Abstract: A non-intrusive objective speech quality assessment is performed on a degraded speech signal. The methods are well suited for systems where random and bursty packet losses may occur and/or packet stream regeneration may also occur prior to speech signal quality assessment. In one embodiment received packetized speech is analyzed to determine to an overall final signal quality score. A limited set of trained neural networks, e.g., 5, corresponding to different signal features, each determine a signal feature quality score. A trained joint quality score determination module determines a joint quality score based on the signal feature quality scores. Packet loss is estimated based on received packet header information and/or detected gap durations. The determined joint quality score is adjusted, based on estimated packet loss information obtained from examining the speech signal, network level statistics and/or codec parameters to generate the final quality score.

Abstract: Mechanisms are provided for performing context based synonym filtering for natural language processing. Content is parsed into one or more conceptual units, wherein each conceptual unit comprises a portion of text of the content that is associated with a single concept. For each conceptual unit, a term in the conceptual unit is identified that has a synonym to be utilized during natural language processing of the content. A first measure of relatedness of the term to at least one other term in the conceptual unit is determined. A second measure of relatedness of the synonym of the term to the at least one other term in the conceptual unit is determined. A determination whether or not to utilize the synonym when performing natural language processing on the conceptual unit is made based on the first and second measures of relatedness and natural language processing on the content is performed accordingly.

Abstract: A predictive pattern high-frequency reconstruction system and method that finds patterns in high-frequency components of an audio signal, encodes the audio signal into an encoded bitstream along with pattern information, and then uses the patterns to reconstruct the high-frequency components during decoding. The high-frequency components can be reconstructed using the pattern information alone. Embodiments of the system and method map normalized subband signals of the audio signal to a scaled representation of a time-frequency grid containing multiple tiles and perform statistical analysis on each tile to estimate subband parameters and determine whether a pattern exists. If a pattern does exist, it can be encoded in the encoded bitstream, transmitted, and used to reconstruct the high-frequency components at the decoder. A direct search technique and a fast Fourier transform (FFT) technique may be used to perform the statistical analysis.

Abstract: The present invention is a server-based translation protocol for improving translation performance for cases where a large number of documents are generated in a source language context but the controversies are adjudicated in a different language context. The protocol is intended to improve terminology consistency, offset the effects of contextual shift on perceived facts in translations, and improve task-tracking order. If the protocol is used by well trained and motivated document reviewers in a collaborative and harmonic environment, it can reduce unnecessary translations, improve translation accuracy, minimize the needs for amendments, control translation costs, and help the client significantly improve its litigation position.

Abstract: An automatic speech recognition (ASR) system and method are provided for using sub-lexical language models together with word level pronunciation lexicons. These approaches operate by introducing a transduction between sequences of sub-lexical units and sequences of words.

Abstract: An acoustic model adaptation system includes a memory device and a model selector engine coupled to the memory device. The model selector engine is configured to compile information of environmental conditions to identify a current speech environment for audio input into a speech recognizer on a device. The model selector engine is further configured to compare the information of the environmental conditions with profiles of acoustic models. Each profile associates with an acoustic model. Each acoustic model compensates for background noise or acoustical distortions of the audio input. The model selector engine is further configured to select a first acoustic model for the speech recognizer based on the information of the environmental conditions exclusive of audio input from the user.

Abstract: A sound analysis apparatus includes a sound information obtaining section chat obtains information relating to a sound acquired by a sound acquiring section that acquires the sound and distinguishes a spoken voice of a wearer from a spoken voice of another person, a phase difference deriving section that derives a relationship between a frequency and a phase difference with respect to the sound acquired by the plural sound acquiring sections, a dispersion deriving section that derives a dispersion that is the level of irregularity of the derived phase difference, and a distance deriving section that derives a distance between the wearer and the other person using a first dispersion derived in a case where the sound is distinguished as the spoken voice of the other person and a second dispersion derived in a case where the sound is distinguished as the spoken voice of the wearer.

Abstract: The present document relates to audio forensics, notably the blind detection of traces of parametric audio encoding/decoding. In particular, the present document relates to the detection of parametric frequency extension audio coding, such as spectral band replication (SBR) or spectral extension (SPX), from uncompressed waveforms such as PCM (pulse code modulation) encoded waveforms. A method for detecting frequency extension coding history in a time domain audio signal is described.

Abstract: Disclosed are a method and apparatus for decoding a an audiospeech signal using an adaptive codebook update. The method for decoding speech an audio signal includes: receiving an N+1-th normal frame data that is a normal frame transmitted after an N-th frame that is a loss frame data loss; determining whether an adaptive codebook of a final subframe of the N-th frame is updated or not by using the N-th frame and the N+1-th frame; updating the adaptive codebook of the final subframe of the N-th frame by using a the pitch index of the N+1-the frame; and synthesizing an audio a speech signal of by using the N+1-th frame.

Abstract: According to one embodiment, an apparatus for applying pitch features in automatic speech recognition is provided. The apparatus includes a distribution evaluation module, normalization module, and random value adjusting module. The distribution evaluation module evaluates the global distribution of pitch features of voiced frames in speech signals, and the global distribution of random values for unvoiced frames in speech signals. The normalization module normalizes the global distribution of random values for unvoiced frames based on the global distribution of pitch features of voiced frames. The random value adjusting module adjusts random values for unvoiced frames based on the normalized global distribution, so that the adjusted random values can be assigned to unvoiced frames in speech signals as pitch features of the unvoiced frames.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for discovery of problematic pronunciations for automatic speech recognition systems. One of the methods includes determining a frequency of occurrences of one or more n-grams in transcribed text and a frequency of occurrences of the n-grams in typed text and classifying a system pronunciation of a word included in the n-grams as correct or incorrect based on the frequencies. The n-grams may comprise one or more words and at least one of the words is classified as incorrect based on the frequencies. The frequencies of the specific n-grams may be determined across a domain using one or more n-grams that typically appear adjacent to the specific n-grams.