Sunday 23 April 2017

Transcription Boundaries

We all know about transcription, right … or do we? What are the ultimate goals? What are the limits, and are they inherent ones or self-imposed ones? I’m taking this opportunity to expand on some important transcription breakthroughs in the recent STEMMA V4.1 release.

Transcription Sources
Most people would begin by transcribing textual sources paragraph-by-paragraph, or sometimes line-by-line, dependent upon the actual source. It would quickly become apparent, though, that various scenarios cannot be transcribed directly as literatim text, such as uncertain characters or words, crossed-out text, text inserted or changed, and marginal annotation. What those people then have to do is decide on some form of mark-up to represent those scenarios (see Power of Annotation), but which one?

There are many schemes, ranging from old-style manuscript mark-up[1], through simple ASCII-character mark-up, to full-blown mark-up languages such as TEI (Text-Encoding Initiative). This latter technology, for instance, can represent semi-diplomatic or full diplomatic transcription of textual sources to digital form. Diplomatic transcription might be valuable for preservation but is that what we need for analysis?

Typescript Sources

This should be the easiest of the cases; when given a page of typed text then we might employ OCR to automate the conversion to a digital form. This is all very well if it is perfectly readable, but barely-readable sections, or additional hand-written annotation, would require a mark-up scheme.

And yet there are some subtle, but profoundly important, situations that rarely get mentioned. The presence of different fonts or typefaces in a printed electronic document would be taken for granted as indicating some semantic difference (e.g. a heading, abstract, or a footnote), but what about documents produced on an old-style typewriter? The presence of different typefaces might then indicate that a document was written on different machines at different times. Similarly with the alignment of the lines, or the marginal indent. But how do we indicate that in the digital form?

Suppose that there was a difference in the sophistication of the grammar in different sections, one that might provide a vital clue to different authors. How would that be represented?

A more important question is who would be the beneficiary of those indications? Schemes concerned with preservation will employ software taxonomies to categorise every eventually, but those subtleties — which could be crucial to the analysis and interpretation of a document  by a researcher — would almost certainly be excluded as unimportant in the digital representation.

Manuscript Sources

When transcribing manuscript documents then the points I’ve just raised become much more prominent. Contributions from different authors are generally more obvious because of their handwriting styles, and these obviously need to be distinguished in order to support any analysis, but what about stylistic variations?

Suppose that someone had underlined a word. That would clearly be an indication of emphasis, and the transcriber might represent it using some mark-up language (e.g. <u>word</u>) or some lightweight mark-up language (e.g. __word__), but what if a different word was underlined twice, or more times? This question also applies to text that has been struck-out. My point is that this is an important piece of information to capture, but how much more is required for analysis than for preservation?

As another example, consider if the author had used different coloured inks. James Joyce and Virginia Woolf both used different coloured pens or crayons in their work. Should a mark-up scheme have taxonomies for the basic colours, or all possible shades and hues? Character size and intensity (e.g. from a firm hand) can also be indicative of something. Who would benefit, though, from knowing that one paragraph was in dark green and another in light green: the software or the researcher? Is there a practical limit to the number of important variations that software taxonomies can distinguish, and if so then why do we insist on that route?

Audio Sources

Schemes that deal with audio transcription are generally specialist, and distinct from those related to textual transcription. The main reason is that those stylistic variations multiply exponentially. Not only do the transcriptions have to distinguish between contributions from different speakers, but they also need to indicate such things as speaking quickly/slowly, loudly/softly/whispered, singing, false accents, mimicry, and even different intonation. Schemes for audio transcription try to define taxonomies for these cases — although there will always be cases that aren’t covered — and the area of intonation is treated in a very formal way by linguistic analysis.

There may be cases of unknown words, slang, or strange pronunciations, each of which may need clarifying annotation.

While it is clear that the field is complex, I want to make an argument that there is a broad categorisation of the scenarios that has parallels in textual transcription, and that a single approach can deal with all three transcription source types. First, let’s look at some further complexities for audio.

There may be utterances or sounds from a given contributor that cannot be transcribed directly as text. For instance, a sneeze, cough, sniff, yawn, whistle, laugh, or swallow.

There may be a significant pause in someone’s speech that is important in the context of their words.

There may be any number of gestures or items of non-verbal communication that are equally important to capture within the transcript. For instance, a nod, smile, head-shake, squint, frown, or applause.

There may be instances where different voices — each of which is being transcribed — are overlapping each other, or where there is some untranscribed background contribution.


We can group all the above scenarios into the following broad categories:

  1. Language from different contributors. Distinguishing different hands, voices, etc.
  2. Stylistic differences from any particular contributor. Different emphasis, emotional delivery, typeface, handwriting, etc.
  3. Annotation where explanation or clarification is needed. Examples are unusual words, unknown words, slang, or local pronunciations.
  4. Contributions that cannot be transcribed directly or wholly as text. This includes changes, marginal notes, noises, gestures, and pauses.
  5. Parallel Contributions. This category is specifically related to audio.

STEMMA’s transcription support is designed to make material searchable, but also to support deep analysis. Some of these categories were already catered for in the cases of textual transcription, but supplementing them to cater for the remaining categories implicitly addressed audio transcription too. For instance, the <Alt> and <NoteRef> elements already catered for category #3 and needed no changes. The <Anom> element already represented textual anomalies, and so was extended to address the other anomalies in category #4.

The way that <Anom> was extended set the scene for the other extensions I will describe in a moment. Its existing taxonomy (see the namespace) was given extra items of Gesture, Noise, and Pause. Within these, though, the specific gestures and noises are described using text, by and for the researcher, and not by using some limitless software taxonomy.

The STEMMA transcription elements <ts> (typescript sources) and <ms> (manuscript sources) were supplemented by <voice> (audio sources), and each were enhanced to cope with categories #1 and #2. They were extended with new attributes of ‘id’ and ‘scheme’, For instance:

<ms id=’id’ scheme=’scheme’>An example sentence</ms>

What these attributes do is attach a key representing the contributor (e.g. a hand, or a voice) and a specific stylistic variation of that contributor. There are no taxonomies used here since the differentiation and description may be subjective; the differentiation is designed to support analysis, not simply a matter of rendition; and there need to be no constraints.

The last category (#5) is addressed by specific variations of the <voice> element that allow it to be used as a container for multiple contributions.

A small example of an audio transcription employing these features may be found at Dialogue Transcription. The <ts>, <ms>, and <voice> elements are documented at Descriptive Mark-up.

The rationale behind this approach is actually quite a well-known one, although not in this field. In the area of Web mark-up, HTML5 tries to separate structure and content from presentation, the latter being left to something like CSS. For the formatting of Web pages, this avoids cluttering the mark-up describing the structure and content of page information, and ensures a consistent presentational style is applied across the pages. For transcription, it avoids cluttering the mark-up describing the structure and content from various contributors, but leaving complete freedom to the researcher to describe these in narrative as part of their analysis process.

[1] Rarely usable in a computer-based transcription because the old symbol set does not correspond with available symbols in an electronic document.

Wednesday 19 April 2017


An original goal of STEMMA was to be able to represent rich-text narrative that could be used for authored works, including essays, memories, and reports. In addition, it aimed to support transcription, including transcribed extracts, which has quite specific requirements of its own.

STEMMA V4.1 has concentrated on its mark-up in these areas and has solved a number of long-standing issues with some novel approaches. Such was the success of the approach to textual transcription that this version also addresses audio transcription as a companion to it. I know of no other system that addresses both of these in a consistent manner, and certainly not when including rich-text authored work and semantic mark-up.


A goal of HTML5 was to separate structure and content from presentation in Web pages. STEMMA has applied a similar principle to its descriptive mark-up for both authored work and transcription.

STEMMA is not a presentation format. It therefore concerns itself with narrative structure, content, and semantics, but not the finer details of the presentation such as colours and fonts. STEMMA narrative may be transformed into any number of presentational formats for visualisation (e.g. HTML+CSS), and it is in these formats that such things would be configured, including page size, style galleries, choice of footnote/endnote/tablenote indicators, heading and cell formatting in tables, caption position, paragraph separation, styles for semantic elements, and so on.

Unusually, it has also applied this principle to both textual and audio transcription. Identifying the structure and content is more important that the finer details of their style and presentation, and the interpretation of any stylistic differences requires analysis rather than it simply being a display matter. For instance, marking where a manuscript used different colours in different places is more important than the specific colours and shades — that level of detail can be written in narrative for the reviewer rather than trying to use some limitless taxonomy for the software. Similarly with different written styles, which may or may not have been evidence of multiple authors. In a typescript document, it would equally apply to different fonts, font-sizes, ink intensity, marginal alignment, or even usage of grammar; all of these could have a bearing on the analysis of that document.

In audio transcription, this approach simplifies a complex area by giving freedom to the transcriber to detail the different voices, intonations, noises, and gestures.

Authored Work

The functionality of STEMMA’s descriptive mark-up has now evolved to the level where I can automatically generate blog-posts for research articles directly from my internal representation.

In order to demonstrate the new version, I have used the recent 5000-word article entitled Jesson Lesson to generate a fully-worked STEMMA example, available at JessonLesson.xml. This genuine research article included precise layout, transcribed extracts, tabulation, endnotes and tablenotes, and hyperlinked images. Its 47 endnotes included examples of reference-note citations, discursive notes, analytical commentary, and multi-source references — the handling of which was outlined previously at Cite Seeing — but also included examples of conflated citations where details of multiple people are placed in a single note for readability.

It was always a personal goal to produce better quality research articles, and so force STEMMA to address real-world scenarios rather than “desktop scenarios”. As a result of this, STEMMA’s general approach to citations has shifted slightly. Although support of citation-elements — implemented using its Parameter mechanism — has been enhanced, the focus of the computer-readable form is now on correlation and interrogation rather than mere formatting. The number of real-world cases (see list under Citations) is just too great for authored works to delegate formatting entirely to software that acts blindly from mere values. This version, therefore, finds a bridge between preferred hand-crafted forms and computer-readable citation-elements.

Another area that has been enhanced greatly is tables, which now support control over table width, column widths and alignment, captions, and tablenotes (i.e. citations deposited at the foot of a table).

Textual Transcription

The existing <ts> element, used to mark text transcribed from a typescript document, and the <ms> element, used to mark text transcribed from a manuscript document, both have new ‘id’ and ‘scheme’ attributes. These label the respective contributions with user-defined tags — ‘id’ for distinct contributions, such as different authors, and ‘scheme’ for stylistic variations — that can be described separately for the benefit of the reviewer.

For instance:

<NoteRef><Text Class=’Legend’>
bold-blue – text was written with a broad-tipped turquoise felt marker.

<ms scheme=’bold-blue’>This section is now out-of-date and is being reworked</ms>

The elements <page>, <col>, <p>, <line>, and <posn> now take SVG-like image coordinates (percentage displacement from top-left image corner) for linking transcription elements to a copy of the original document. One use of this is to support parallel scrolling of image and transcription for the end-user.

The associated image is specified by a preceding <ResourceRef> element identifying a Resource entity using the mode ‘SynchImage’.

Audio Transcription

For audio transcription, the <voice> element provides the analogy to <ts>/<ms>, and it similarly takes ‘id’ and ‘scheme’ attributes. This allows different vocal (or other audio) contributions to be distinguished, and also their intonation, emotional delivery, artificial accents, etc.

Additional features are supported in a way analogous to textual transcription:

  • Anomalous contributions from an individual that cannot be represented as text, including noises, pauses, and gestures – see <Anom>
  • Alternative word meanings, clarifications, or other notes – <Alt> and <NoteRef>, exactly as with textual transcription
  • Time synchronisation – time-stamping with <time>. This is analogous to the <posn> element, and other x/y coordinates, used for textual transcription.

For time-stamping, the associated recording is specified by a preceding <ResourceRef> element identifying a Resource entity using the mode ‘SynchAudio’, analogous to ‘SynchImage’ for images.

As well as marking distinct voices, these features include the ability to mark overlapping contributions and background contributions. An example demonstrating many of these features may be found at Dialogue Transcription.

Change Details

Specific changes to the data model include the following:

  • ‘WhereIn’ attribute added to Citation Parameter definitions. This finally provides the missing criteria necessary for the automatic generation of shortened subsequent reference-note citations. ‘Subst’ attribute added to Citation parameter values in order to override formatting, or provide a substitution for cases on of a value being unavailable.
  • <ParentCitationLnk> now allowed in both <CitationLnk> and <CitationRef> elements in order to create transient chained citations.
  • Quality element, within Source entity, moved inside the Frame element.
  • Review of entries in citation-layer-type namespace.
  • DataControl element of Resource entity supports attribution text.
  • Control of table widths, and individual column widths and alignments.
  • Ability to align images when embedded within narrative.
  • Ability to hyperlink images embedded in narrative.
  • Requirement for enclosing Narrative element dropped for Text elements, except for top-level Narrative entities. Text elements can now be nested.
  • <cb> replaced with <col>, and relationship between paragraphs and columns now reversed (paragraphs now within columns).
  • ResourceRef Mode=SynchImage allows synchronisation between images and transcriptions.
  • Corresponding SVG-x/y coordinates added to elements <page>, <col>, <p>, and <line>. Additional <posn> element defined to associate coordinates with arbitrary text locations.
  • <Page>/<Line> renamed to <page>/><line> and moved alongside <p>/<col> as related to structure and content rather than semantics.
  • Mode=Tablenote attribute supplementing Foonote and Endnote in various places.
  • Text-element Header=boolean attribute replaced with Class=Header | H1 | H2 | H3 | Caption | Footnote | Endnote | Legend | Tablenote.
  • Text-element Class=Caption attribute used in Resource/ResourceRef and tables for generating captions.
  • Text-element Class=Footnote | Endnote | Tablenote attribute used in CitationRef to allow pre-formed (preferred) citations.
  • Deprecated the <Text> attributes Abstract=boolean, Extract=boolean, Manuscript=boolean, and Transcript=boolean..<voice> mark-up added to supplement existing <ts>/<ms> mark-up. <ts>/<ms>/<voice> all enhanced to cope with different hands, voices, fonts, colours, etc.
  • In transcripts of audio recordings, support for multiple voices, overlapping dialogue, intonation, gestures, noises, pauses, timestamps, etc.
  • ResourceRef Mode=SynchAudio allows synchronisation between audio recordings and transcriptions, analogous to SynchImage for textual transcription (above).
  • Complete revision of Mode values for CitationRef element.
  • Relaxation of Date Parameters in order to cover the full range of calendars. One requirement was to represent the date-of-issue for newspaper sources that predated the Julian-to-Gregorian changeover.

Further refinements to this data model are uncertain as it has now achieved the level of stability and functionality that was required for its serious usage.