US 5799277

Patent Overview — US 5,799,277 A

Title: Acoustic model generating method for speech recognition

Patent number / application: US 5,799,277 A (application US 08/547,794)

Assignee: Victor Company of Japan, Ltd. (original assignee listed); inventor Junichi Takami

Key dates (as listed): Priority 1994-10-25; filed 1995-10-25; granted/published 1998-08-25; anticipated expiration 2015-10-25; status "Expired – Fee Related."

Purpose and plain-language scope: The patent addresses speech recognition using hidden Markov models (HMMs). Its stated aim is to generate acoustic models with high representation efficiency — a "soft state network structure" representing maximum acoustic phenomena at high precision and robust stability — by deciding, under optimum conditions, the model unit, states of signal sources, state-sharing structure, and signal-source parameters, so that maximum speech phenomena are modeled with the minimum possible model parameters.

The claimed method iteratively performs splitting or merging on at least one signal source of an initial model, selecting one processing successively to form multiple signal sources, until the number of generated signal sources reaches a predetermined value for optimum speech recognition. When that number is reached, it decides — under a common evaluation criterion — the phoneme context category of the model unit, the number of states per model, sharing of states among models, sharing of signal sources among states, and the shape of each output probability distribution.

Merging may combine two similar signal sources into one without losing precision; it may calculate synthesized distribution magnitude and merge the pair with minimum distribution. Merging may precede splitting and is adopted only if its evaluation value (e.g., sum total likelihood) exceeds that of already-obtained models with the same number of signal sources; otherwise it is discarded and splitting is re-run. Splitting may split a source into two, allocate the original's two mixture distributions, and copy self/successor transition probabilities. Three splitting options are described: splitting two signal sources and reconstructing their shared structure; splitting one state into two phoneme-context states; and splitting one state into two time-direction states — with selection based on maximum likelihood.

Source: https://patents.google.com/patent/US5799277/en

What the excerpt does not establish: The supplied text is explicitly marked as excerpted (e.g., FIG. 4 description and later portions are omitted), so it does not provide the full specification, the exact wording or count of claims, all drawing contents, or the detailed flowchart steps. It also does not establish examination history, actual claim scope as allowed, or the accuracy of listed dates, which the source itself labels assumptions.