WO2005034083A1

WO2005034083A1 - Letter to sound conversion for synthesized pronounciation of a text segment

Info

Publication number: WO2005034083A1
Application number: PCT/US2004/030468
Authority: WO
Inventors: Gui-Lin Chen; Jian-Cheng Huang
Original assignee: Motorola, Inc.
Priority date: 2003-09-29
Filing date: 2004-09-17
Publication date: 2005-04-14
Also published as: RU2320026C2; CN1604184A; RU2006114705A; CN1308908C; KR100769032B1; KR20060056404A; EP1668629A1; EP1668629B1; EP1668629A4; DE602004019949D1

Abstract

There is described a method (200) for text to speech synthesis, the method (200) includes receiving (220) a text string and selecting at least one word from the string. Then a step of segmenting (240) the word into a sub-words forming a sub-word sequence with at least one of the sub-words comprising at least two letters. The step of identifying (250) provides for identifying phonemes for the sub-words and step (260) effects concatenating the phonemes into a phoneme sequence. A performing speech synthesis (280) on the phoneme sequence is then conducted.

Description

LETTERTO SOUND CONVERSIONFORSYNTHESIZED PRONUNCIATIONOF ATEXT SEGMENT

FIELD OF THE INVENTION The present invention relates generally to Text-To-Speech (TTS) synthesis. The invention is particularly useful for letter to sound conversion for synthesized pronunciation of a text segment. BACKGROUND OF THE INVENTION Text to Speech (TTS) conversion, often referred to as concatenated text to speech synthesis, allows electronic devices to receive an input text string and provide a converted representation of the string in the form of synthesized speech. However, a device that may be required to synthesize speech originating from a non-deterministic number of received text strings will have difficulty in providing high quality realistic synthesized speech. One difficulty is based on letter to sound conversion in which identical letters or groups of letters may have different sounds and vowel stress/emphasis depending on other adjacent letters and position in a text segment to be synthesized. In this specification, including the claims, the terms 'comprises', 'comprising' or similar terms are intended to mean a non-exclusive inclusion, such that a method or apparatus that comprises a list of elements does not include those elements solely, but may well include other elements not listed.

SUMMARY OF THE INVENTION According to one aspect of the invention there is provided a method for text to speech synthesis, the method including: receiving a text string and selecting at least one word therefrom; segmenting the word into a sub-words the sub-words forming a sub-word sequence with at least one of sub-words comprising at least two letters; identifying phonemes for the sub-words; concatenating the phonemes into a phoneme sequence; and performing speech synthesis on the phoneme sequence. Suitably, The sub-word sequence is determined by analysis of possible sub-words that could comprise the word Preferably, each one of the possible sub-words has an associated predefined weight. Suitably, the sub-words with the maximum combined weights that form the selected word are chosen to provide the sub-word sequence. The sub-word sequence is suitably determined from analysis of a Direct Acyclic Graph. Suitably, the identifying phonemes use a phoneme identifier table comprising a phonemes corresponding to at least one said sub-word. Preferably, the identifier table also comprises a position relevance indicator that indicates the relevance of the position of the sub-word in the word. There may also suitably be a phoneme weight associated with the position relevance indicator.

BRIEF DESCRIPTION OF THE DRAWINGS In order that the invention may be readily understood and put into practical effect, reference will now be made to a preferred embodiment as illustrated with reference to the accompanying drawings in which: Fig. 1 is a schematic block diagram of an electronic device in accordance with the present invention; Fig. 2 is flow diagram illustrating a method for text to speech synthesis; Fig. 3 illustrates a Direct Acyclic Graph (DAG); Fig 4 is part of a mapping table that maps symbols with phonemes; Fig 5 is part of a phoneme identifier table; and Fig 6 is part of a vowel pair table. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT OF THE INVENTION Referring to Fig. 1 there is illustrated an electronic device 100, in the form of a radio-telephone, comprising a device processor 102 operatively coupled by a bus 103 to a user interface 104 that is typically a touch screen or alternatively a display screen and keypad. The electronic device 100 also has an utterance corpus 106, a speech synthesizer 110, Non Volatile memory 120, Read Only Memory 118 and Radio communications module 116 all operatively coupled to the processor 102 by the bus 103. The speech synthesizer 110 has an output coupled to drive a speaker 112. The corpus 106 includes representations of words or phonemes and associated sampled, digitized and processed utterance waveforms PUWs. In other words, and as described below, the Non Volatile memory 120 (memory module) in use for Text-To-Speech (TTS) synthesis (the text may be received by module 116 or otherwise). Also the waveform utterance corpus comprises sampled and digitized utterance waveforms in the form of phonemes and stress/emphasis of prosodic features. As will be apparent to a person skilled in the art, the radio frequency communications unit 116 is typically a combined receiver and transmitter having a common antenna. The radio frequency communications unit 116 has a transceiver coupled to antenna via a radio frequency amplifier. The transceiver is also coupled to a combined modulator/demodulator that couples the communications unit 116 to the processor 102. Also, in this embodiment the non- volatile memory 120 (memory module) stores a user programmable phonebook database Db and Read Only Memory 118 stores operating code (OC) for device processor 102 Referring to Fig. 2 there is illustrated a method 200 for text to speech synthesis. After a start step 210 a step 220 of receiving a text string TS from the memory 120 is performed. The text string TS may have originated from a text message received by module 116 or by any other means. Step 230 provides for selecting at least one word from the text string TS and a segmenting step 240 provides for Segmenting the word into sub-words the sub- words forming a sub-word sequence with at least one of sub-words comprising at least two letters. An identifying step 250 then provides for identifying phonemes for the sub-words. A concatenating step 260 then provides for concatenating the phonemes into a phoneme sequence. The sub-word sequence is determined by analysis of all possible sub-words that could comprise the selected word. For instance, referring briefly to the Direct Acyclic Graph (DAG) of Fig. 3, if the selected word was "mention", then the Direct Acyclic Graph DAG is constructed with all possible sub-words that could comprise the selected word "mention". With each sub-word a pre-defined weight WT is provided, for example as shown the sub-word "ment", "men" and "tion" have respective weights 88, 86 and 204. Accordingly, the concatenating step 260 traverses the DAG and selects the sub-words with the maximum combined (summed) weights WT that form the selected word. In the case for the word "mention" the sub-words "men" and "tion" would be selected. The step 250 of identifying phonemes uses two tables, stored in memory 120, one table part of which is illustrated in Fig. 4 is a mapping table

MT that maps symbols with phonemes. As shown, phoneme ae is identified by symbol @, whereas phoneme th is identified by symbol D. The other table is a phoneme identifier table PIT as part of which is illustrated in FIG. 5. The phoneme identifier table PIT comprises a sub-word field; phoneme weight field; position relevance field(s) or indicators; and a phoneme identifier field(s). For instance, in Fig 5, the first line is aa 120 A_C, where aa is the sub- word; 120 is the phoneme weight, the letter A is the position relevance and "C" is the phoneme identifier corresponding to the sub-word aa. The position relevance may be labeled as: A with the meaning relevant for all positions; I with the meaning relevant for sub-words at the beginning of a word; M with the meaning relevant for sub-words in the middle of a word; and F with the meaning relevant for sub-words at the end of a word. Thus, using the phoneme identifier table PIT and with the sub-words position in a word then the step 250 of identifying phonemes is effected. The phoneme weights and DAC weights pre-defined weight WT are the same weights obtained from Fig 5. These weights were determined such that if we choose the occurring time as a weight, one sub-string has higher weight than the string itself. As a consequence, if we select the segmentation form with maximal weight as the result, short morpheme-like string is always preferable. For instance, the word seeing will be segmented as s ee|m|g instead of s\ee\ing. But overall, the relationship between long string and phoneme sequence is more confident. To ensure the high priority of long morpheme-like string, we consider the following aspects: affix If one short string is a prefix or suffix of a long string, we add its occurring time to the long string; but other sub-strings are not being considered. ambiguity In some cases, one morpheme-like string can correspond to multiple phoneme strings; for instance, en can pronounce as ehn and axn. To decrease uncertainty, we employ the string positions such as word initial, word medial and word final. Even under this condition, the morpheme-like string can correspond to more than one phoneme string. To overcome this problem, we choose the phoneme string with maximal occurring time and calculate the ratio r as follows: _{r ^} max{N_llk} (3) where u is the string index while k is the position index, if r < α ( α is a threshold, α = 0.7) , we exclude this morpheme- like string. For example, the word-final en can pronounce as ehn and αxn, if the total time is 1000, if the time corresponding to αxn is 800 (of course, it is the maximal time), r=0.8. Hence, we will add the word- final en to the list. minimal occurring time. We also set the minimal occurring time, min (min-9) as the threshold. Each string whose occurring time is less than this value has been discarded. Under these constraints, we assign each string weight W_s in the following way: W_s = 101nN_s , N_s is the adjusted occurring time. The method 200 next effects a step 265 of performing stress or emphasis assignment on the phonemes that represent vowels. This step 265 identifies vowels from the suitably identified phonemes identified in the pervious step 250. Essentially, this step 265 searches a relative strength/weakness vowel pair table stored in the memory 120. Part of this vowel pair table is illustrated in Fig 6. For instance consider 3 vowels that could be identified as phonemes in a word, these vowels being identified by the symbols (obtained fro the mapping table MT) 'ax; aa; and ae. Then by analysis of the vowel pair table when 'ax occurs before aa then a stress weight of 368 is indicated in contrast to a, stress weight of 354 when aa occurs before 'ax. Hence, by analyzing the vowel pair table for 'ax; aa; and ae the following analysis results: the vowel identified by symbol ae has primary (the most) stress; the vowel identified by symbol 'ax has secondary stress; and the vowel identified by symbol aa has no stress. Essentially, the stress weights are determined by using a training lexicon. Each entry in this lexicon has word format and its corresponding pronunciation, including stress, syllable boundary and letter-to-phoneme alignment. So based on this lexicon, stress was determined by statistical analysis. In this regard, stress reflects strong/weak relationship between vowels. To generate required data, statistical analysis for all entries in the lexicon were therefore conducted. Specifically, within the scope of a word, if vowel vf is stressed, VJ is unstressed, we assign one point for the pair (v , ) and zero point for pair (VJ,V ). If both are unstressed, the point is also zero. A test step 270 is then performed to determine if there are any more words in the text string TS that need to be processed. If yes then the method 200 returns to step 230, otherwise a performing speech synthesis on the phoneme sequence is effected at a performing step 280. The performing speech synthesis is effected by the synthesizer 110 on the phoneme sequence for each of the words. The method 200 then ends at an end step 290. During the performing speech synthesis of step 280 the stress (primary, secondary or no stress as appropriate) on the vowels is also used to provide an improved synthesized speech quality by appropriate stress emphasis. Advantageously, the present invention improves or at least alleviates sounds and vowel stress/emphasis depending on other adjacent letters and position in a text segment to be synthesized. The detailed description provides a preferred exemplary embodiment only, and is not intended to limit the scope, applicability, or configuration of the invention. Rather, the detailed description of the preferred exemplary embodiment provides those skilled in the art with an enabling description for implementing preferred exemplary embodiment of the invention. It should be understood that various changes may be made in the function and arrangement of elements without departing from the spirit and scope of the invention as set forth in the appended claims.

Claims

WE CLAIM:

1. A method for text to speech synthesis, the method including: receiving a text string and selecting at least one word therefrom; segmenting the word into a sub-words the sub-words forming a sub-word sequence with at least one of sub-words comprising at least two letters; identifying phonemes for the sub-words; concatenating the phonemes into a phoneme sequence; and performing speech synthesis on the phoneme sequence.

2. A method for text to speech synthesis, as claimed in claim 1, wherein the sub-word sequence is determined by analysis of possible sub- words that could comprise the word.

3. A method for text to speech synthesis, as claimed in claim 1, wherein each one of the possible sub-words has an associated pre-defined weight.

4. A method for text to speech synthesis, as claimed in claim 1, wherein the sub-words with the maximum combined weights that form the selected word are chosen to provide the sub-word sequence.

5. A method for text to speech synthesis, as claimed in claim 4, wherein the sub-word sequence is suitably determined from analysis of a Direct Acyclic Graph.

6. A method for text to speech synthesis, as claimed in claim 1, wherein the identifying phonemes uses a phoneme identifier table comprising a phonemes corresponding to at least one said sub-word.

7. A method for text to speech synthesis, as claimed in claim 6, wherein the identifier table also comprises a position relevance indicator that indicates the relevance of the position of the sub-word in the word.

8. A method for text to speech synthesis, as claimed in claim 7, wherein there is a phoneme weight associated with the position relevance indicator.