Diphone

Diphone, developed at Ircam in Paris, is a program for making transitions between sounds an effect commonly referred to as sound morphing. The program runs on Macintosh computers and is part of a set of programs supplied to the subscribers of the Ircam Forum. Originally, the term diphone refers to a transition between two phonemes, but for musical purposes this notion has been expanded to mean the transition between any two sounds, not necessarily of vocal origin. 

Ircam s Diphone addresses one of the most challenging synthesis problems of both artificial speech and sound composition the concatenation of sound sequences. In computer-synthesised speech the algorithm used to utter a specific syllable, such as mu, may not produce satisfactory results on different words. For instance, take the words music and mutation. The computer s rendition of the syllable mu, for the word music, used at the beginning of the word mutation sounds artificial because the transition (i.e. the diphone) between u and s in the word music and the transition between u and t in the word mutation involve different spectral behaviour. This transition problem is also evident in music composition where certain articulations and musical passages are clearly more appropriate to our ears than others. This problem is even more challenging for electronic musicians because they deal with a much larger repertoire of sounds to combine and articulate. 

Although this book does not focus on music composition. Diphone is introduced here as an example of a synthesis system that was primarily designed to address a problem pertinent to all electronic musicians, and more specifically to those working with sampled 

The program concatenates the sounds by applying an algorithm that interpolates the analysis data of neighbouring segments. It is important to note that Diphone does not manipulate the sounds directly, only the analysis data. The whole sequence is synthesised only when the user commands the computer to do so. 

Although originally designed to work with additive synthesis. Diphone also supports other synthesis techniques such as formant synthesis (discussed in Chapter 3) and source modelling (Chapter 4). 

---

Moulines and Charpentier, 1990] Moulines, E. and Charpentier, F. (1990). Pitch-synchronous waveform processing techniques for text-to-speech synthesis using diphones. Speech Communication, 9(5/6) 453-467. 

Poirot et al., 1988] Poirot, G., Rodet, X., and Depalle, P. (1988). Diphone sound synthesis based on spectral envelopes and harmonic/noise excitation functions. Proc. of International Computer Music Conference, Koln. 

DIAPHENYLSULPHON DIAPHENYLSULPHONE DIPHONE DISULONE DSS DUBRONAX DUMITONE EPORAL 1358F F 1358 MALOPRIM... 

Further oxidation of dibenzoylsinomenolquinone [xciv, R = CO] with 30 per cent, hydrogen peroxide in glacial acetic acid affords 4 5 -dimothoxy-fi 6 -di-(benzoyloxy)-diphonic acid [xov, R = -CO],... 

Note that multisampled wavetable synthesis, phoneme speech synthesis, and diphone speech synthesis all take into account a priori the notion of consonants and attacks, versus vowels and steady states. This allows such synthesis techniques to do the right thing, at the expense of someone having to explicitly craft the set(s) of samples by hand ahead of time. This is not generally possible, or economically feasible, with arbitrary PCM. 

One of the most the most popular choices of unit is that of the diphone. A diphone is defined as a unit which starts in middle of one phone and extends to the middle of the next phone. The justification for using diphones follows directly from the target-transition model where we have a stable target region (the phone middle) whieh then has a transition period to the middle of the... 

A diphone synthesiser requires a different t5q)e of s mthesis specification Irom the phone centred one used previously. For a phoneme string of length N, a diphone specification of phone pairs of length A — 1 is required. Each item in the specification is comprised of two states known here as half-phones, named state 1 and state 2. Each of these half-phones has its own duration. We can define the start and end points of the diphones in a number of ways, but the simplest is to place them in the half way between the start and end of each phone. A number of options are available for the FO component as before, but one neat way is to have a single FO value for each half-phone, measured Irom the middle of that half phone. This is shown graphically in Figure 13.6. A single item in a diphone specification would then for example look like ... 

For convenience, we often denote a single diphone by both its half-phones joined by a hyphen. Hence the phone string /h eh 1 ow/ gives the diphone sequence /h-eh eh-11-ow/ or /h-eh/,... 

A key task is to carefully determine the full set of possible diphones. This is dealt with in full in Section 14.1, but for now, it suffices to say that we should find a database where there is a high degree of consistency, to ensure good continuity and joins, and where there is a broad range of phonetic contexts to ensine that all diphone combinations can be found. While it is possible to use existing databases, it is more common to specifically design and record a speech database that meets these particular requirements. 

Once the diphone tokens have been identified in the speech database we can extract the LP parameters. This is normally done with fixed frame analysis. First we window the section of speech into a number of frames, with say a fixed frame shift of 10ms and a window length of 25ms. The coefficients can be stored as is, or converted to one of the alternative LP representations. 

The full list of diphone types is called the diphone inventory, and once determined, we need to find units of such diphones in real speech. As we are only extracting one unit for each diphone type, we need to make sure this is in fact a good example for our purposes. We can state three... 

We need to extract diphones that will join together well. At any join, we have a left diphone and a right diphone which share a phoneme. For each phoneme we will have about 40 left diphones and 40 right diphones, and the aim is to extract units of these 80 diphones such that the last frame of all the left diphones is similar acoustically to the first frame of all the right... 

We wish for the diphones to be typical of their type makes sense to avoid outliers. 

E q)erience has shown that with a suitably recorded and analysed diphone set, it is usually possible to concatenate the diphones without any interpolation or smoothing at the concatenation point This is to be expected if the steady-state/transition model is correct (see Section 13.2.6). As we shall see in Chapter 16 however, this assumption is really only possible because the diphones have been well articulated, come from neutral contexts and have been recorded well. It is not safe to assume that other units can always be successfully concatenated in phone middles. 

Once we have a sequence of diphone units which match the phonetic part of the specification, we need to modify the pitch and timing of each to match the prosodic part of the specification. The rest of the chapter focuses on techniques for performing this task. 

In addition to these advantages we also have the possibility to modify the spectral characteristics of the units. One use of this would be to ensure that the spectral transitions at diphone joins are completely smooth. While careful design and extraction of the units from the database should help ensure smooth joins, they can t always be guarenteed, and so some spectral manipulation can be useful. 

This takes the form of a set of units where we have one unit for each unique type. Diphones are the most popular type of unit. 

The prosodic content is generated by explicit algorithms, and signal processing techniques are used to modify the pitch and timing of the diphones to match that of the specification. 

Within one t5q)e of diphone, all variation is accountable by pitch and timing differences... 

One way of realising this is as a direct extension of the original diphone principle. Instead of recording and analysing one version of each diphone, we now record and analyse one version for each combination of specified features. In principle, we can keep on expanding this methodology, so that if we wish to have phrase initial, medial and final units of each diphone, or a unit for every type or variation of pitch accent, we simply design and record the data we require. 

As we use more features, we see that in practical terms the approach becomes increasingly difficult. This is because we now have to collect significantly more data and do so in just such a way as to collect exactly one of each feature value. Speakers cannot of course not utter specific diphones in isolation, and so must do so in carrier words or phrases. This has the consequence that the speaker is uttering speech in the carrier phrases that is not part of the required list of effects. If we adhere strictly to this paradigm, we should throw this extra speech away, but this seems wasteful. The unit selection approach offers a solution to both these problems, which enables us... 

---