Present invention relates to the field of synthesizing of speech or music, and more particularly without limitation, to the field of text-to-speech synthesis.
The function of a text-to-speech (TTS) synthesis system is to synthesize speech from a generic text in a given language. Nowadays, TTS systems have been put into practical operation for many applications, such as access to databases through the telephone network or aid to handicapped people. One method to synthesize speech is by concatenating elements of a recorded set of subunits of speech such as demi-syllables or polyphones. The majority of successful commercial systems employ the concatenation of polyphones.
The polyphones comprise groups of two (diphones), three (triphones) or more phones and may be determined from nonsense words, by segmenting the desired grouping of phones at stable spectral regions. In a concatenation based synthesis, the conversation of the transition between two adjacent phones is crucial to assure the quality of the synthesized speech. With the choice of polyphones as the basic subunits, the transition between two adjacent phones is preserved in the recorded subunits, and the concatenation is carried out between similar phones.
Before the synthesis, however, the phones must have their duration and pitch modified in order to fulfil the prosodic constraints of the new words containing those phones. This processing is necessary to avoid the production of a monotonous sounding synthesized speech. In a TTS system, this function is performed by a prosodic module. To allow the duration and pitch modifications in the recorded subunits, many concatenation based TTS systems employ the time-domain pitch-synchronous overlap-add (TD-PSOLA) (E. Moulines and F. Charpentier, “Pitch synchronous waveform processing techniques for text-to-speech synthesis using diphones,” Speech Commun., vol. 9, pp. 453-467, 1990) model of synthesis.
In the TD-PSOLA model, the speech signal is first submitted to a pitch marking algorithm. This algorithm assigns marks at the peaks of the signal in the voiced segments and assigns marks 10 ms apart in the unvoiced segments. The synthesis is made by a superposition of Hanning windowed segments centered at the pitch marks and extending from the previous pitch mark to the next one. The duration modification is provided by deleting or replicating some of the windowed segments. The pitch period modification, on the other hand, is provided by increasing or decreasing the superposition between windowed segments.
Despite the success achieved in many commercial TTS systems, the synthetic speech produced by using the TD-PSOLA model of synthesis can present some drawbacks, mainly under large prosodic variations.
Example of such PSOLA methods are those defined in documents EP-0363233, U.S. Pat. No. 5,479,564, EP-0706170. A specific example is also the MBR-PSOLA method as published by T. Dutoit and H. Leich, in Speech Communication, Elsevier Publisher, November 1993, vol. 13, N. degree. 3-4, 1993. The method described in document U.S. Pat. No. 5,479,564 suggests a means of modifying the frequency by overlap-adding short-term signals extracted from this signal. The length of the weighting windows used to obtain the short-term signals is approximately equal to two times the period of the audio signal and their position within the period can be set to any value (provided the time shift between successive windows is equal to the period of the audio signal). Document U.S. Pat. No. 5,479,564 also describes a means of interpolating waveforms between segments to concatenate, so as to smooth out discontinuities. In prior art text-to-speech systems a set of pre-recorded speech fragments can be concatenated in a specific order to convert a certain text into natural sounding speech. Text-to-speech systems that use small speech fragments have many such concatenation points. Especially when the speech fragments are spectrally different, these joins produce artefacts that reduce the intelligibility. In particular, when two speech segments from different recording times are to be concatenated, the resulting speech can have a discontinuity at the joint of the two segments. For example, when a vowel is synthesized, the left part mostly comes from a different recording than the right part. This makes it impossible to reproduce the exact color of a vowel.
The slight differences in the formant trajectories produce a sudden jump at the joint location. What is mostly done in the prior art to reduce this effect is to re-record the speech fragment until it matches with the rest or add different versions (extra fragments) to minimize the difference.
The present invention therefore aims to provide an improved method of synthesizing of a speech signal, the speech signal having at least a first diphone and a second diphone. The present invention further aims to provide a corresponding computer program product and computer system, in particular text-to-speech system.
The present invention provides for a method of synthesizing of speech signal based on first and second diphone signals which are superposed at their joint. The invention enables a smooth concatenation of the diphone signals without any audible artefacts. This is accomplished by appending periods of an end interval of the first diphone signal in inverted order at the end of the first diphone signal and by appending periods of a front interval of the second diphone signal at the beginning of the second diphone signal. The end and front intervals are overlapped to produce the smooth transition.
In accordance with an embodiment of the invention the end and front intervals of the first and second diphone signal are identified by a marker. Preferably the end and front intervals contain periods which are about steady, i.e. which have approximately the same information content and signal form. Such end and front intervals can be identified by a human expert or by means of a corresponding computer program. Preferably the first analysis is performed by means of a computer program and the result if reviewed by a human expert for increased precision.
In accordance with a further embodiment of the invention the last period of the end interval and the first period of the front interval are not appended. This has the advantage that no periodicity is introduced into the signal by the immediate repetition of two identical periods.
In accordance with a further embodiment of the invention a windowing operation is performed on the end and front intervals as well as on the respective appended periods by means of fade-out and fade-in windows, respectively. Preferably a raised cosine window function is used for voiced end intervals and the appended periods, whereas for unvoiced end intervals and the appended periods a sine window is used as a fade-out window. Likewise a raised cosine is used as a window function for smoothening the beginning of a voiced segment of the second diphone or a sine window for unvoiced segments.
In accordance with an embodiment of the invention a duration adaptation is performed for the intervals to be overlapped. Especially if the intervals have different durations this is advantageous in order to avoid the introduction of abrupt signal transitions.
In accordance with a further embodiment of the invention, text-to-speech processing is performed by concatenating diphones in accordance with the principles of the present invention. This way a natural sounding speech output can be produced.
It is important to note that the present invention is not restricted to the concatenation of diphones but can also be advantageously employed for the concatenation of other speech units such as triphones, polyphones or words.
In the following embodiments of the invention are described in greater detail by making reference to the drawings in which:
In step 102 periods within the end interval of the diphone signal A are repeated in inverted order in order to provide a fade-out interval which is appended at the end of the end interval. In step 104 the end interval with its' appended fade-out interval are windowed by means of a fade-out window function in order to smoothly fade out the diphone signal at its' end. Likewise a diphone signal B is provided in step 106. The diphone signal B has at least one associated marker in order to identify a front segment of the diphone signal B. In step 108 at least some of the front intervals periods are appended at the beginning of the front interval of the diphone signal B in inverted order. This way a fade-in interval is provided. In step 110 the front interval and the appended fade-in interval are windowed by means of a fade-in window. This way a smooth beginning of the diphone signal B is provided. In step 112 a duration adaptation is performed. This means that the durations of the end and front intervals of the diphone signals A and B are modified such that the end and fade-in intervals have the same duration. Likewise the durations of the fade-out and front intervals are adapted. In step 114 an overlap and add operation is performed on the diphone signals A and B with the processed end and fade-in intervals and the fade-out and front intervals. This way a smooth concatenation of the diphone signals A and B is accomplished. For voiced segments usage of the following raised cosine window function is preferred:
For unvoiced segments, a sine window is used:
The advantage of using a sine-window is that this ensures that the total signal envelope in power-domain remains constant. Unlike a periodic signal, when two noise samples are added, the total sum can be smaller than the absolute value of any of the two samples. This is because the signals are (mostly) not in-phase. The sine-window adjusts for this effect and removes the envelope-modulation.
Number | Date | Country | Kind |
---|---|---|---|
0207887205 | Sep 2002 | EP | regional |
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/IB03/03624 | 8/8/2003 | WO | 3/15/2005 |