Abstract: The described implementations relate to speech interfaces and in some instances to speech pattern recognition techniques that enable speech interfaces. One system includes a feature pipeline configured to produce speech feature vectors from input speech. This system also includes a classifier pipeline configured to classify individual speech feature vectors utilizing multi-level classification.

Abstract: An encoder/decoder uses intra-coded B-fields [“BI-fields”]. For example, rather than encoding many of the macroblocks in a field as intra macroblocks at a scene change, an encoder can choose to encode the entire field as intra. Encoding an entire field as a BI-field allows increased coding efficiency through reduced coding overhead at macroblock level, and without intra-frame dependency, the BI-field can still be dropped if appropriate in low-bitrate applications.

Abstract: For interlaced B-frames, an encoder/decoder computes direct mode motion vectors for a current macroblock by selecting at most one representative motion vector for each of the top and bottom fields of the co-located macroblock of the previously decoded, temporally subsequent anchor. For example, the selecting is performed based at least in part on the mode of coding the current interlaced B-frame's macroblock (e.g., 1MV mode, 2 Field MV mode, etc.). For interlaced B-fields, an encoder/decoder selects direct mode motion vectors using logic that favors the dominant polarity if the corresponding macroblock in the corresponding field of the next anchor picture was coded using four motion vectors. For example, if the corresponding macroblock's same polarity motion vectors outnumber its opposite polarity motion vectors, the encoder/decoder calculates the median of the same polarity motion vectors to obtain a motion vector for deriving direct mode motion vectors.

Abstract: Techniques and tools for coding/decoding of video images, and in particular, B-frames, are described. In one aspect, a video encoder/decoder determines a fraction for a current image in a sequence. The fraction represents an estimated temporal distance position for the current image relative to an interval between a reference images for the current image. The video encoder/decoder processes the fraction along with a motion vector for a first reference image, resulting in a representation of motion (e.g., constant or variable velocity motion) in the current image. Other aspects are also described, including intra B-frames, forward and backward buffers for motion vector prediction, bitplane encoding of direct mode prediction information, multiple motion vector resolutions/interpolation filters for B-frames, proactive dropping of B-frames, and signaling of dropped predicted frames.

Abstract: An encoder sends binary information indicating whether a prediction mode is forward or not-forward for one or more macroblocks in an interlaced B-field. For example, the encoder sends forward/not-forward decision information at B-field level in a compressed bitplane. Sending forward/not-forward prediction mode decision information in a compressed bitplane at B-field level can reduce coding overhead for prediction mode coding. A decoder performs corresponding decoding.

Abstract: For interlaced B-fields or interlaced B-frames, an encoder/decoder uses 4MV coding. For example, 4MV is used in one-direction prediction modes (forward or backward modes), but not in other available prediction modes (e.g., direct, interpolated). Using 4MV allows more accurate motion compensation for interlaced B-fields and interlaced B-frames; limiting 4MV to forward and backward modes reduces coding overhead and avoids decoding complexity associated with combining 4MV with modes such as direct and interpolated.

Abstract: An indexing system uses a graph-like data structure that clusters features indexes together. The minimum atomic value in the data structure is represented as a leaf node which is either a single feature index or a sequence of two or more feature indexes when a minimum sequence length is imposed. Root nodes are formed as clustered collections of leaf nodes and/or other root nodes. Context nodes are formed from root nodes that are associated with content that is being indexed. Links between a root node and other nodes each include a sequence order value that is used to maintain the sequencing order for feature indexes relative to the root node. The collection of nodes forms a graph-like data structure, where each context node is indexed according to the sequenced pattern of feature indexes. Clusters can be split, merged, and promoted to increase the efficiency in searching the data structure.

Abstract: Techniques and tools for encoding and decoding motion vector information for video images are described. For example, a video encoder yields an extended motion vector code by jointly coding, for a set of pixels, a switch code, motion vector information, and a terminal symbol indicating whether subsequent data is encoded for the set of pixels. In another aspect, an encoder/decoder selects motion vector predictors for macroblocks. In another aspect, a video encoder/decoder uses hybrid motion vector prediction. In another aspect, a video encoder/decoder signals a motion vector mode for a predicted image. In another aspect, a video decoder decodes a set of pixels by receiving an extended motion vector code, which reflects joint encoding of motion information together with intra/inter-coding information and a terminal symbol. The decoder determines whether subsequent data exists for the set of pixels based on e.g., the terminal symbol.

Abstract: Indexing methods are described that may be used by databases, search engines, query and retrieval systems, context sensitive data mining, context mapping, language identification, image recognition, and robotic systems. Raw baseline features from an input signal are aggregated, abstracted and indexed for later retrieval or manipulation. The feature index is the quantization number for the underlying features that are represented by an abstraction. Trajectories are used to signify how the features evolve over time. Features indexes are linked in an ordered sequence indicative of time quanta, where the sequence represents the underlying input signal. An example indexing system based on the described processes is an inverted index that creates a mapping from features or atoms to the underlying documents, files, or data. A highly optimized set of operations can be used to manipulate the quantized feature indexes, where the operations can be fine tuned independent from the base feature set.

Abstract: Techniques and tools for encoding and decoding video images (e.g., interlaced frames) are described. For example, a video encoder or decoder processes 4:1:1 format macroblocks comprising four 8×8 luminance blocks and four 4×8 chrominance blocks. In another aspect, fields in field-coded macroblocks are coded independently of one another (e.g., by sending encoded blocks in field order). Other aspects include DC/AC prediction techniques and motion vector prediction techniques for interlaced frames.

Abstract: A decoder receives luma motion vector information for plural luma motion vectors for a macroblock (e.g., a 4:2:0 macroblock). The decoder derives a chroma motion vector for each of the plural luma motion vectors by performing at least one calculation on the luma motion vector information, maintaining a 1:1 ratio of chroma motion vectors to luma motion vectors for the macroblock. For example, the decoder receives four luma (frame or field) motion vectors for a macroblock and derives four chroma motion vectors for the macroblock. The deriving can comprise sub-sampling and/or rounding (e.g., using a field-based rounding table).

Abstract: Biasing of language model customization due to repetitious data is substantially reduced by introducing novelty screening to data harvesting process. Novelty detection based filtering is added to ensure that an adaptation system gives more weight to representative adaptation data that is not repetitious. The value of the adaptation data is preserved and the process prevented from being polluted when the same data is seen multiple times, such as the original posting in an email thread, various versions of the same document, and the like. The screening technique may be built on top of existing data harvesting mechanisms as already seen data is used to determine the novelty of a particular portion of the data. A window into the new data, fixed or variable size, is compared against the already collected data to determine the likelihood that the data is novel.

Abstract: Repetition of content words in a communication is used to increase the certainty, or, alternatively, reduce the uncertainty, that the content words were actual words from the communication. Reducing the uncertainty of a particular content word of a communication in turn increases the likelihood that the content word is relevant to the communication. Reliable, relevant content words mined from a communication can be used for, e.g., automatic internet searches for documents and/or web sites pertinent to the communication. Reliable, relevant content words mined from a communication can also, or alternatively, be used to automatically generate one or more documents from the communication, e.g., communication summaries, communication outlines, etc.

Abstract: A clustering tool to generate word clusters. In embodiments described, the clustering tool includes a clustering component that generates word clusters for words or word combinations in input data. In illustrated embodiments, the word clusters are used to modify or update a grammar for a closed vocabulary speech recognition application.

Abstract: Various techniques and tools for chrominance motion vector rounding are described. For example, during motion compensation, a video encoder or decoder converts luminance motion vectors into chrominance motion vector having quarter-pixel accuracy. Or, during motion compensation, a video encoder or decoder converts luminance motion vectors into chrominance motion vectors using one of multiple available chrominance motion vector rounding modes.

Abstract: In one aspect, for a first interlaced video frame in a video sequence, a decoder decodes a bitplane signaled at frame layer for the first interlaced video frame. The bitplane represents field/frame transform types for plural macroblocks of the first interlaced video frame. For a second interlaced video frame in the video sequence, for each of at least one but not all of plural macroblocks of the second interlaced video frame, the decoder processes a per macroblock field/frame transform type bit signaled at macroblock layer. An encoder performs corresponding encoding.

Abstract: High-fidelity transcoding techniques are described. For example, a video transcoder obtains re-coding data set elements for source format compressed video and uses the elements for coding decisions and/or equivalent parameters when compressing the video in a target format. This allows syntax elements and coding decisions to be maintained across the two formats, including picture types, intra/inter macroblock coding types, field/frame coding decisions, and/or quantization levels. This helps the transcoder match quality in the compressed video between the source and target formats, reducing transcoding losses. At the same time, the transcoder gives the target format encoder the freedom to exploit additional compression opportunities to reduce overall bitrate. The transcoder may apply proportional rate control so as to produce output at a constant or relatively constant bitrate.

Abstract: Predictive lossless coding provides effective lossless image compression of both photographic and graphics content in image and video media. Predictive lossless coding can operate on a macroblock basis for compatibility with existing image and video codecs. Predictive lossless coding chooses and applies one of multiple available differential pulse-code modulation (DPCM) modes to individual macro-blocks to produce DPCM residuals having a closer to optimal distribution for run-length, Golomb Rice RLGR entropy encoding. This permits effective lossless entropy encoding despite the differing characteristics of photographic and graphics image content.

Abstract: Techniques and tools for signaling reference frame distances are described. For example, a video encoder signals a code for a reference frame distance for a current field-coded interlaced video frame. The code indicates a count of frames (e.g., bi-directionally predicted frames) between the current frame and a preceding reference frame. The code may be a variable length code signaled in the frame header for the current frame. The encoder may selectively signal the use of a default value for reference frame distances rather than signal a reference frame distance per frame. A video decoder performs corresponding parsing and decoding.

Abstract: A decoder processes a first bitstream element (e.g., a pull-down flag) in a first syntax layer (e.g., sequence layer or entry point layer) above frame layer in a bitstream for a video sequence, the bitstream comprising encoded source video having a source type (e.g., progressive or interlace). The decoder processes frame data in a second syntax layer (e.g., frame layer) of the bitstream for a frame (such as an interlaced frame or progressive frame, depending on source type, or a skipped frame) in the video sequence. The first bitstream element indicates whether a repeat-picture element (e.g., a repeat-frame element or a repeat field-element) is present or absent in the frame data in the second syntax layer.