Guidelines for RT-02 Transcription

Overview

• word-for-word transcription, using standard orthography and standard capitalization.
• Removal of all text not present in the audio - i.e. copyright information, closed-caption speaker identification
• use end-of-sentence punctuation (periods, question marks), commas, and hyphens for hyphenated words but no other punctuation (please remove all other extraneous marks, ie quotation marks, double hyphens, etc.)
• speaker turns are indicated & timestamped, and provided with a unique speaker id, containing gender information and the speaker name if possible.
• intermediate breakpoint that are automatically inserted  - are not initially accurate, and no effort is being made to ensure accuracy aside from turn boundaries, and non-speech

     <sr 115.45> <male, spkr_1>
  <t 129.54> <male, Bill_Clinton>
  <t 135.67> <female, spkr_2>
 

Labels

sr	start of <section type=report> news story section
sf	start of <section type=filler> filler section that was classified as < sn
sf	start of <section type=non-news> non-news section - commercials,etc
t	start of (non-initial) speaker turn within section
b	turn interval breakpoints (not accurate)
e	end of turn within section, followed by a non-speech region
o	start of overlap region (speaker one is interrupted by speaker two)
e1	end of overlap where speaker one stops and speaker two continues
e2	end of overlap where speaker two stops and speaker one continues





Punctuation	Include punctuation for ease of transcription	,.?
Numerical sequences	Written in full	twenty-five, one hundred and six
Non-standard contractions	spell out in full  - gonna, wanna	going to, want to
Misc speaker noises	indicate occurrences	{breath}, {laugh}, {cough}, {sneeze}
Titles of address	Written in full	Doctor, Mister
Proper names	??	??
Acronyms, pronounced letters	use tilde ~	~FBI, Washington ~DC
Interjections		uh-huh, mhm, yeah, uh-oh, whoa
Non-lexemes	%	ah, eee, eh, ew, ha, hee, huh, hm, oh, oo,  um, uh
Word fragments	indicate partial word with  - followed by restart --	inspecto- --
Pronounced Acronyms	@	@NAFTA, @AIDS
Made-up word	+	+poodlish
Difficult to understand words . Best guess attempt.	(( ))	we went to (( )) to grab a bite to eat.
Background noise,non-speech events within a turn.(also see usage of <e convention)	Indicated	[[NS]]






Speaker Identification

For broadcast speech the goal is to identify speakers as precisely as possible. At the very least, each unique speaker in a recording should have a unique identification. Further information to be added includes speaker gender and proper name if possible.

    Gender
    The possibilities are "male" , "female", "child" "altered", "unison". "Unison" occurs in situations where two or
    more individuals say the same thing at the same time.

               SHORTCUTS  for gender insertion

            Alt-f           female
            Alt-m         male
            Alt-c          child
            Alt-u       unison  (more than two speakers with identical utterences)


    Proper names
    Whenever possible, include the proper name of the speaker. Examples of proper names include Jacques_Cousteau, William_Cohen, and Madeleine_Albright.

    If a speaker is not identified within a recording, a unique numerical index is used. For the convenience of the transcribers, a broad categorical identification may be used. The two categories currently supported are Reporter and Speaker.

    Reporter refers to either the anchor of the news broadcast, or the reporter on location giving the story Speaker,  on the other hand, refers to anyone interviewed on tape by the Reporter, when that person is not identified by name. Numbers do not overlap. Each successive anonymous speaker  has a unique number, regardless of the category the speaker is assigned to. For example, the following sequence is entirely possible:

        reporter_1
        reporter_2
        spkr_3
        spkr_4
        reporter_2 (assuming it is the same voice as the previous Reporter_2)
        reporter_6 (a new reporter distinct from the two above)

    Native, non-native, and altered
    In English broadcast, "native" speakers are standard North American dialects. These are not marked.
    "Non-native" speakers, are determined as foreign accented speakers, including British-english
    speakers."Altered" is used to tag deliberately altered voice patterns, for instance in the case of a disguised
    informant's speech, or for machine generated speech.
 

Examples

         <sr 1.402> <<male, Leon_Harris>>
         <sr 158.244> <<female, Joie_Chen>>
         <t 196.813> <<male, spkr_1>>
         <t 498.314> <<male, non-native, spkr_3>>
         <t 567.215> <<male, altered, spkr_4>>




  Overlapping speech

during interruptions,  <o> & timestamp will be used to indicate the start of the overlapping speech region corresponding to the interrupting speaker.

- an overlapping speech region is determined by overlapping word boundaries, rather than the exact point in the waveform which may require splitting words.

<t 1122.443> <male, Jacques_Cousteau>
about the same thing I
<o 1123.276> <female, spkr_2>
SPEAKER1:understand {laugh}
SPEAKER2:well,
<e1 1124.256>
oceanography is new exploration and we're not
 
• <e1 indicates that the first speaker has now stopped, while the second speaker has continued to speak. If BOTH speakers stop  at the same point in time, the next speaker turn indicated by a if overlap ends with non-speech section (silence, music, etc.), mark beginning of non-speech section with a terminating  <e> & timestamp

The [[NS]] tag can be used when there is an area within a turn that has no speech within it , i.e. a musical interruption, or extended background noise.
 

<b 123.456 >
The crowd was furious.
<b 124.567>
[[NS]]
<b 128.987>
Calm was soon restored by the arrival of the riot police.

 

• indicate disfluencies by using hyphen to mark partial words; transcribe pause fillers, e.g.

   We're jus--  just waiting for that uh tha-- that report to to come in.
 

• transcribe standard English contractions as they're spoken: they're, won't, isn't, don't, etc.

 
• for non-standard contractions like "gonna" and "wanna" spell out the entire word: going to, want to.

 
• identify extended non-speech sections (music, dead air, sound effects) with <e> and timestamp at beginning of section, followed by <t> and timestamp when the speech resumes, e.g.

 

<t 148.57>
Gunfire filled the air.
<e 154.50>
<t 170.89>
That sound greeted early morning visitors on tuesday.


 
NOTE
Multiple speakers
In situations when you have several people speaking at once, (more than two) and it is very difficult to make them out, insert an <e tag at the start of the confused section. Then start the new turn at the next availabe clear section. This is the same treatment that is used for extended period of non-speech.
 

        <t 223.456> <<male, spkr_1>>
        <b 225.678>
        <e 230.302>
        <t 232.563> <<female, spkr_3>>>

Speakers start simultaneously
Create start times for the speakers that are about one tenth of a second apart, and insert the overlap tag.



<t 1123.176> <male, Jacques_Cousteau>
about the same thing I
<o 1123.276> <female, spkr_2>
SPEAKER1:understand {laugh}
SPEAKER2:well,
<e1 1124.256>
oceanography is new exploration and we're not
 

 

Checking and separation of unintelligible (( )) speech

<t 859.405> <<male>>
Not only do we methodically destroy the coastal fringe
<b 863.598>
but we also throw back our toxic *1(( ))*2 directly in the sea
<b 868.453>
or under the sea when we feel ashamed.

i) Use Ctrl-c s to "send" the segment to the waveform window. Then find the section which cannot be understood and isolate it as you would if you were placing breakpoints
ii) Return to the text area, and place the cursor in front of the brackets (*1)
iii) Hold Alt and the middle mouse button (M2), drag the cursor over the (( )) region, and release M2(*2). There is no highlighting over the region to show you that it is enabled, but you will receive a prompt for implementation of the change if it happens correctly.




Syntax Checking

Common Messages include:



time-stamp without text data?
 -The timestamp does not contain corresponding transcript data



time-stamp follows non-empty line
 - an empty line should follow each transcribed timestamp.



turn should be on single line
 -only one turn permitted for each line.



<English ...> should not be inside (( ))
 -foreign speech should not be contained within "guess" brackets.



<English (()) > has no text content)
 -rather the "guess" should be contained within the foreign language bracket



closing angle (`>') should be followed by space
self evident



bracket error with '[]'
may be a number of possibilities



bracket error with '()'
may be a number of possibilities



punctuation should be inside `((...))'
If completely necessary, punctuation should go inside these brackets - in most case, no punctuation will be necessary.



punctuation should be inside `<... >'
If there is punctuation immediately outside an < , please place on the inside of the bracket.



bad spacing around punctuation `.'
There should exist a space after punctuation



bad spacing around punctuation `?'
There should exist a space after punctuation



closing paren (`))') should be followed by space
self explanatory



turn contains ILLEGAL CHARACTER `!'
Some characters are not allowed within the text - for instance, exclamation points - please let your language leader know when you come across this error warning.



digits found in text
There should not be any numerals in the text outside of the timestamps