Speech recognition method for all languages without using samples

The speech recognition method for all languages without using samples provides speech recognition on all languages comprising:

• (1). a raw database of unknown words of any languages or unknown continuous voices with samples having the sample means and sample variances of all unknown words or voices spread (better uniformly) in the E×P dimensional vector space or a raw database of more unknown words or voices without samples;
• (2). a permanent database of known words pronounced by a speaker with a standard, clear and distinct utterance or by the testing speaker if the testing speaker has his or her special accent;
• (3). a pre-processor to delete noise and the time interval without speech signal;
• (4). a method to normalize the whole speech waveform of a word or a continuous voice, using E elastic frames (windows) without filter and without overlap and to transform the waveform into an equal-sized E×P matrix of the linear predict coding cepstra (LPCC) such that the same words have about the same LPCC at the same time position in their equal-sized E×P matrices of LPCC;
• (5). from the samples of each unknown word or each unknown continuous voice in the raw database, compute the sample means and sample variances of linear predict coding cepstra (LPCC);
• (6). for each known word in the permanent database, in the raw database of unknown words or voices with sample means and variances, use the Bayesian distances to find the top N unknown words from the raw database with the N shortest Bayesian distances to the known word in the permanent database or in the raw database of unknown words or voices without samples, find the top N unknown words or voices in the raw database with the N shortest absolute distances to the known word in the permanent database;
• (7). if the raw database has samples for unknown words, for each known word in the permanent database, compute the weighted average of the N sample means of the top N unknown words with the N shortest Bayesian distances to the known word and the linear predict coding cepstra (LPCC) of the known word as the mean for the known word, compute the weighted average of the N sample variances of the top N unknown words as the variance for the known word and the E×P matrix of the means and the variances is called the primary feature of the known word and is saved in the word database, which may have words of several languages saved together in the word database;
• (8). if the unknown words or voices in the raw database do not have samples, consider the N linear predict coding cepstra (LPCC) of the top N unknown words in the raw database with the N shortest distances to the known word and the LPCC of the known word in the permanent database as (N+1) data, compute the mean (weighted average) and the variance of (N+1) data and the E×P matrix of the means and the variances is called the primary feature of the known word and is saved in the word database;
• (9). a method to re-compute and stabilize the features of all known words several times in the word database such that all known words have better and more relatively stable features to each other in the word database and the final feature is called the standard pattern (denoted by a E×P matrix of means and variances) to denote a known word in the word database;
• (10). a method to normalize the whole waveform of an input unknown word using E elastic frames (windows) without filter and without overlap and to transform the waveform into an equal-sized E×P matrix of linear predict coding cepstra (LPCC), which is called the categorizing pattern to represent the unknown word;
• (11). a simplified Bayesian decision classifier is used to match the standard pattern of every known word in the word database with the categorizing pattern of an input unknown word and to find a known word in the word database with the least Bayesian distance to the unknown word to be the unknown word;
• (12). a method to partition an unknown sentence or name into D unknown words;
• (13). a Bayesian classifier to find the top F similar known words from the word database for an unknown word and an unknown sentence or name is represented by a D×F matrix of known similar words, which may belong to several different languages;
• (14). the D×F matrix of similar known words denoting an unknown sentence or name matches all known sentences and names in the sentence and name database and find a known sentence or name in the sentence and name database with the highest probability to be the unknown sentence or name;
• (15). a method to re-compute and improve the feature of an unknown word in the input unknown sentence or name such that the input unknown word, sentence or name is guaranteed to be recognized correctly.