The basic idea
Respeaking is a method of live captioning in which a trained person listens to a speaker and repeats what they hear into speech recognition software. The software, which has been adapted to that person’s voice, converts the repeated speech into text. The text is then displayed to viewers as captions or subtitles.
The method is also known as voice writing in some places, and in some regions it has been widely used for live television subtitling. The person doing it is called a respeaker.
The idea sounds roundabout. Why not let the software listen to the original speaker directly? The answer lies in what recognition software finds easy and hard.
Why a person in the middle helps
General-purpose recognition software has to cope with any voice, microphone and room. Real events often provide difficult conditions: several people talking, background noise, strong accents, fast speech, or audio from a telephone line. A respeaker acts as a translator from that messy sound into a clean one. They are in a quiet space, speak into a good microphone, and use a voice that the software has been trained to recognize. The software then has a much easier job.
The human also brings understanding. They can interpret unclear audio from context, know when a name should be capitalized, and decide how to word something so the text reads well.
What a respeaker does
A respeaker’s task combines several skills at once, which is why training is usually considerable.
- Listening and speaking at the same time. The respeaker is repeating one phrase while already listening to the next, a demanding mental skill similar to simultaneous interpreting.
- Condensing. Where the speaker is very fast, the respeaker may shorten or simplify, dropping repetition and filler while keeping the meaning. This is a form of text condensation done in real time.
- Speaking punctuation. Most recognition software does not add punctuation reliably, so the respeaker often says commands aloud, such as “comma” or “full stop”, or uses other cues defined by the software.
- Marking speakers and sounds. They may add speaker changes or indicate important sounds through spoken commands or quick keystrokes.
- Correcting. If the software misrecognizes a word, the respeaker may fix it using keyboard or voice commands, though correcting costs time.
Preparation
Respeakers usually prepare before a session. The software’s vocabulary can be extended with names, places and specialist terms that are likely to come up. Having access to scripts, running orders and slides makes this far more effective. Some respeakers also practice with recordings from the same speaker or topic. Because the software is trained to the respeaker’s voice, they generally use their own profile rather than a shared one.
This is also why briefing material matters. A glossary shared in advance can change the outcome noticeably.
Where respeaking tends to work well
- Live television and broadcasts. Where speech is fast and varied, a respeaker can condense sensibly and maintain readable text.
- Poor or varied audio. Because the respeaker’s voice is what the software hears, problems with the source audio are partly absorbed by the person.
- Meetings and events with specialized vocabulary. With a prepared vocabulary and a person who understands the subject, results are often better than with fully automatic recognition.
- Settings where stenographic writers are scarce. Respeakers can be trained for some languages and regions where shorthand keyboard methods are less widespread. Availability varies widely by place.
Where it has limits
Delay. The respeaker has to hear and then repeat, so there is an extra step compared with direct recognition. The lag is generally noticeable, though it varies by person and setup.
Errors from software. The recognition step can still produce wrong words, particularly homophones and unusual names. Errors may reflect the software as much as the person.
Condensing is not a transcript. Because respeakers shorten fast speech, the text is not always word for word. That suits following an event but is not a precise record. If an exact record is needed, a later review against a recording is advisable.
Fatigue. The work is intense, and respeakers generally work in shifts, often short ones, so longer events need more than one person.
Speaking environment. The respeaker needs a quiet place, which may mean a separate room or a soundproofed booth. For remote work, the audio link must be good.
Cross-talk and overlap. Several people speaking at once is hard for a person to follow too.
Compared with other methods
Respeaking sits between two other approaches. Stenography uses a shorthand keyboard and is often favored where very fast or very exact output is needed. Fully automatic captions are cheaper and instant but depend heavily on audio quality and vocabulary. Respeaking offers human judgment with a method that many people can be trained to do, though the details depend on language, software and local practice. None of these is best everywhere, and many providers combine them.
Checking quality
If you are choosing a respeaking service, it is reasonable to ask how the respeaker prepares, how many people will be used for a long event, how corrections are handled and what delay to expect. Asking for a short test with your speakers, or sample output, is more informative than a general claim of accuracy. Where accuracy matters, have important material reviewed by human reviewers afterward, as with any live method.
Requirements for access vary by location. A qualified professional or the relevant authority can say what applies to your event.