperforms a latest alternative which is not based on NMF.

Method for synthesizing a voice waveform which includes compressing voice-element data in a fixed length scheme and expanding compressed voice-element data of voice data sections

A method for synthesizing a voice waveform includes compressing voice-element data in a fixed length scheme that uses data from a preceding or succeeding frame. The compressed voice-element data of each voice section is expanded, and the preceding or succeeding frame of the expanded voice-element data is discarded. The remaining voice-element data is synthesized after discarding portions of the expanded voice-element data.

Detection of anomaly acoustic scenes based on a temporal dissimilarity model

This paper proposes detection of anomaly acoustic scenes based on a temporal dissimilarity model. The periodicity in the temporal variation of acoustic scenes is first pointed out and then used to build a new stochastic model. In the new model, the temporal variation is expressed by dissimilarity between current and previous acoustic scenes. Anomaly acoustic scenes are detected based on the 24-hour periodic dissimilarity model. Evaluation results using 40-day (1000-hour) data show that the proposed method can detect unknown anomaly acoustic scenes with 82.3% F-measure in 0 dB signal-to-noise-ratio conditions.

TDOA estimation based on phase-voting cross correlation and circular standard deviation

This paper proposes a new TDOA estimation based on phase-voting cross correlation and circular standard deviation. Based on phase delay and kernel function, the proposed method generates a probability density function (PDF) of TDOA for each frequency bin. TDOA estimate is determined by voting the PDFs generated for all frequency bins. Peak positions of the bin-wise PDFs for the target signal are concentrated only at the target time difference because peak positions for the noise totally differ among bins and periodicity of peaks depends on frequency. Therefore, by voting the PDFs for all frequency bins, the peak position for the target can be easily identified. The kernel width of PDF is determined by circular standard deviation of cross spectral phase for each frequency bin. This width control enhances peaks of PDFs for high SNR frequency bins since phases for high SNR bins are more stable than those for low ones. Evaluation with ship and drone sounds shows that the RMSE of TDOA estimation by the proposed method reaches 0.37 times that by GCC-PHAT.