Mietta's Labeling Page

About this page

I am a phonetician who is interested in Finnish spontaneous speech. For my work, I need to collect large amounts of conversational speech data of a high acoustic quality, and for the most part, the material has to be segmented and transcribed manually. So far, I have been doing a lot of labeling work myself. Labeling is also quite error-prone, and it is easy to "forget" or change your own segmentation principles. People tend to disagree on segment boundary placement and the phonetic labels for segments, and yet they are not very consistent themselves. This page is to provide documentation of the principles that I have been using (or those that I think I have been using). Primarily, it is a notebook for myself, but I hope it will also help others to learn the exhausting secrets of labeling work... Many of the principles I state below have been either inspired by or jointly agreed upon with my friend Nina Alarotu, who undoubtedly still is the number one speech labeler in Finland today, considering the amount of labeling hours she spent with the Finnish speech database! The symbol lists are based on the Worldbet references that are available on the web, and I have only made slight modifications in order to have a useful and easy-to-remember symbol set for my purposes.

This page is constantly under construction: I will elaborate on the different aspects and problems of transcription as I go along. Please feel free to comment!

Mietta Lennes
mietta.lennes@helsinki.fi

A few general principles in speech segmentation and phonetic transcription

Labeling tools

I use Praat for the segmentation and labeling of my sound files.

Praat has the possibility of using very long sound files, whose size can be as big as 2 gigabytes. These sounds can be used as LongSound objects, meaning that the sound file is not fully loaded into memory, but it can be viewed and accessed with a corresponding TextGrid object. TextGrids are nothing but text files, so you can use long conversational speech files from CD-ROMs or other read-only media and only edit and save the small TextGrid files each time when labeling the speech file. You can easily share the TextGrids with others, and each person can modify the TextGrid for herself, as long as the sound file is not changed.

Top of this page

Levels of transcription

In Praat, you can use many IntervalTiers for different levels of transcription. I have tried to utilize this possibility to the full, since many different levels allow many kinds of searching criteria later on. Here, I discuss those levels which I think are useful. I also point out some segmentation and transcription problems and give my solutions to them (for the time being; I may change my mind...)

1. Phone level: My transcription method is rather phonetic: I try to find the closest possible Worldbet symbol for each speech sound I find in the signal, adding all necessary diacritics. I determine the phone boundaries and labels in an acoustic-phonetic basis: partly by inspecting spectrograms, intensity curves, etc., and partly by listening to shorter and longer segments of speech. A phonetic transcription is supposed to describe what is actually produced, so I make as little reference to phonemes or "what-there-should-be" as possible.

2. Phoneme level: This level I do not use, because it has turned out to be too difficult and frustrating to look for the corresponding speech segments for a Finnish phoneme string. The phonemic structure of a word of casual Finnish speech cannot always be defined in a straightforward way, since the corresponding written word forms may not exist. Moreover, we do have an orthographic transcription of the utterances, and it is quite simple to convert the orthographic symbol strings into phoneme-like strings, since Finnish orthography is said to follow the phonemic structure rather closely. However, I am working on a Praat script that would automatically convert my narrow phonetic segmentation and labeling to a "quasi-phonemic" segmentation and labeling for situations where this representation might be needed.

3. Syllable level: I have attempted to mark all syllable boundaries. This should be done after segmenting and labeling the phonetic level, perhaps at the same time as marking word boundaries. The problems with this level are similar to those with word boundaries: In Finnish spontaneous speech, the definition of a syllable and its boundaries often relies completely on the intuition of the language speaker. Syllable boundaries in the middle of geminates are marked as geminates at word boundaries.
   A boundary in the syllable level does not have to coincide with a boundary in the phonetic level, but word boundaries have to coincide with syllable boundaries.
   

4. Word level: I have attempted to mark all word boundaries.This should be done after segmenting and labeling the phonetic level. A boundary in the word level does not have to coincide with a boundary in the phonetic level, but it has to coincide with a syllable boundary.
   Problems:
   • A geminate (a double consonant) at the boundary, in case it is clearly a case of final/initial doubling (see explanation in Finnish). I mark the whole consonant segment as belonging to one of the words.
   • A geminate (a double consonant) at the boundary, in case it seems that the consonant "belongs" to both words and is not necessarily a result of final/initial doubling. First, I try to look for any visible acoustic changes in the spectrogram that would give a legitimate place for the boundary. These include voice quality changes or movements in spectral structure. If there are no such changes during the phonetic-level consonant segment, I place the word boundary halfway of the segment.
   • Previous word ends in the same vowel than the next word begins = a long vowel at the boundary. First, I try to look for any visible acoustic changes in the spectrogram that would give a legitimate place for the boundary. These are most often voice quality changes. Especially if the word on the right side is accented, it is quite common to find creak/ a phase of glottalization at the boundary, sometimes even a glottal stop. In the case of a "trough" of glottalization, I mark the boundary after the longest glottal period, just before the F0 starts to rise again. This turning point is often reflected in vowel quality. If there is a glottal stop between words, the glottal closure is not part of either word, but the release belongs to the rightmost word. If all else fails, I mark the word boundary halfway of the vowel segment.

   • Words at the end of utterance: If there is, e.g., a [h]-like sound at the end of the last word, I take this as belonging to the utterance, but not to the word. However, where this [h]-sound is clearly coloured by the final vowel of the utterance-final word, it belongs to the word as well, and should be phonetically labeled as the voiceless counterpart of the vowel (e.g., "A_0").

5. Accent level: The boundaries in this level should be identical to word level boundaries. I have marked with an 'x' all words that to me are clearly prominent (bear a sentence accent). As in any perceptual prominence judgments, my decisions are not the ultimate truth, but I consider it useful to have even a preliminary and subjective indication of accentuation. At this point, no distinction is made between different sentence accent types etc. Several or none accented words are allowed within one utterance.
   Thus, (sentence) accent is now marked roughly as a property of a whole word, and no indication about the more specific domain of accent is given. In Finnish, there is no lexical stress (lexical items are generally not distinguished with regard to accent placement), but in case a word receives a (sentence) accent, it is usually the first syllable that is perceived as most prominent. So, although the real situation is more complicated than that, we may assume that the first syllable of an accented word is the prominent one.

6. Intonation Phrase level (IP): This level is not very well developed. I have just been trying to cut utterances into smaller phrases which would be somewhat coherent prosodically. No fine definitions so far.

7. Utterance level: An utterance is a stretch of speech by a single speaker between two pauses. A pause is any period of silence where the speaker is not articulating a speech sound. The labels of utterance level intervals are the orthographic transcriptions of the utterances. Boundaries must coincide with intonation phrase boundaries.

8. Nonspeech level: This level describes long-term "non-speech" properties which may overlap with speech sound production: breathing (in or out; ingressive speech can occur for short utterances in Finnish), laughter; or external sounds, like different background noises. Those phenomena that cannot overlap with speech articulations can be marked in the utterance level or phone level (the Worldbet symbols for these features always begin with a dot '.', so they can be distinguished from phones).
   Actually, I have not started to use this level yet, but I plan to...
   

9. Topic level: A rather loosely defined tier, where I mark any big changes in the discourse topic, e.g., "Grandma's new house", "New job", etc. I mark the topics in English, to make the TextGrid file more legible to the international audience.

10. "Ready?" level: I label with 'ok' all the stretches of the sound file which have been segmented, labeled and checked to the full. This way, when running analysis scripts, I can only select good and checked data.

Top of this page

Worldbet

For the phonetic transcription of speech, I use Worldbet, which is an ASCII version of the International Phonetic Alphabet (IPA). There are a few basic Worldbet documentation sources on the web (please let me know if you spotted better ones):

• James L. Hieronymus, ASCII Phonetic Symbols for the World's Languages: Worldbet. Technical report, Bell Labs, 1993. (PostScript document, 294 K. The download site of the above file is Center for Spoken Language Understanding at OGI.)
  This is the basic paper on Worldbet! Unfortunately, the printing quality is not very high.

• Chen Tao and Guo Shuang: Adaptation of IPA to World Languages

• IPA, Worldbet and OGIbet (by Ed Kaiser)
  

However, these sources are not quite consistent with regard to symbol definitions, and I have had to make a few slight modifications and additions. Below, I will try to explain each Worldbet symbol and convention that I use for Finnish. I will also refer to the corresponding IPA symbols.

In Worldbet, the base symbol is one ASCII character or sometimes a combination of two characters (see Worldbet symbols). Additional features or feature changes are indicated with diacritics, in the same manner as in IPA. Diacritics are separated from the base symbol with an underscore '_'. I have not avoided using diacritics whenever necessary. In case I need to give several diacritics for a sound, I have decided to add an underscore in front of each diacritic. This way, it is easy to refer to diacritics that have been given in a different order from within a script or a program.

It seems to me that a certain impression of vowel quality (e.g., if a vowel is centralized, or when indicating the voiced-voiceless distinction for consonants) can often be indicated with two or even more different symbol combinations (a base symbol, or base symbol plus diacritics). I often prefer to use a common and simple base symbol and to add more diacritics. Since we are not making specific claims about phonemic structure, and thus not "deriving" the phones from phonemes, it does not matter which symbol is used as a base, as long as the diacritics and the base symbol combined will imply the same set of perceived phonetic features. (Moreover, different transcribers will use slightly different symbols anyway!)

The list of Worldbet symbols

Top of this page

Worldbet symbols

as used for the transcription of spontaneous Finnish

General information about Worldbet and my transcription practices

Vowels
Consonants
Diacritics
Non-speech and other sounds

A couple of my old maps of Worldbet symbols:
Consonants
Vowels and diacritics

Top of this page

Vowels

The symbols I have more or less systematically used for Finnish are marked in bold.
(Unfortunately, the corresponding IPA symbols are still missing. I'm working on it...)

Top of this page

Consonants

Stops
Nasals
Fricatives
Affricates
Approximants
Laterals
Flaps/ taps
Trills
Ejectives
Implosives
Clicks

Top of this page

Stops

The symbols I have more or less systematically used for Finnish are marked in bold.
(Unfortunately, the corresponding IPA symbols are still missing. I'm working on it...)

Top of this page

Nasals

The symbols I have more or less systematically used for Finnish are marked in bold.
(Unfortunately, the corresponding IPA symbols are still missing. I'm working on it...)

Top of this page

Fricatives

The symbols I have more or less systematically used for Finnish are marked in bold.
(Unfortunately, the corresponding IPA symbols are still missing. I'm working on it...)

Top of this page

Affricates

Under construction...

The symbols I have more or less systematically used for Finnish are marked in bold.
(Unfortunately, the corresponding IPA symbols are still missing. I'm working on it...)

Top of this page

Approximants

Under construction...

Top of this page

Laterals

Under construction...

Top of this page

Flaps / taps

Under construction...

Top of this page

Trills

Under construction...

Top of this page

Ejectives

Under construction...

Top of this page

Implosives

Under construction...

Top of this page

Clicks

Under construction...

Top of this page

Diacritics

Top of this page

Non-speech and other sounds

Top of this page